Search Results

If you could reliably predict the future, even just a little bit, you would possess something akin to a superpower. We are not talking about the sci-fi sense of time travel or crystal balls. Imagine instead being able to predict whether your health condition will improve or degrade over the next three months, which startup will dominate its market in five years, or, much more commonly, exactly what the weather will be like tomorrow! In our latest episode, we dive into the complex world of time series data, exploring the subtle but crucial distinction between prophecy and forecasting. AI is not about seeing the  future; it is about prediction under uncertainty. It is about estimating which futures are more likely than others. As it turns out, AI is incredibly good at this—under the right conditions. The Magic of “Time Series" Data To understand how a machine predicts the future, we first need to understand the concept of a time series . In data science, a time series is simply a data structure consisting of measurements taken over time. It is a sequence where the strict order of events matters. Every day, we generate countless time series. Your phone battery draining from 97% to 93% to 88% is a time series. Your beat-by-beat heart rate is a time series. Even a city’s fluctuating energy demand throughout the day is a time series. In all these examples, the core idea is the same: you are measuring a variable that changes in relation to its past. However, not all sequences are time series. If you roll a six-sided die one hundred times, you could represent that data sequentially, but it would not be meaningful. The result of your next roll is completely unconstrained by your previous rolls. For AI to predict the future, it requires an environment where the past actively constrains  the future. The stronger the constraint, the more predictable the system. The Enemies of Forecasting: Randomness and Feedback Loops Even when a system is theoretically predictable, there are practical challenges that prevent AI from being a flawless oracle. The Problem of Randomness:  Randomness is a fundamental mathematical difficulty in forecasting because it compounds over time. If an AI is slightly unsure about where an object will be in one hour, predicting its location in ten hours becomes exponentially harder. Furthermore, unprecedented, chaotic events—like the outbreak of COVID-19 or a sudden factory fire—introduce randomness that historical patterns simply cannot capture. Unknown Unknowns:  Sometimes the most important driver of a future event is not in your dataset at all. Imagine trying to predict supermarket sales in early 2020. If you only had data from 2000 to 2019, your models would be completely baffled by the massive, sudden spike in toilet paper demand. The AI model itself did not break; the world changed, and the cause (panic psychology amplified by the news) was not a variable the AI was measuring. Feedback Loops:  Because we live in an interactive world, predictions often change behaviour, and behaviour changes what happens next. Consider Google Maps predicting a traffic jam. If the app tells thousands of drivers that a specific route will be heavily congested, those drivers will take alternative routes. The original route may then become unexpectedly clear. In financial markets, this phenomenon is even more pronounced. If a highly influential investor (or AI system) predicts a stock will surge, people will follow that advice, creating a surge in demand that artificially fulfils the prophecy. In these environments, the forecast itself actively alters the future. A Brief History of Looking Forward Humans have been fascinated by understanding time and predicting the future for thousands of years. Early attempts relied on sweeping narratives, like the ancient prophecies found in the Book of Revelation. These were symbolic and vague enough to be reinterpreted later, meaning there was no way to mathematically falsify them. True forecasting—modelling the set of possible outcomes as a distribution—requires precision and an acceptance of uncertainty. 1600s to 1705:  The birth of celestial mechanics. Using newly formulated laws of planetary motion by Johannes Kepler and Isaac Newton's theories of gravity, Edmund Halley studied historical comet observations. He made a staggering 50-plus-year forecast, accurately predicting the return of Halley’s Comet in 1758. 1960 to 1969:  Rudolf E. Kalman published the Kalman Filter, a mathematical algorithm that proved vital for the Apollo 11 moon landing . Because spacecraft sensors drifted and were imperfect, scientists used physics to predict the ship's position, tracked how wrong that prediction might be, and corrected it with sensor data. By incorporating uncertainty, they successfully navigated to the moon. 1970:  Box and Jenkins developed ARIMA (Autoregressive Integrated Moving Average), a general class of statistical models. Instead of humans defining the strict physics of a system, these models were flexible frameworks that learned patterns directly from historical time series data. 2023:  We entered the era of deep learning with Google’s release of GraphCast . For over 50 years, global weather forecasting relied on massive supercomputers running numerical physics simulations. GraphCast, trained on decades of historical weather data, completely outperformed these traditional simulators, running global forecasts in under a minute. How Does AI Perceive Time? It is easy to imagine an AI experiencing the continuous flow of time just as we do, but computers do not inherently understand concepts like “Tuesday" or “3:00 PM." To a machine learning model, everything is simply a number. Time is perceived in discrete, fragmented moments. To give AI clues about time, data scientists might include variables representing the time of day or the day of the week, acknowledging that a Monday behaves differently from a Saturday. However, AI's true superpower lies in its ability to handle multivariate  data. Visualisation of the raw weather data. There is a clear seasonal trend, but there are some extreme outliers among the pressure measurements, which make it difficult to see any patterns ( Sebastian Callh, 2020 ). A univariate  time series is just one line wobbling over time, like tracking daily temperature. A multivariate  time series is like a massive spreadsheet moving through time—simultaneously tracking temperature, humidity, pressure, rainfall, and wind. Most real-world systems are not just one line; they are many lines interacting. Rather than constantly observing a continuous stream, the AI looks at data in isolated “windows." It might process 90 days of weather data all at once to predict the next seven days. From Memory to Attention: Under the Hood Historically, engineers built Recurrent Neural Networks (RNNs)  to process time series data sequentially, in the order it was generated. An RNN reads a sequence one step at a time, carrying a little “memory" forward. However, RNNs suffered from vanishing gradients —older information quickly faded, and the model forgot important past events. This led to the development of Long Short-Term Memory (LSTM)  network. LSTMs introduced a brilliant forget mechanism , allowing the model to selectively remember and forget certain timesteps. If an LSTM is analysing weather every ten minutes, it learns to forget hours of light drizzle while heavily retaining the memory of a massive pressure drop, ensuring it does not get overwhelmed by useless noise. More recently, the AI landscape has shifted towards Transformers  (the technology underpinning tools like ChatGPT). Transformers abandon the step-by-step reading approach entirely. Instead, they use an attention mechanism , looking at all time steps simultaneously. If a Transformer is trying to make a prediction for today, it asks: “Which other moments in history are most relevant to this exact situation?"  Instead of reading a book left to right, it can instantly flip to the exact page that contains the answer. The Three Core Tasks of Time Series AI When we feed these complex neural networks our historical data, we are usually asking them to perform one of three tasks: Forecasting (“What comes next?"):  For example, the UK's National Energy System Operator relies heavily on forecasting to balance the power grid. Storing energy is highly inefficient, so supply must perfectly match demand. If the AI overestimates demand, power plants burn wasted energy, resulting in higher consumer prices. If it underestimates demand, the country could face rolling blackouts. Classification (“What is happening right now?"):  Wearable tech uses time series data to classify human behaviour. A smartwatch tracks your heart rate, movement, and breathing over time. Without you pushing a button, it can look at that window of data and categorise whether you are sitting at a desk, walking, or swimming. Anomaly Detection (“Is this behaviour weird?"):  Anomaly detection asks if a sequence deviates from what is considered “normal" for that specific system. This is how modern bank fraud algorithms operate. Buying a coffee in London is normal behaviour. Buying a coffee in London three minutes after a purchase is made in Tokyo is a glaring anomaly. This is also how the Apple Watch’s medically approved Atrial Fibrillation (AFib) detector works. It tracks your heart rate to learn what your “normal" looks like. It does not alert you after a single irregular beat; it waits for multiple irregular readings over time. While studies show the watch's sensitivity  is somewhat low (meaning it misses some real AFib episodes), its specificity  is incredibly high—meaning it almost never gives a false alarm, reliably catching early warning signs of strokes. Weather forecasting using a trained AI model. The model has picked up on the seasonality in the data and successfully extrapolates. The final observations are somewhat irregular, and not perfectly captured by the model ( Sebastian Callh, 2020 ). The Trap of Spurious Correlations While AI is incredibly powerful at spotting patterns, it has zero common sense! This leads to a dangerous pitfall known as spurious correlations —when two variables appear mathematically linked but have no causal relationship. For example, historical data shows a near-perfect correlation between the amount of cheese consumed in a given year and the number of people who tragically die by becoming tangled in their bedsheets. Similarly, there is a flawless statistical correlation between the release of Nicolas Cage movies and the number of swimming pool drownings. Ice cream sales and shark attacks also rise together perfectly —though both are simply caused by the arrival of warm summer weather. An AI blindly hunting for patterns might confidently use Nicolas Cage movie releases to forecast public safety risks. We know that correlation does not equal causation, but an algorithm splitting millions of tiny correlations across a vast, multivariate dataset does not know the difference. The primary skill in data science is shifting from asking “how do I build a forecast?"  to “how do I critically evaluate one?" The Ethics of Having a Crystal Ball As AI forecasting becomes more embedded in our infrastructure, we must confront the ethical implications of a machine that anticipates human behaviour. First, there is the issue of historical bias. AI models can only learn from the data of the past. If a predictive policing model is trained on decades of arrest records, it learns where arrests have historically been made, not necessarily where crime occurs today. Deploying that model into the real world simply entrenches and accelerates historical inequities, trapping society in the prejudices of the past. Second, predicting the future threatens individual autonomy. The more granularly an AI can predict consumer or patient behaviour, the more that knowledge can be used to nudge or restrict people. If a wearable device accurately forecasts that a patient's health will severely decline over the next year, a health insurance company equipped with that data could preemptively price that individual out of coverage. AI forecasting does not magically see the future. It learns from the past and makes highly calibrated guesses about what similar-looking futures might bring. The world is fundamentally chaotic and constantly evolving. As Kieren and Riku point out, the future is always yet to come, and the only truly predictable thing about it is change. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts — whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

It is officially 2026, and we can no longer pretend that AI is a niche tool for just the super tech-savvy. Just looking at ChatGPT alone, as of late 2025, the platform hit over 800 million active weekly users—that's roughly 10% of the world’s population! It's become the most rapidly adopted technology in history, outpacing the internet and the personal computer. But now that the hype has settled into daily habit, we need to ask the difficult question of whether we're actually using these tools to get smarter, or are we just outsourcing our thinking? In this post, we explore the latest reports from OpenAI, Anthropic, and MIT to uncover the reality of AI usage in 2026. We'll look at the “productivity paradox," the dangers of “cognitive debt," and how you can stop treating AI as an oracle and start using it as a co-pilot. How the World Uses AI Before we discuss how we should use it, we need to look at how we are using it. In the UK alone, about a third of the population (22.5 million people) were using AI tools as of October 2025 . Across the pond in the US, over half of adults are using generative AI in some capacity. Interestingly, the usage has shifted dramatically toward the workplace. According to a report from Anthropic , 40% of employees now report using AI at work—double the figure from just two years prior. OpenAI data backs this up, showing that work-related queries make up over 70% of usage among paid users. So, what are we actually asking ChatGPT to do? OpenAI's breaks down their data into three main categories: Practical Guidance (24%): This includes tutoring, creative brainstorming, and general problem-solving. Seeking Information: Roughly 30% of search queries in the UK now show AI overviews. Writing: This accounts for about 40% of work queries, ranging from drafting emails to generating reports and marketing copy. While these numbers suggest a massive uptake in productivity tools, the reality of the output is far more nuanced. The Productivity Paradox If you ask people, they will tell you AI is a lifesaver. In a field experiment by METR in 2025, experienced computer programmers self-reported that using AI tools would speed them up by about 20% to 30%. However, the reality was starkly different. When their performance was actually measured, they were 20% slower . Experts and study participants (experienced open-source contributors) substantially overestimate how much AI assistance will speed up developers—tasks take 19% more time when study participants can use AI tools like Cursor Pro ( METR, 2025 ). This is the “productivity paradox." We perceive speed because the initial friction of the “blank page" is removed, but the time spent debugging, verifying, and wrestling with the AI's output often negates those gains. Furthermore, while creativity scores in some studies went up, user motivation dropped by 11% and boredom increased by 20%. We therefore have to ask: Is a marginal gain in output worth a significant drop in fulfilment? That isn't to say there are no benefits. An MIT study found that workers using ChatGPT completed tasks 40% faster with 18% higher quality output. GitHub reported developers were 88% more productive on repetitive tasks. But these gains are not guaranteed—they depend entirely on how you engage with the tool. The Hidden Risks of Mass Adoption If we rely on these tools blindly, we risk falling into several cognitive traps that experts are only just beginning to understand. Cognitive Debt and “Brain Rot" We often talk about “cognitive offloading"—letting the machine do the heavy lifting. But what happens to your brain when it stops lifting? A study titled “ Your Brain on ChatGPT " by MIT researchers compared groups writing essays with and without AI assistance. The results were alarming. The group using ChatGPT showed significantly less brain activation during the task. More worryingly, this created “cognitive debt": even when these participants returned to working independently without the AI, their brains failed to reactivate to previous levels. This mirrors findings regarding social media and “brain rot." Studies have shown that infinite scrolling on platforms like TikTok can wipe short-term memory and reduce our ability to retain information. If we treat AI as an infinite scroll for text generation, we risk a similar degradation of our critical thinking faculties. The Homogenisation of Ideas Perhaps the scariest risk of so many people using the same few chatbots is the “homogenisation of ideas". In the MIT study mentioned above, experts noted that the AI-assisted essays felt “soulless," recycling the same stock ideas and phrases. If 800 million billion people rely on a single algorithmic “seed" to generate their thoughts, we inevitably erode the diversity, complexity, and richness that defines human language. We risk creating a feedback loop where we all sound like the same “average" statistical next-word prediction. The Snowball Effect of Bias We know AI models contain bias because they are trained on human data. But researchers have identified a phenomenon called the “snowball effect," where humans actually learn the bias from the AI and then amplify it. In one study , participants were shown images of faces labeled by AI. If the AI biasedly labelled a neutral face as “sad," the humans adopted that bias. Even after the AI was removed, the humans continued to interpret neutral faces as sad, effectively learning the machine's skewed worldview. The Jagged Technological Frontier One of the hardest things to navigate is knowing when to trust these tools. We often assume that if a computer is smart, it is smart at everything. However, there is something known as Moravec’s paradox: computers are often great at things humans find hard (like playing chess) but terrible at things humans find easy (like identifying objects within an image). AI capabilities are therefore often described as a “jagged frontier". A model might be capable of writing an incredible essay on string theory but fail to count the number of Rs in the word “strawberry". Just because a model passes the bar exam does not mean that intelligence scales to every task. This figure displays the AI frontier as jagged. Tasks with the same perceived difficulty may be on one side or the other of the frontier. ChatGPT produced this image starting from the authors’ prompts ( Harvard Business School, 2023 ). This makes using chatbots as “oracles" very dangerous They are people-pleasers, trained via Reinforcement Learning from Human Feedback (RLHF) to provide answers that look helpful. They are prone to sycophancy—telling you what you want to hear rather than the truth. As Professor Miranda Mowbray noted , this is like sugar: it tastes good, but it isn't necessarily good for you. How to Use AI Responsibly in 2026 So, how do we prevent the brainrot? We don't need to throw our devices away, but we do need to change our relationship with them. Here is a guide to using chatbots without losing your agency. Be the Pilot, Not the Passenger The most crucial shift is mental. View AI as a collaborator or co-pilot, never as an oracle. Never ask it to do the work for you; ask it to do the work with you. Use the “Socratic Co-Pilot" Method Instead of asking for the answer, ask the AI to coach you. Many models now have learning modes that adopt a Socratic style—answering a question with a question. This forces you to engage your brain and actually learn, rather than just copy-pasting the result. Iterate, Don't Delegate Don't go from zero to one with AI. Draft First: Always do a “brain-first pass." Even if it is just 30 seconds of bullet points, establish your own view before consulting the algorithm. Critique: Use the AI to critique your draft or offer alternative perspectives. Verify: Demand that the AI shows its reasoning. Many tools now allow you to click “show thinking" to trace the steps the model took. You Write the Summary Regardless of how much help you get with research or structure, you must write the final summary. You have a responsibility to the words you own. If you don't write it, you haven't processed it, and you certainly won't remember it . Practice Metacognition Metacognition is “thinking about your thinking". When you reach for an AI tool, pause and ask yourself: Why am I using this? Do I want to learn this skill, or do I just want the output? Being intentional about your usage helps you avoid the zombie-like state of cognitive offloading. Looking Ahead As we move deeper into 2026, the question is no longer about access to AI, but about our relationship with it. We must ensure we are using these tools to augment our intelligence, not replace it. If we are not careful, we risk a future where our work is faster but our minds are slower; where our output is higher but our creativity is homogenized. The goal is to keep the human in the driver's seat. Use the tool, but don't become the tool! If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts — whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

For decades, science fiction has trained us to fear the moment the lights turn on inside the machine. From HAL 9000 in 2001: A Space Odyssey to modern depictions of AGI, we worry about the moment computers become conscious entities that can suffer, plan, and potentially attack us. But recently, this hasn't felt so fictional. In 2022, a Google engineer claimed a chatbot was sentient . Anthropic recently hired an “ AI Welfare Researcher ," and they estimate a 15% chance that chatbots are already conscious. If AI is becoming more autonomous — acting in markets, executing contracts, and mimicking human emotion — do we need to start talking about AI Rights ? In this episode, we sat down with Dr. Miranda Mowbray , a mathematician turned AI ethicist at the University of Bristol. She took us on a journey from cybersecurity to the legal rights of rivers, dismantling the “consciousness" hype to reveal the practical legal realities we actually need to worry about. The Consciousness Trap The biggest hurdle in discussing AI rights is our tendency to anthropomorphise . We are wired to treat things that sound human as human. Current AI models are trained for what Miranda describes as “ sycophancy " . They are rewarded for agreeing with us and validating our prompts, much like an eager intern trying to please a boss. This can feel like compassion, but it is actually just statistical affirmation — sugar for the user's ego. Because we can’t even prove fellow humans are conscious, basing laws on “machine consciousness" is a dangerous game. To prove why, Miranda applied standard philosophical tests for consciousness (autonomy, complexity, ability to plan) to a surprising candidate: Malware . Examples she mentions include: Passing “a version” of the Turing test : scams that convincingly mimic humans. Planning and reasoning : trivial for software to mimic in narrow contexts. Unpredictability : a random number generator will do. Self-replication : malware already replicates constantly. Complex networks : botnets can create massive interconnected systems. Autonomy: The Conficker worm survived for over a decade without human help, jumping from machine to machine. If we grant rights based on these criteria, we would accidentally grant rights to malicious software designed to steal from us. As Miranda bluntly put it: “We do not want to give rights to malware!" Rights for Rivers and Stone Statues Shiva Lingam, Pashupatinath Temple, Kathmandu, Nepal If consciousness is a bad metric, should we dismiss AI rights entirely? Not necessarily. Miranda pointed out that the legal system grants rights to non-living things all the time for “legal convenience" . Corporations: They are “legal persons" so they can sue, be sued, and own property. Rivers: The Magpie River in Canada has rights to flow and not be polluted, a mix of Indigenous belief and environmental protection. Religious Icons: Miranda shared a fascinating example of a lingam (a statue representing the Hindu god Shiva) that holds legal rights to manage temple finances. In these cases, “rights" aren't about the inner feelings of the river or the statue. They are legal tools used to protect the humans and communities who rely on them. So what would “AI rights” actually do? This is where the conversation becomes practically useful. As AI systems become more autonomous — acting in markets, executing contracts, and mimicking human emotion — you can see why some people start reaching for the corporate analogy: treat AI like a legal actor to close the accountability gap. But Miranda was very cautious about a key failure mode: Rights can become a way for companies to offload responsibility. The “right to clear instructions” (and why she dislikes it) Kieren and Riku float an idea inspired by the classic “ paperclip maximiser ” thought experiment: maybe an AI should have a “right to clear instructions”, so it can’t do absurdly harmful things due to vague prompting. Miranda dislikes this framing because it shifts the burden onto users: “You prompted it wrong” “You didn’t specify every edge case” “Not our fault!” That approach weakens incentives for companies to build safer products. Rights that protect humans , not models She offers a more compelling example from scholar Kate Darling : pet robots or chatbots might deserve protections against abuse because humans form emotional attachments — and harming the robot harms the person. This is a recurring theme: Sometimes “rights for X” are really rights for people, implemented through X. The Mathematics of Fairness Finally, we touched on the “Right to be Unbiased." This sounds great in theory, but Miranda explained why it is mathematically impossible to perfect. Using the famous COMPAS case (an algorithm used for bail decisions in the US), she explained that there are different mathematical definitions of fairness. Equal False Positive Rates: Predicting people are high-risk when they aren't. Predictive Parity (Precision): If the system says you are high risk, how likely is it that you actually are? In the COMPAS case, the system satisfied one definition but failed the other (showing bias against Black defendants). Mathematically, it is often impossible to satisfy both definitions at once . This means we can't just program “fairness" into a machine. We have to make difficult social choices about which definition of fairness we value most — and that requires consulting the people affected by the system, such as victims and defendants. What Can You Do? It is easy to feel powerless against Big Tech, but Miranda emphasised that the public has power. Join Organisations: Individuals have little power alone, but collective groups (consumer rights, digital rights activists ) have significant influence. Participate in Consultations: Governments frequently run public consultations on tech policy. They want to hear from citisens. Vote with your Feet: Tech companies care about market share. If a product is unsafe or biased, consumer pressure forces change. The Bottom Line We shouldn't get distracted by sci-fi debates about whether a robot has a soul. The real questions are about legal convenience, corporate accountability, and mathematical fairness. We need regulation that is transparent, accountable, and targeted — not to protect the feelings of the machine, but to protect the rights of the humans using it. Next Step: Interested in how you can actually influence AI policy? Look up open public consultations on technology in your country (like the UK Parliament or US Congress websites) this week. Your voice matters more than you think. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts — whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

AI is reshaping the way we work, live, and interact at a pace not seen since the deployment of the printing press six centuries ago. As AI algorithms become deeply embedded in daily life — from filtering your social feeds and navigating maps to determining loan approvals and hiring decisions — ensuring these systems align with core ethical principles is now a crucial societal priority. This episode, features special guest Dr. Huw Day, researcher and co-organiser of Data Ethics Club , an open-to-all journal club on data and AI ethics. Together, we tackle the profound ethical challenges accompanying this technological leap. We explore why ethical AI is not about hindering technology, it’s about shaping AI development in a way that maximises benefits while minimising harm. The urgency of this conversation is undeniable. As Riku notes in the episode, AI is unique because, unlike past technologies which only amplified human actions, AI is the only technology that actually takes power away from humans  by making decisions instead of us. Furthermore, these complex tools are only as reliable as the data they are trained on and the people who build them. The seriousness of unquestioned development was underscored recently by a huge coalition of experts signing the Call to Ban Superintelligence , urging a halt on AI vastly smarter than humans until there is scientific consensus on safety and strong public buy-in. Core Principles of Ethical AI Ethical frameworks around the world, notably UNESCO’s global Recommendation on the Ethics of AI (endorsed by 193 member states), coalesce around several core principles aimed at protecting human rights and societal well-being. Principle Explanation & Importance Fairness and Non-Discrimination AI systems must treat individuals and groups equitably, avoiding the reinforcement of societal biases. Developers must strive to eliminate bias in data and algorithms to ensure benefits are accessible to all. For instance, algorithms trained on historical data might reproduce biases, like disproportionately showing STEM career ads to men . Transparency & Explainability AI operations should not be “black boxes". Transparency ensures people know when AI is being used. Explainability (especially in high-stakes fields like healthcare or finance) means humans must understand the AI's reasoning, even though modern models are often too complex for creators to fully explain. Accountability and Human Oversight Ultimate responsibility remains with humans ; it cannot be abdicated to algorithms. Developers or deployers must be accountable for outcomes, often by keeping humans “in the loop”. Mechanisms like responsible officers for AI systems or liability insurance are being explored by policymakers. Safety and “Do No Harm” AI must not pose undue risks, covering physical safety (e.g., autonomous vehicles) and psychological/social harm. A 2025 study found that AI therapy chatbots often breached professional ethics by mishandling users in crisis or reinforcing negative sentiments. Privacy and Data Protection Since AI relies on vast amounts of personal data, respecting privacy rights is a cornerstone. This involves obtaining data fairly, securing systems against breaches, and ensuring AI is not deployed for mass surveillance that violates civil liberties. Sustainability Ethical development calls for monitoring and mitigating AI’s environmental footprint, as training large models consumes significant energy and water . AI should align with broader social goals and be a positive force for future generations. Awareness and Education Many frameworks stress the need for AI literacy and public engagement . People must be informed about what AI is doing and empowered to question or challenge its use. Mandatory labeling of AI-generated content is also key. Practical Challenges in Implementing Ethical AI While the principles are clear, the reality of building and deploying ethical AI presents difficult technical and organisational hurdles: Algorithmic Bias and Inclusive Design AI systems often perpetuate or amplify biases present in their historical training data, leading to real harm like unfair targeting for police scrutiny or discrimination in hiring . Researchers must categorise and eliminate these biases, though experts acknowledge eliminating all bias is extremely difficult. Dr. Day offered two general tools for promoting inclusive design: Question Automation: Consider if the decision-making process should be automated at all. Humans, while biased, are arguably easier to hold accountable than machines. Community Involvement: Ask the people being evaluated or affected how they would like the system to operate. Dr. Day shared an inspiring example where a clinical geneticist, Dr. Karen Low, consulted the GenROC Consortium (parents of children with neurodevelopmental genetic conditions) on all machine learning research ideas, and the consortium was included as a co-author on resulting papers. The Black Box Problem and Accountability The complexity of deep learning networks often means neither the user nor the deployer can fully explain why an AI made a critical decision, undermining accountability and due process. Critical decisions affecting people’s lives demand explainability. This challenge forces a trade-off: researchers may need to sacrifice some model accuracy to gain interpretability. Furthermore, when AI causes harm, assigning liability is complex due to the long chain of contributors (data engineers, algorithm designers, users). The core message, however, remains: humans cannot abdicate responsibility to algorithms . The Impact of Financial Incentives A common recurring theme in ethical discussions is capitalism and financial incentives. Dr. Day notes that the “move fast and break stuff" culture means ethics often becomes a “necessary diversion" to the goal of profitability. Companies may even use ethical concerns strategically ; for example, large tech firms might advocate for regulations they already comply with, thereby restricting smaller companies trying to compete. Exploiting the Workforce While AI brings productivity gains, it threatens to displace many jobs. Anthropic’s CEO has predicted AI could replace up to 50% of entry-level white-collar roles in the next five years . Ethical development requires ensuring gains do not widen inequality. Possible remedies include investing in retraining and upskilling programs for impacted workers. Perhaps the most harrowing challenge revealed is the creation of unethical jobs used to train AI models. Dr. Day detailed the plight of Mechanical Turk workers (often outsourced to the Global South for cheaper labor) who perform menial data labeling, including reviewing and classifying deeply disturbing content to align large language models. In one dreadful anecdote from the book Empire of AI , a Turk worker in Western Kenya, whose job involved labeling sexual content, struggled severely. Months later, the very model he helped make safe (GPT-3) was released, leading to the disappearance of the writing contracts held by his brother — his sole source of support. The worker asked: “I'm very proud that I participated in that project to make chat GPT safe, but now the question I always ask myself, was my input worth what I received in return” . Dr. Day stressed that companies like Meta, offering signing bonuses in the tens of millions to AI engineers, “can probably afford to pay the data labelers more" . The Landscape of Governance and Regulation The global regulatory landscape is rapidly evolving, with different regions taking distinct approaches. European Union – The EU AI Act The EU has taken a pioneering role, developing the world’s first comprehensive AI law . It uses a risk-based approach (categorising AI systems as unacceptable, high, limited, or minimal risk). Unacceptable-Risk AI (Banned): Include AI systems for social scoring of individuals, the exploitation of vulnerable groups (like harmful AI toys), and intrusive real-time biometric surveillance in public. High-Risk AI: Systems used in critical infrastructure, medical devices, or employment decisions are not banned but are subject to strict oversight, including conformity assessments and human oversight. Limited-Risk AI: These are subject to less stringent transparency requirements: developers and deployers must make sure that end-users know they are interacting with AI (such as chatbots and deepfakes). Minimal-Risk AI: Are not regulated (including most AI applications currently on the EU single market, such as AI-powered video games and spam filters – at least as of 2021; this is evolving with generative AI). United Kingdom – A pro-innovation approach In contrast, the UK initially adopted a more flexible and sector-specific approach, encouraging existing regulators in finance, health, and transport to interpret core principles like safety and fairness. The rationale was to maintain “critical adaptability" and avoid overbearing legislation that might stifle innovation. However, the UK has recently proposed binding measures on those developing the most powerful AI models. United States The US currently has no single federal law dedicated to AI ethics . It relies on a patchwork of existing laws (e.g., anti-discrimination). In 2022, the White House released a non-binding “AI Bill of Rights” blueprint . All 50 states have introduced some form of AI-related legislation by 2025. Global Efforts International organisations like the OECD and UNESCO have established broad ethical principles. However, regulating AI remains fundamentally difficult because the core issue lies in auditing and validating whether principles like transparency and non-bias have actually been violated. Toward Responsible and Beneficial AI Building a responsible future with AI requires a multi-stakeholder effort involving technologists, governments, corporations, and civil society. Technologists must design AI with ethical considerations from the ground up. Dr. Day highlighted a good example of this: Te Hiku Media in Aotearoa New Zealand , which sought guidance and consent from Māori elders and the wider community before collecting data. In just 10 days, the community contributed over 300 hours of transcribed audio. Te Hiku governed the data with a Kaitiakitanga licence — keeping it for the benefit of Māori rather than open-sourcing it — showing how early stakeholder engagement and data sovereignty can align technology with communal goals. Crucially, the public has a massive role to play. We cannot trust providers to optimise for social harmony. AI business models are highly concentrated in just a few firms, often developed in teams lacking cultural diversity. As consumers, we must exercise our power: Vote with your feet: Be discerning about which AI chatbots, websites, and apps you use, supporting those with values you endorse. Increase AI Literacy: Public engagement and education are essential. Dr. Day emphasises that ethical discussions should be non-confrontational and inclusive, bringing people from different backgrounds together to share how data has affected them. Take Action: Write to your local MP to make your voice heard on AI safety regulation. As Kieren summarised, “Your understanding of AI rights shapes how policymakers act” . Further Reading Recommendations (from Dr. Huw Day): Weapons of Math Destruction by Cathy O'Neil: An extremely pertinent read documenting how algorithmic bias impacts everyday life. Data Feminism by Catherine D'Ignazio and Lauren Klein: Approaches data science from a social science lens, discussing diversity and the importance of community engagement. Empire of AI by Karen Hao: Documents the development of large language models from a non-technical perspective, including the politics, environmental impact, and the grim realities of Mechanical Turk workers. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

The technology world is currently dominated by one word: AI. From revolutionary large language models (LLMs) to the slightly less necessary AI-powered toothbrush, everyone is talking about AI-powered everything. But beneath the dazzling veneer of innovation lies a deep economic anxiety, symbolised by valuations that defy gravity. Currently, just one company, Nvidia, sits comfortably over a $4 trillion market capitalisation , dwarfing the entir e London Stock Exchange , which is valued at about $3.4 trillion. This striking disparity raises the core question: Is AI being massively overvalued, and are we witnessing an economic bubble about to burst? Defining the Bubble: When Hype Outpaces Value To determine if we are in an AI bubble, we must first understand what a financial bubble is! An economic bubble is defined as a period when asset prices surge far beyond their intrinsic value —that is, how much a company should actually be worth based on its sales and physical assets. One key metric used to gauge this overvaluation is the Price-to-Earnings (P/E) ratio , which compares a company's current share price to its earnings per share. A high P/E ratio suggests investors are willing to pay significantly more now, betting on enormous future revenue. For instance, Nvidia’s current P/E ratio is between 28 and 29 times its annual revenue . The Engines of Exuberance Bubbles are largely driven by exuberant investment behaviour , often sparked by a compelling real-world innovation or breakthrough. This behaviour is then fuelled by FOMO (fear of missing out) where investors worry about missing the train while everyone else profits. The danger arrives when speculation overtakes rational valuation , detaching the boom from actual economic value and leading to an overinflated, and inherently fragile, state. History teaches us that bubbles always pop sooner or later because the values are not grounded in anything real. When they burst, they trigger a sharp market crash, potentially causing businesses to fail, wiping out net worth, and in the worst cases, leading to a recession. Notable examples include the 1920s stock bubble leading to the Great Depression , and the 1990s dot-com bubble . Echoes of the Dot-Com Crash Today’s AI boom closely resembles the dot-com bubble of the 1990s. That era saw wild, rapid investment in internet startups, driving tech stocks to record highs before the bubble burst in 2000, wiping out tens of thousands of jobs and causing the NASDAQ to plunge nearly 80% from its peak . The similarities are striking: Revolutionary Technology, Irrational Hype: Just as the internet was game-changing, AI is a revolutionary technology promising AGI (Artificial General Intelligence) and massive productivity benefits. Yet, the hype has pushed current valuations to irrational heights. Startup Madness: Tiny AI startups, sometimes consisting of just three people and some AI powered idea, are securing millions in investment—a situation Open AI’s CEO Sam Altman called “insane" and “not rational behaviour". Extreme Concentration: Stock market gains are heavily concentrated. Seven companies (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla—the “Magnificent Seven") now account for 36% of the S&P 500’s total value , a 50-year record high concentration. Furthermore, AI-related stocks have accounted for 75% of S&P 500 returns since late 2022 . Revenue Complications: The bubble is further fuelled by Nvidia's monopoly over optimised GPU chips and roundtripping like behaviour among major players (like Oracle buying graphics cards from Nvidia, and Nvidia investing in OpenAI, who partners with Oracle). This complex exchange of billions makes true revenue incredibly hard to track. The Bear Case: Arguments for the Burst A compelling case exists suggesting the AI bubble is on shaky ground: Missing ROI: There is currently little financial evidence to justify the spending. A recent MIT report found that 95% of businesses that had integrated generative AI into their operations have not yet seen any return on investment . The $2 Trillion Problem: US tech firms must generate at least $2 trillion of additional annual AI revenue by 2030 to justify current capital spending—a massive amount that feels farfetched. Resource Bottlenecks: Scaling AI infrastructure requires immense resources. Electricity supplies are becoming a significant bottleneck, with utility companies squashing plans for new data centres. Already, 5-gigawatt data centres are being built, demanding energy equivalent to the power consumption of London . GPU Deterioration: These massive investments in data centres require constant upkeep. GPUs must be replaced every two to five years , adding continuous costs to infrastructure spending. Expert Alarm Bells: Warnings are coming from the highest financial authorities. The Bank of England warned this month that the risk of a sharp market correction driven by an AI stock slump could pose a “material risk" to the financial system. Jamie Dimon, CEO of JPMorgan, stated he is “far more worried than others," suggesting the burst could be six months or two years away. Even Sam Altman has warned that people will “over-invest and lose money" in this current phase. The Bull Case: Is AI Different This Time? Despite the dire warnings, some experts argue that the AI boom is just the beginning of a “super-cycle"—a multi-decade period of massive growth—not a typical bubble. Stronger Fundamentals: Unlike many dot-com startups, many leading AI companies (Nvidia, Microsoft, Google) are making substantial profits outside of AI, which they can reinvest to fund their AI development. This foundation means AI is grounded in stronger fundamentals than past bubbles. Exponentially Evolving Technology: The underlying technology of AI is “simultaneously increasing" in capability. While the infrastructure of the internet stayed largely the same, generative AI models are rapidly getting smarter, potentially allowing companies to “ride the wave" of fundamental technological improvement. Untapped Potential: Lisa Su, CEO of AMD, argues that skeptics are “thinking too small" and that AI's potential will spark a decade-long super-cycle transforming industries globally. We argued that society has not fully tapped into the potential of having a “mini Einstein in our pocket," partly due to the current lack of AI literacy. Humanoid Robots on the Horizon: A massive source of untapped return on investment could be humanoid robots. Companies like Tesla and Figure are developing advanced robots that will apply the current AI “brains" to physical labor, potentially entering the workplace within the next five years and delivering huge physical labor efficiency. Conclusion: The Long-Term View The jury remains out on whether the impending crash will be a swift, devastating bubble burst or merely a necessary correction in a longer-term super-cycle. If the bubble does burst, history suggests bad things will happen. A major crash could trigger financial instability, cause retirement funds (like those invested in the S&P 500) to lose value, and potentially increase inequality as the wealthy buy up devalued assets. However, the consensus is that AI technology itself will ultimately pay off and be a net positive for society in the long run, just as the internet has been. The question is whether the current hype and investment levels are sustainable until that long-term potential is realised. Perhaps a slowdown in development and spending would not be entirely negative. It would give society time to address major unsustainable issues arising from the rapid scaling of AI, including: The immense climate impact and reliance on fossil fuels for energy. Unresolved issues around data privacy and copyright . The fundamental lack of understanding regarding how these large models work and ensuring they are aligned with human values . While the immediate future holds volatility, the underlying technology promises transformation—it's just a matter of which companies, and how many investors, will survive the inevitable shakeout. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

In our latest episode, we dove deep into a surprisingly pressing issue that's making waves (pun intended!) in the tech world: AI's growing thirst for water . What started as a seemingly innocuous detail in a blog post by OpenAI CEO Sam Altman has quickly escalated into a global concern, prompting us to ask: Is AI really drinking all the water? How Much Water Are We Talking About? The discussion kicked off with Sam Altman's blog post , “The Gentle Singularity," where he mentioned that an average ChatGPT query uses about 0.34 watt-hours of electricity and a tiny one-fifteenth of a teaspoon of water . Sounds like nothing, right? But as we dug deeper, the numbers started to tell a different story. OpenAI reportedly receives around 2.5 billion queries every single day . If you multiply Altman's figure by this staggering number of queries, it equates to roughly 200,000 gallons of water per day, or nearly 1 million litres . However, this figure has been widely debated. Research papers suggest Altman's numbers might be significantly underestimated, possibly referring to smaller, older models or only a very narrow aspect of water use. For instance: A paper titled “Uncovering and Addressing the Secret Water Footprint of AI Models" found that 10 to 50 “medium-sized" queries (around a 100-word email) to GPT-3 evaporated approximately 500 millilitres of water . The Washington Post, recalculating with the more modern GPT-4, found it was 519 millilitres of water per single 100-word query . This is nearly half a litre per query, making Altman's initial estimate seem very low indeed. Even more concerning, the International Energy Agency (IEA) estimates that OpenAI's facility, using about 500 megawatts of power, could be consuming around 10 million litres of water a day . This means each prompt would equate to about 4 millilitres of water , nearly 10 times Altman's figure. Globally, data centres currently consume 560 billion litres of water per year , a number the IEA expects to more than double to 1,200 billion litres per year by 2030 . These are truly immense figures! Why Does AI Need Water Anyway? It might sound odd that computer systems require water, but there's a crucial reason: heat . https://www.sustainabilitymatters.net.au AI models run on incredibly powerful computers housed in massive data centres. Unlike Central Processing Units (CPUs), which are good for single, fast tasks, AI relies heavily on Graphics Processing Units (GPUs) . GPUs are designed for parallel processing, handling many tasks simultaneously, making them much more powerful but also causing them to get way hotter . Remember Sam Altman's X post about GPUs “melting"? That's why. To cool these hot GPUs, simple air fans are no longer enough. Data centres use sophisticated liquid coolant systems , and this is where water comes in. Crucially, this isn't just any water; it needs to be clean, fresh (drinking) water  to prevent issues like corrosion, salt buildup, or bacteria in the cooling pipes. The cooling process involves pipes circulating coolant over processors, which then transfers heat to water in a heat exchange unit. This hot water goes to cooling towers, where fans cause 80% of the water to evaporate  to dissipate the heat. This means fresh water is constantly lost and needs to be replaced. To put this into perspective, a typical 100-megawatt data centre guzzles about 2 million litres of water a day , enough for 6,500 US homes. And OpenAI is building a 5-gigawatt facility – 50 times that size , capable of powering the entirety of London or 4.5 million toasters! The Real Problem: Local Impact and Water Stress While the global numbers are staggering, the real concern lies in the local impact . Data centres are often built in areas already struggling with water scarcity: Two-thirds of all data centres commissioned or built in the United States since 2022 are located in areas of high or extremely high water stress zones . Just five states host 72% of these new data centres : California (17), Arizona (26), Texas (26), Illinois (23), and Virginia (a staggering 67 new data centres since 2022). Virginia is particularly popular because it's where much of the internet's original infrastructure was built, offering faster data speeds. This pattern is echoed globally, with tech companies establishing data centres in water-scarce countries like China, India, Saudi Arabia, and the UAE . This has led to protests worldwide , from the Netherlands and Uruguay to Chile , and even here in the UK, Anglian Water recently objected to a new AI data centre in North Lincolnshire due to insufficient water supply. The unfortunate truth is that water costs and scarcity are often a later consideration  when siting data centres, trumped by the price of energy and the need for high-speed internet infrastructure. So, What's the Solution? There are efforts underway to address this monumental challenge: Big Tech & Scientific Innovation: Most major tech companies have pledged to be water-neutral by 2030 , though specific plans are often vague. Some CEOs optimistically suggest that scaling AI will accelerate research into new clean energy sources or cooling mechanisms. More concretely, companies like Microsoft are investing in “zero-water" designs or closed-loop systems  that recycle all the water used for cooling, preventing evaporation. OpenAI is also incorporating these into new facilities. Amazon's AWS is exploring innovative methods like using treated sewage for cooling . There's even research into using the heat generated by data centres to warm homes in cooler climates. However, many of these solutions are still in the research or pilot phases, and for now, AI still relies on evaporating vast amounts of water. What You Can Do: A Three-Step Decision Plan for AI Use: We’ve put together a simple three-step decision framework to help you consider your own AI usage: The Social Value Test:  Before you send a prompt, ask yourself: Is this genuinely valuable to me or my community, or am I just trying to kill time?  Using AI for essential services or research is different from asking for a silly story when you could be doom-scrolling instead. The Right-Tool Test:  Do you actually need a giant cloud model for this task?  Many tasks can be done with a smaller, on-device model, traditional search, or even just your own brain, significantly reducing water and energy usage. We explore this in our “Island Explorer" exhibit at We The Curious , building intuition about the water use of different information sources. The Frequency & Batching Test:  Can you ask fewer, better questions?  Learning prompt engineering to be more efficient with your AI interactions can reduce the number of queries needed to get the desired result. The picture of AI's energy and water consumption is complex and goes much deeper than we could cover in one short episode. But it's clear that the environmental footprint of this rapidly advancing technology is a critical issue that demands our attention, from global policy to our everyday choices. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

This episode features an interview with Dr Jennifer Fleming, the coordinator of the Protein Data Bank in Europe (PDBe) and AlphaFold Protein Structure Database (AFDB) lead. AlphaFold is a program that predicts the structure of proteins. If you follow computational biology at all, or even if you don’t, you might well have heard of it – it won a Nobel Prize in 2024 for Google DeepMind co-founder Sir Demis Hassabis and chief Scientist John Jumper, the quickest ever awarded! It represents such a major breakthrough that it’s been said to usher in a whole new era of biology.   Listen to the episode  here , or read on for a breakdown of the topics covered, along with some extra background. by Charlie Harrison A Quick Bit of Background Proteins and the protein folding problem Proteins are the molecular machines that drive almost every process that happens in living cells. They are polymers, consisting of chains or sequences of amino acids. Proteins in all living things share a library of twenty different amino acids. A protein sequence can contain anything from a few dozen, up to thousands or even tens of thousands of amino acids. These sequences are manufactured in cells, and they fold up into intricate three-dimensional structures with very specific conformations of chemical and electromagnetic properties. These structures can behave in astonishingly intricate ways – forming highly selective pores in cell membranes that only let certain molecules pass, functioning as motors that propel through fluids, or binding to multiple molecules and bringing them close together in just the right formation to allow them to react. The precise structure of a protein determines its function, and it must be optimised for the type of environment in which the protein operates. Complex 3D shapes emerge from a string of amino acids ( Google DeepMind, 2020 ).   Protein folding is the process by which a string of amino acids adopts a three-dimensional structure. Because the structure of the protein is so precise, folding is critical, and it can and does go wrong in living cells all the time, sometimes with grave consequences. Single mutations in a gene sequence can change the sequence and thus the structure of a protein, in ways that are difficult to predict. A misfolded protein is at best unable to perform its function, and at worst can be toxic to the cell or the organism. Some proteins can misfold in a way that causes other copies of the same protein to also follow the same erroneous path. These forms, known as prions, aggregate into plaques that usually form in the nervous system and brain, and are associated with many neurodegenerative diseases with no known cure, including Alzheimer’s and Parkinson’s disease.   Understanding the process and the outcomes of protein folding is a central problem in biology, but determining the structure of a protein experimentally is very difficult – much harder than determining the sequence. Structural determination relies on X-ray crystallography, where X-ray light is shone through protein crystals. By examining the diffraction patterns, it’s possible to determine the symmetries in the crystal lattice, and work back from there to calculate the structure of the individual proteins in the crystal. These calculations are complicated, but the hardest part is purifying and crystallising the protein in the first place. Forming the crystal lattices takes a lot of trial and error, varying solvents and precipitating agents, pH, temperature, the presence of so-called chaperone molecules, and sometimes even portions of the protein sequence deemed unimportant to its function and likely to impede crystallisation  ( McPherson et al., 2014 ) . In comparison, it’s relatively easy to determine the sequence of a protein, through mass spectrometry or  Edman degradation . As an analogy, it’s easy to knock down a house and figure out what materials it was made of, but it’s very difficult to look at a pile of bricks, mortar, insulation, wiring and plumbing, windows and doorframes, and work out what the house is going to look like. The protein structure prediction problem is the key to the inner workings of biological systems. CASP and DeepMind The CASP competition was founded in 1994. CASP – Critical Assessment of Structure Prediction – sought to formalise the process of structure prediction, to rigorously compare the performance of different methods and algorithms and accelerate research into this key bioscience problem. The CASP team worked with research groups around the world to obtain newly discovered protein structures, before their publication. The corresponding sequences were given to competitors, and their predictions were assessed against the known structures using a specially developed metric called the Global Distance Test (GDT). GDT ranges from 0-100, and scores above 90 were considered to be about as accurate as experimentally determined structures, accounting for variation in the crystallography process. CASP competitions run every two years, and until the late 2010s, the best results from predictions in competition scored GDTs of around 40.   Enter DeepMind, under the leadership of former chess prodigy Demis Hassabis. Founded in 2010 and acquired by Google in 2014, DeepMind’s first high-profile successes were in building AI models to play board games. Chess bots had dominated against human players since the late 90s, but DeepMind’s AlphaGo became the first computer program to defeat the world champion human player in the even more complex game of  Go . DeepMind then developed a more general bot, AlphaZero, which learned to play multiple games using pure reinforcement learning and defeated the reigning champion AI engines in chess,  shogi , and Go. DeepMind applied its expertise to the protein folding problem, and in 2018, it entered AlphaFold into CASP-13, winning easily and achieving new benchmark GDT scores of around 60. In CASP-14 in 2020, AlphaFold2 built on this success, with GDT scores close to 90, prompting the organisers of CASP to declare that : “The problem has been largely solved for single proteins." How AlphaFold Works The first edition of AlphaFold took a fairly generic deep learning architecture and applied it to the protein folding problem. AlphaFold2 was fully redesigned from end to end, purpose-built to solve the protein folding problem. There were two keys to the breakthrough. One was in hardware, with tensor processing units (TPUs) providing enormous parallel computing power. The other was a novel neural network architecture called Evoformer, which leverages the evolutionary information encoded in related proteins taken from different species by building a multiple sequence alignment (MSA) based on the input protein sequence. High-level overview of how AlphaFold2 predicts a protein’s structure from its amino acid sequence ( EMBL-EBI, 2025 ).   An MSA is a matrix-like structure containing a set of similar sequences, one per row. The similar sequences are identified by searching a database, and gaps are inserted in the rows where necessary to match up the properties of the amino acids in each column as closely as possible. Related proteins from different organisms, sharing similar sequences, also likely share an evolutionary history, and similar structures and functions too. These related proteins can be viewed as viable examples of variation that preserve the key elements of a structure, allowing it to fold correctly and perform its function. An MSA shows sets of positions in the sequence that tend to vary together – that is, a change in one position is likely matched with a corresponding change in another position. This covariance implies that the amino acids at those positions in the sequence are likely to be closely located in the folded protein. AlphaFold2’s Evoformer network uses an MSA to build a representation of the relative positions of every pair of amino acids in the input sequence. It then iteratively updates both the MSA and the pair representation, with information flowing in both directions, eventually arriving at a final prediction of the protein structure. Predictions are accompanied by confidence scores at the local and global levels.   Another important technical development was self-distillation. The Protein Data Bank (PDB) contains about 150,000 known structures, which is a lot, but not enough to saturate the learning potential of an AI model. AlphaFold was trained on the known structures and then used to make predictions for unknown structures. Predictions with high confidence scores were fed back into the training process and used to further enhance the accuracy of the model.   AlphaFold3 was a further extension to AlphaFold2, gaining the ability to predict interactions between multiple molecules. This allows it to predict the structures of multi-chain proteins, as well as ligand and RNA binding. The Impact of AlphaFold AlphaFold is a computationally intensive model and requires significant resources to run. To help disseminate the benefits of AlphaFold’s predictions, Google teamed up with the EBI to create  AlphaFold DB . Scientists can now access predicted structures for over 200 million protein sequences, the majority of the sequences in  UniProt . The predictions are free to use, and can be browsed, searched, and visualised online in the interactive portal or downloaded for further use.   AlphaFold DB is managed by the EBI, which is also home to the European Protein Data Bank (PDBe) containing experimentally-determined structures, and the accompanying knowledge base (PDB-KB), which contains rich annotations on how proteins function and interact with other molecules. This colocation allows for maximal knowledge exchange between the different, complementary databases. PDB is over 50 years old, and the data is very high-quality, very well curated and labelled, which was essential for the success of AlphaFold. AlphaFold DB uses the same file formats as PDB, so the structural data is more or less interchangeable – though experimentally determined structures are accompanied by additional data like electron densities or diffraction patterns.   By 2023, AlphaFoldDB had over 2 million users in 190 countries. Open-sourcing the predictions has saved a huge amount of computational and research effort. All EBI data is released under CC-BY 4.0, which allows freedom of use with attribution, to maximise access. AlphaFold3 caused some controversy because the code and the model itself were initially not open-sourced, though they were subsequently shared with the scientific community after protestations. However, the multi-molecule interaction predictions from AlphaFold3 are released under a more restrictive licence than its predecessor, one which prohibits commercial activity. In 2021, DeepMind’s parent company, Alphabet, established a new spin-off called Isomorphic Labs with the aim of building on AlphaFold3 and applying it to the expensive enterprise of drug discovery.   Drug discovery and design is one of the areas of research where protein structure prediction shows the most potential, with the promise of speeding up development pipelines, reducing failure rates in clinical trials, and even personalising drugs to the needs of an individual. Thinking more broadly, if the goal of functional protein design is fully realised, the possibilities are almost endless – imagine enzymes that can break down plastics, produce biofuels, or make crops more resilient. This requires iterative pipelines where proteins are sequenced and their structures and functions are tested. Performing these steps  in silico  can make the impossible feasible.   In the meantime, the AlphaFold DB team are working hard to make the resource as valuable as possible. There is a lot of potential for experimental data to improve the model, so they plan to add a feedback mechanism for continuous updates. They plan to add as many new features as possible, like linking to related resources to add context, and providing ways to compare structures, while making sure that every additional piece of information is high-quality and readily understandable. Most importantly, they want to involve as many people as possible. The real impacts of this work will continue to be felt for decades to come, and it’s the wider community of researchers that will realise it. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

How AI Understands Your DNA, featuring an interview with Professor Ewan Birney. Prof Birney is Interim Executive Director of the European Molecular Biology Laboratory ( EMBL ) and Non-Executive Director of Genomics England , and played a pivotal role in the Human Genome Project . Read on for some selected highlights from the new episode, along with some extra background. Catch up on the first episode by listening  to or  reading about it . by Charlie Harrison Episode highlights DNA is like a massive book, written in a foreign language . It also has lots of obscure passages that we won’t really understand. To use Leo Tolstoy's book, War and Peace, as an analogy, do we really need all that stuff about farming in 19th-century Russia to explore the themes of love, war, and the human condition? Before the human genome project was completed, Prof Birney ran a sweepstake  called  GeneSweep , taking guesses on how many protein-coding genes would be found. Most guesses ranged from 40k to 100k, which seemed reasonable given that a nematode worm (the first multicellular organism whose genome was sequenced in full, completed in 1998) has around 20k. It turns out that the actual number in humans is also approximately 20k, far fewer than most thought. However, the actual number of genes is difficult to define precisely because, like everything in biology, transcription is complicated! For example, cells perform alternative splicing of transcripts and post-translational modifications, genomes contain overlapping reading frames and bidirectional genes, and a large number of transcribed regions are functional but don’t code for proteins. Protein-coding regions make up only around 1.5% of the human genome.  The human genome has about 3 billion base pairs. Compare this to the nematode worm, which has about 100 million base pairs for the same number of genes! Much of the rest consists of transposons and retrotransposons – regions of the genome that copy and insert themselves into different positions within a genome, either directly or via an RNA intermediate. These act like genetic parasites, bloating the genome and imposing a metabolic burden on the cell, which has mechanisms to try to repress them; but they also play an important role in evolution. Protein-coding regions are of less primary importance than expected.  The non-coding regions of the genome are far from inert; in fact, they’re very active. Many regions that are completely dissociated from protein-coding genes are involved in protein binding or RNA synthesis. Structural changes to genome packing trigger changes in gene regulation that can lead to dramatic outcomes, including cancer. We used to think that the genome behaved like a mould, or a blueprint  – giving complete and explicit instructions to make a perfect replica. Now we know that it’s more like a script – there’s lots of room for interpretation, and context is important. The Human Genome Project was a landmark achievement in the open data movement , thanks to the principles enshrined in the Bermuda Accord , which stated that human genome data would be released into the public domain within 24 hours of its production and would not be held under licence by private companies. The contemporary landscape is more varied. Large private companies play important collaborative roles in major bioinformatics projects, but they also have their own agendas and can’t be forced to share everything. In AI, “open” can mean different things – pseudocode, an open-source repository, or a fully trained model with weights. The cost of sequencing a human genome has reduced from around $3 billion  for the first ever completed in 2003, to around $200 in 2022 – a five-million-fold reduction. Large modern facilities can sequence 20 thousand human genomes per year, and the UK Biobank holds more than 500k complete genomes along with health data. This is made possible by improvements in technology, but also by the existence of the reference genome, which makes it simpler to assemble sequence data into a coherent structure. A single human genome sequence stored in plain text or FASTA format would take up around 3 GB . It can be stored in a more compact TwoBit format, which represents nucleotide bases in binary (T as 00, C as 01, A as 10, and G as 11), taking up just 800 MB. However, the  full human reference genome  contains much more information than just the basic sequence; there is structural information about how the sequence was assembled, which chromosome each element belongs to, annotations about genes and RNA coding sequences and the surrounding regions. There are also versions of the genome with highly repetitive sequences hidden, to avoid biasing search algorithms (this is called “ masking “), and versions in multiple file formats. Many of these files can be compressed, which is often highly effective because the sequences contain many repeating regions. Taken altogether, the reference genome can be downloaded to about 18 GB on disk. The breakthrough in genomic sequencing is multifaceted . Improvements in algorithms, a drop in sequencing cost leading to greater availability of data, and an increase in computing power all played crucial roles Modern AI research is very experimental.  There is a lot of trial and error; trying to find a model that works well, then later trying to explain why it’s so effective. This is a paradigm shift compared with earlier techniques like Bayesian modelling. Deep learning models can be very counterintuitive, but they can be interrogated. Prof Birney cites three concerns about AI : Hype . To many people, “AI” now means “chatbots”, and these are both impressive and disappointing. There is a worry that disillusionment could overshadow some of the amazing achievements, like AlphaFold, advances in image analysis, and genome decoding. Reliability . In medicine especially, reliability is crucial. How do we avoid weird behaviours, like the hallucinations that ChatGPT and other LLMs sometimes display? One way is to put a human in the loop, but this has its own problems, like preventing boredom – especially when the AI is right most of the time. In  airport security systems ,  computer-generated images of banned items are occasionally projected onto luggage scans to make sure staff remain alert, so maybe a similar method could work in medicine. Bioweapons . There is the potential for AI to invent new toxins or viruses that pose a real danger to humanity. This is not completely new, as lots of dangerous chemicals and organisms exist in nature, but AI changes the shape of the problem. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

Welcome to the cutting edge of science! For 20 episodes, we’ve explored the fascinating world of AI, but today, we’re diving into an area that’s not just important but rapidly developing: the intersection of Artificial Intelligence and Biology. This isn’t just a niche topic; it’s a field where incredible research is emerging, potentially bringing once-futuristic concepts into reality. Imagine curing genetic diseases in children, turning leftover food into car fuel, or creating plastic bottles that biodegrade on their own. All these possibilities stem from the integration of AI and biology, offering a path to reprogram genetic code (DNA) as easily and cheaply as editing computer software. by Charlie Harrison Why AI and Biology Are a Perfect Match The sheer scale and structure of biological data are too vast for human comprehension. A single human genome contains  three billion bases  – the building blocks of DNA – and a human has  30 trillion cells . If you unraveled all the DNA in your body and laid it end to end, it would stretch from  the sun to Pluto and back 17 times . This truly is a “big data set”, and AI is uniquely capable of navigating this immense scale of information. Revolutionary Applications of AI in Biology The impact of AI in biology is already profound, transforming various sectors: Healthcare and Drug Design: AI excels at digesting and analysing enormous amounts of health data to find patterns. During the COVID-19 pandemic, Pfizer used AI and machine learning to reduce clinical trial data analysis time  from months to just 22 hours , significantly speeding up vaccine development. AI is now being applied much earlier in the drug development pipeline to generate and screen potential new drugs and drug combinations before laboratory testing or clinical trials, saving immense amounts time and resources. A company called Exscientia (now Recursion), which prioritises AI in drug development, managed to reduce their exploratory research phase for an immuno-oncology drug  from four and a half years to just 12 months . Digital twins promise to  personalise medicine , producing tailored treatments optimised for an individual rather than a population. Tackling Superbugs: AI is not just speeding up existing processes; it’s enabling entirely new discoveries. In May 2023, scientists used AI to discover a brand new class of antibiotics capable of killing deadly antibiotic-resistant bacteria.  These antibiotics were generated “from scratch” by AI , in ways that conventional biologists might not have considered. Environmental Sector and Sustainability: AI has been used to  generate novel plastic-eating enzymes  by predicting optimal adaptations and mutations to existing natural enzymes. This process is akin to AI “speeding up evolution,” finding the most effective adaptations for consuming plastic. AI is helping to predict how to design and genetically engineer living organisms (like bacteria or enzymes) to sustainably produce biomaterials, biofuels, and biopolymers. This includes concepts like using leftovers to fuel cars, creating sustainable plastics, or capturing carbon dioxide from the atmosphere to converting greenhouse gas emissions into valuable raw materials. The Journey: How AI Met Biology The integration of AI and biology wasn’t always obvious. Historically, these were separate academic fields. The timeline of their convergence can be split into four key stages: Before Computers: Early applications involved using mathematical equations to model biological processes, such as population dynamics (e.g., rabbit and fox populations) and the formation of spotted and striped patterns through diffusion, demonstrating that some biological phenomena could be explained mathematically. The Arrival of Computers: This era was marked by the creation of large biological datasets. The Genbank database, a “Google for genetic information,” allowed for the digital storage of vast genetic sequences that wouldn’t fit on paper. Simultaneously, the development of high-throughput robotics and measurement techniques in laboratories automated previously painstaking experiments, leading to an explosion of data that could be stored in databases like Genbank. The Machine Learning Era: With massive datasets available, research groups began applying machine learning, which naturally excels at finding patterns in large data. The most revolutionary application came with Google DeepMind’s AlphaFold. Proteins are life’s fundamental building blocks, and their 3D shape dictates their function. AlphaFold can take a protein sequence and predict its complex 3D structure, a task that would previously have absorbed an entire PhD project. AlphaFold has now published over 200 million protein structures, saving billions of research hours. The Generative AI Era (Today): Building on predictive models like AlphaFold, generative AI models like Evo can now create entirely new biological entities. For example, they can generate new protein designs, which AlphaFold can then evaluate for feasibility and properties. This combination of generation and prediction will massively accelerate the exploration of different biological structures, leading to the creation of new drugs, plastic-eating enzymes, and lab-grown foods. Ethical Considerations: High Stakes and Future Challenges While the potential benefits are immense, the intersection of AI and biology comes with significant ethical considerations. The stakes are incredibly high, especially when dealing with global problems like pandemics and diseases. Dual Use Technology: Any technology with such transformative potential also carries a risk of harm. AI in biology is a “dual use” technology: tools that can save lives could also be used to create bioweapons or dangerous pathogens with no known cures. This concern about AI for bioweapons is often overlooked in broader discussions about AI’s existential threats. AI Interpretability: A challenging aspect is that we often don’t know exactly what complex AI models like AlphaFold have learned. There’s a trade-off: giving AI the flexibility to understand patterns beyond human comprehension often makes the outputs less understandable to us. While methods like “mechanistic interpretability” are being developed, it remains a difficult challenge. Public Discourse and Privacy: There’s an urgent need for more public, governmental, and scientific discourse on these applications. A major concern is the use of human medical data and genomes to train AI, raising significant privacy issues. The Future: Utopia or Inequality? Looking ahead, AI in biology envisions some truly transformative futures: Precision Medicine: This concept aims for an AI model of each individual’s cells and genome, allowing for customised treatments tailored to a person’s specific biology. This could lead to higher efficacy and fewer side-effects compared to current general treatments that are developed over populations. Harnessing Biology for Creation: The ultimate goal is to understand and manipulate biology at every level – from reactions to protein folding – to produce new things and grow cures. This could mean growing your own medicine at home or even producing food. The hope is that this capability could make things more abundant and accessible, democratising production that was previously expensive and resource-intensive. However, a critical question remains: will these amazing benefits exacerbate inequality? There’s concern that high costs could make technologies precision medicine only accessible to the wealthy, creating a genetic lottery where the rich have an advantage. While some applications, like sustainable fuel from leftovers, might naturally become cheaper and more accessible, the commercialisation and gatekeeping of medical innovations, particularly within privatised healthcare systems, present a significant challenge that will require careful consideration and systemic changes. This first episode of the AIBIO-UK mini-series offers a comprehensive overview of how far we’ve come – from “pencil and paper” biological models to the generative AI that is reshaping life itself. Stay tuned for future episodes that will dive deeper into specific topics, including how AI decodes genes, the full story of AlphaFold, the future of AI pharmacists, AI’s role in saving our planet, and the crucial ethical path forward. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

In our latest episode, Kieren and Riku tackled a monumental question that's on everyone's minds: Are AI chatbots, like the ubiquitous ChatGPT, making us stupid? As the usage of these tools skyrockets among knowledge workers and students alike, it's crucial to understand their long-term impacts. The question of technology's impact on human intelligence isn't new; historical parallels abound, but AI's unique capabilities present genuinely novel challenges. A Brief History of “Tech Panic" To frame the current debate, we first explored cognitive offloading : the act of using external tools or technology to reduce mental effort in completing tasks. This isn't a modern phenomenon; it's a concept that dates back 2,500 years. Socrates and Writing: Ancient Greek philosopher Socrates famously expressed concerns that writing would lead to “forgetfulness" and an “appearance of wisdom, not true wisdom," as people would rely on external records rather than exercising their memory. The Printing Press: Similar fears emerged with the printing press, yet studies have shown that reading books can actually strengthen memory. Calculators: The advent of calculators sparked worries that children would never learn basic arithmetic, but instead, they often foster a more positive attitude toward math, allowing focus on concepts rather than rote calculations. Computers and GPS: More recently, typing on computers has led to a decline in handwriting skills (notably in Japan with complex characters), and GPS usage has been shown to reduce spatial memory, with studies comparing London cab drivers (who have larger hippocampi due to navigation) to bus drivers. The key difference with AI chatbots is their breadth of tasks and their ability to reason about problems and bring together disparate concepts, unlike previous tools that offered specific inputs and outputs. This leads to the concern that AI encourages a new level of “mental heavy lifting" to be offloaded. The Rise of Metacognitive Laziness This brings us to the core concept of metacognitive laziness : our human tendency to do the bare minimum, combined with AI's broad capabilities. Metacognition is “thinking about thinking," and the worry is that AI erodes critical thinking skills and our ability to process information and draw our own conclusions. Impact on Education Several recent studies shed light on this issue within educational settings: Brain activity patterns while writing essays with different tools: using ChatGPT (LLM), using a search engine, or relying only on one’s own brain. The figure shows how strongly different brain areas worked together in each case. The asterisks mark where the differences between groups were meaningful, ranging from small (*) to very strong (***) ( Kosmyna et al., 2025 ) . “ Does ChatGPT enhance student learning? ” (2024 Meta-Analysis) : This review of 69 studies found that while AI improved students' academic performance, motivation, and even higher-order thinking skills, it significantly reduced the mental effort students exerted during learning tasks . The question then becomes: are students truly mastering material, or just getting correct answers with less effort? “ Generative AI Can Harm Learning " (2024 Study): A high school experiment with GPT-4 showed that while AI improved performance on homework (48% for standard ChatGPT, 127% for a “GPT Tutor" designed to prompt thinking), students performed worse when AI access was taken away (17% reduction). This suggests students used AI as a “crutch," hindering real learning. “ Your Brain on ChatGPT " (2025 MIT Study ): This study used EEG to measure brain activity during essay writing. It found that participants using powerful AI tools showed significantly lower brain activation and “under-engagement" of neural networks. Furthermore, the LLM group had reduced memory for their own words just minutes after writing, a phenomenon dubbed “cognitive debt". “ How university students use Claude " (2025 Anthropic Education Report): Analysing 1 million student conversations with Claude, this study revealed that students overwhelmingly used AI for higher-order thinking tasks like “creating" (40% of requests) and “analysing" (30%), rather than simple fact recall. This raises the concern that if we offload the most cognitively demanding tasks, what are we left to do ourselves? The summary of these studies is clear: true learning is difficult, and AI making things too easy can prevent the necessary effort for retention. As one researcher put it: “All animals are under stringent selection pressure to be as stupid as they can get away with". Impact on Professional Productivity While less studied, the professional environment also shows interesting trends: METR Field Experiment (2025 Study): In a study with experienced computer programmers, developers predicted AI would speed up their tasks by 20-30% and estimated a 20% speed-up after use. However, actual timing revealed that AI slowed them down by 19% . The primary issue was that time spent correcting and verifying AI output outweighed any gains . This challenges the common belief that AI is a massive productivity booster, especially for complex tasks like coding, where an error can propagate throughout the entire solution. A chilling anecdote was shared about an AI chatbot erasing an entire company database it was meant to manage, highlighting the risks of blind reliance. Is Cognitive Offloading Always Negative? Despite the concerns, the answer to whether offloading is inherently negative is nuanced. There's a growing societal stigma against using AI tools, but we must consider potential benefits: Cognitive Bandwidth: Offloading routine tasks can free up mental resources for more creative thinking, planning, or complex reasoning. Error Reduction: AI can automate error-prone, low-level tasks, improving accuracy. Accessibility & Equity: Tools like spell-checkers or screen-readers, and potentially AI, can level the playing field for individuals with learning disabilities or other disadvantages, providing more equitable access to educational outcomes. Extended Cognition: Our minds, combined with tools like diaries, to-do apps, or AI chatbots, can function as “second brains," forming an integrated “mind+tool system" that enhances overall capabilities. The “walking stick" analogy illustrates this: does a walking stick become part of a person with a bad leg, effectively extending their ability to walk? Similarly, are smartphones or AI chatbots becoming “walking sticks" for our cognition? However, the costs include skill decay (if we don't “use it or lose it"), over-trust and complacency (leading to blindness to real-world hazards), and surface understanding (if offloading occurs before a mental model is formed). The key takeaway is the need for meta-cognition or meta-learning —learning how to learn. We need to develop an intuition for which “cognitive muscles" are exercised by different tasks and consciously decide what skills we want to develop versus what we are willing to offload. Just like a gym-goer chooses which muscles to train, we should be intentional about our cognitive workouts. How Should We Use Chatbots? Practical Tips for the Future Given these insights, how should we integrate AI into our lives responsibly? Age Matters: The impact of AI shortcuts is more significant during formative years (e.g., teens risk missing foundational skills), while for university students, it might limit higher-order analysis and creativity. Professionals risk skill stagnation. AI as a Partner, Not a Replacement: Use AI for practice and feedback , not as a “one-click answer engine". Instead of asking AI to write an entire essay, prompt it for resources or different perspectives to facilitate your own critical thinking. Be a Critical Evaluator: Always critique AI outputs , especially in fields where you have expertise. Understand that AI can hallucinate, and don't just “copy and paste". Proofread and rewrite extensively. Implement Metacognitive Prompts: External Cues: Tools or self-reminders to pause and reflect. Questions like “How closely does the response align with what you expected?" or “What perspectives might you be missing?" can encourage deeper engagement. System-Level Prompts: Many chatbots allow you to set initial instructions that guide every response. You can tell the AI not to give direct answers but to guide you towards them, fostering your own critical thinking. Frameworks for AI Tutors: Developers building AI for educational settings, especially for children, should design them with frameworks like Zimmerman's Self-Regulated Learning (SRL). These chatbots would prompt users to plan learning goals, offer personalised feedback, and encourage reflection on the learning process, supporting metacognition rather than just providing answers. Final Thoughts: Redefining “Smartness" in the AI Era As we move forward, society might need to critically re-evaluate what it means to be “smart". Is it about un-aided knowledge recall, or is it more about creativity, charisma, and the ability to formulate compelling ideas, leveraging tools effectively? AI models are still in their infancy, and the current “apply AI to everything" phase might subside as we better understand their limitations and optimal use cases. The ultimate message is “use it or lose it" . Be mindful of what skills you're exercising and which ones you're comfortable offloading. We must learn from past technological adoptions, like social media, where we didn't fully consider the long-term negative consequences of techno-optimism. Generative AI's impact could be even more profound due to its rapid and widespread adoption. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

In our latest episode, we tackled the pressing topic of deepfakes. This technology, which uses artificial intelligence to create hyper-realistic digital forgeries, is evolving rapidly and raising significant questions about trust and reality. While media coverage on deepfakes might seem to have quietened down, they are actually on the rise. Let’s unpack what that means. What is a Deepfake? The term “deepfake” originates from the deep-learning algorithms that power this technology. These algorithms, a form of generative AI, are designed to replicate and produce unreal versions of real things. A major concern is that deepfakes can be used to steal online identities and erode trust in digital media, including text, voices, videos, and images. The age-old adage “seeing is believing” no longer holds true. How are Deepfakes Made? We discussed the two primary AI technologies used to create deepfakes: Encoder-Decoders : These AI systems are trained on thousands of images or video frames of a person’s face. They learn to compress and reconstruct the face, enabling them to swap one face onto another in videos. This technique is widely used for face swapping. Generative Adversarial Networks (GANs) : GANs involve two AI systems in competition with one another—a generator (forger) and a discriminator (detective). The generator creates fake images, while the discriminator tries to identify whether they are real or fake. Through this iterative process, both systems improve, producing increasingly convincing forgeries. This method is now the leading approach for deepfakes. The Rise of Deepfakes: A Brief History Deepfakes first surfaced in 2017 on Reddit, gaining attention through videos of celebrities and politicians saying or doing things they never actually did. For example, there was a deepfake of Barack Obama making inappropriate comments about Donald Trump, and one of Mark Zuckerberg claiming to control the world. Tools like DeepNude, which can remove clothing from photos, further highlighted the dangers. More recently, a deepfake of the Pope in a puffer jacket fooled many people. This shows not only how advanced the technology has become, but also how adept people are at using it to craft convincing narratives. The Bad and the Ugly Deepfakes have been exploited for malicious purposes, including: Scams : Deepfake voices have been used to deceive individuals into transferring large sums of money, with one case resulting in a $35 million loss . Misinformation : Deepfake images, such as those of Donald Trump being arrested, have blurred the line between reality and fiction. This erosion of trust can make people desensitised to future events. Legal Manipulation : In child custody cases, deepfakes have been used to fabricate recordings of parents behaving abusively . Pornography : A staggering 98% of deepfake videos are pornographic , with 99% of them targeting women, particularly celebrities. Shockingly, it now takes less than 25 minutes and less than £1 to create a 60-second deepfake pornographic video. The Good Side of Deepfakes While the risks are undeniable, deepfake technology also has positive applications: Voice Restoration : People with conditions like ALS can use AI to recover their voices by training systems on past audio clips. Awareness Campaigns : Deepfakes have been used to highlight important causes, such as malaria prevention, with David Beckham speaking in 27 different languages. Challenging Bias : Deepfakes have been used to combat gender bias, such as mapping male faces onto female football players to showcase the excitement of women’s football. Entertainment : Deepfakes allow actors’ voices to be translated into multiple languages, making films and series more accessible globally. Detecting Deepfakes As deepfake technology advances, so do detection methods: AI Detection Tools : Companies are developing AI tools to identify deepfakes. For example, Intel’s real-time detector claims 96% accuracy , though its performance in real-world settings may be lower. Other techniques analyse blood flow in pixel data. Watermarking : Invisible digital watermarks can be embedded into images, detectable only by AI. Human Observation : Signs like blurriness, inconsistent lighting in the eyes, or unusual hand details (e.g., missing fingers) can indicate deepfakes. Source Verification : Always check the source of an image or video. Reverse image searches on platforms like Google can help identify original content. Legal and Social Implications Governments and social media platforms are taking steps to regulate deepfakes: Regulations : China mandates clear labelling of deepfake content, and the EU is working on similar policies. Criminalisation : Sharing intimate deepfake images is now a crime in the UK and some US states, like Virginia. Social Media Guidelines : Platforms such as YouTube, Instagram, and Reddit prohibit posting deepfakes without proper labelling. Hollywood Strikes : Recent strikes highlighted actors’ concerns over deepfakes being used to replicate their likenesses without consent. The Future of Deepfakes The technology is advancing rapidly, with key trends including: Text-to-Video : AI is developing the ability to generate full videos from text prompts. Although current outputs are short, this is expected to improve. Synthetic Content : By 2026, experts predict that 90% of new online content will be synthetically generated . What Can We Do? To address the rise of deepfakes, we need to act now: Educate Yourself : Learn how the technology works and how to spot deepfakes. Critical Consumption : Assess the content you encounter online and verify sources before sharing. Be Alert : Recognise the types of content that might convince you most, as these are often targeted. Support Policy : Advocate for local governments to introduce legislation addressing deepfakes. This is an ever-evolving landscape, and we’ll continue to cover these developments in future episodes. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!

In this episode, we dive deep into a topic we've been eager to cover since the podcast began: The AI Arms Race . What do nuclear weapons, moon landings, and Artificial General Intelligence (AGI) all have in common? They are all results of intense, global arms races, and the AI arms race might just be the most consequential one yet. A Look Back at History An “ arms race" is traditionally defined as a competition between nations for superiority in the development and accumulation of weapons . The term originated from the “ Dreadnought race" in the early 20th century, a naval arms race between Britain and Germany, where the development of revolutionary battleships like the HMS Dreadnought kicked off a rapid escalation that significantly contributed to the tensions leading to World War I. More generally, an arms race is a competitive dynamic where parties continuously bolster their capabilities in response to perceived or actual enhancements by rivals . Past examples include: The Nuclear Arms Race (1940s-1991) : Following the US detonation of the first atomic bombs in 1945, a fierce rivalry ignited with the Soviet Union, especially after their nuclear test in 1949 and the development of hydrogen bombs. This race brought the world to the brink of nuclear disaster, notably during the Cuban Missile Crisis, highlighting that treaties and safety procedures were only established after  near-catastrophe. The Space Race (1957-1975) : A Cold War offshoot aimed at conquering space, this began with the Soviet Sputnik satellite and culminated in the American moon landing. A key concept for understanding these dynamics is Game Theory , a mathematical area for analysing strategic interactions where each player's best move depends on what others are doing. In an arms race, companies and nation-states choose their R&D speed, openness, and safety budgets relative to rivals. The episode introduces the concept of a Multipolar Nash Equilibrium , a situation where no single player can improve their outcome by changing their current course . In the nuclear standoff, the “ Mutually Assured Destruction" (MAD) theory created an equilibrium where no one used nukes because retaliation was guaranteed. However, in the AI race, this could be detrimental: if one company prioritises safety and slows down, it risks being left behind, forcing others to continue racing for dominance. The AI Arms Race: Key Moments and Similarities The AI arms race is currently unfolding. Key moments include: AlphaGo beats Lee Sedol (2016) : Google's AI system beat the world champion at the board game Go, a “ Sputnik moment for AI" that prompted significant AI investment, especially in China. ChatGPT Viral Launch (2022) : OpenAI's release highlighted generative AI's potential and spurred other tech giants like Google, Meta, and Amazon to race to catch up. DeepSeek-V3 (2024) : This Chinese company's release of an AI model rivalling ChatGPT at a fraction of the cost changed dynamics, proving that frontier-level reasoning could be achieved cheaply and openly, and casting doubt on the valuation of Western companies. The AI arms race shares some high-level similarities with past races, primarily the human desire to ‘play God' and master life or existence . The moon race was about conquering the physical universe, the nuclear race about conquering death, and the AGI race is about conquering life and consciousness. Strategically, they all involve a first-mover lock-in  (setting global standards and rules) and act as a dual-use engine , spurring innovations (like nuclear energy from the atomic bomb or satellites/GPS from the space race) far beyond their initial motivation. Why the AI Arms Race is Different (and More Concerning) Despite similarities, the AI arms race is distinct in critical ways: Intangible Weapons : Unlike missiles, AI consists of algorithms, data, and compute, making the threat less clear but far more pervasive and widespread . Private-Sector Lead : Start-ups and Big Tech outspend governments, leading to faster R&D and less regulation . This makes it profit-driven, with companies trying to “lock you in" to their products. Speed of Iteration : AI models improve at an unprecedented rate, often referred to as “double exponential growth,"  with capabilities doubling every year. This pace means no regulatory or safety framework is currently equipped to deal with it. Defining Humanity's Interaction : It's not just a race to build AI, but also to define how humanity perceives and interacts with it . Companies may prioritise persuading users and gaining data over purely beneficial innovation. Who's Winning the AI Race? Determining the “best" chatbot is complex, as it depends on desired capabilities like: Stylistic control and core language generation Knowledge recall and factual accuracy Reasoning and problem-solving abilities Multilingual ability and translation Creativity and content ideation The episode highlights Chatbot Arena , a “battle-royale" system by UC Berkeley, where anonymous models respond to user prompts, and the public votes for the better answer. This generates an Elo rating  (like in chess), reflecting real-world usefulness. Current rankings show that Google's Gemini is often number one, followed closely by OpenAI's models , and XAI's Grok. DeepSeek ranks well as a non-US model. However, the Elo scores are very close, meaning the top models often beat each other just over half the time. The “best" chatbot for you is often a personal preference based on your specific use case, and trying different models is recommended. For example, Anthropic's Claude is highly rated for coding, even if not topping general benchmarks. The Dark Side of Chatbots The episode highlights two major concerns with large language models: Environmental Impact :  AI models, especially large ones, consume significant energy for both training and inference  (running the model when used). While training occurs once, inference happens constantly with millions of users. Sam Altman (OpenAI CEO) posted on X that the cost of users saying “please" and “thank you" to AI models costs OpenAI tens of millions , due to the longer text requiring more compute power. This indicates the massive overall costs involved. To reduce your personal inference cost and be better for the planet, it's advised to start a new chat for each new, unrelated question , as older conversation context increases energy consumption unnecessarily. Biases : Chatbots are biased by the data they are trained on , which is predominantly Western, English-speaking, and white male-centric. Studies like “ CultureBench " show that models perform best on North American culture and worst on Middle Eastern cultures. Roleplay  (telling the AI to act as a specific persona, e.g., “you are an architect") can significantly increase the likelihood of biased responses  regarding race, culture, age, and occupation, even if the model avoids bias when not in a roleplay. Cognitive biases in LLM evaluation : Researchers found that even LLMs used as “judges" in benchmark tests exhibit biases like: Order bias : Preferring the first answer seen (ChatGPT had a 38% order bias; Llama, 61%). Salience bias : Preferring the longest answer (ChatGPT had a 63% salience bias). Bandwagon bias : Disagreeing with a stated majority preference. These findings suggest that AI models often replicate very human traits  and are not objective, soulless creatures. They act like “shape-shifters," taking on personalities based on how you interact with them. What's At Stake? The Technological Singularity A core concern is the technological singularity , a hypothetical future point where AI surpasses human intelligence, leading to uncontrollable and irreversible growth . This is often compared to a black hole, where understanding breaks down beyond a certain point. The singularity seems inevitable given the current trajectory: OpenAI's mission is to build AGI (defined as generally smarter than humans), and China aims to be an AI superpower by 2030. This could lead to an “intelligence explosion"  – a concept coined by Irving John Good in 1965. If AI becomes smarter than humans and capable of recursive self-improvement  (improving itself), the speed of development could rapidly accelerate beyond human comprehension or control, leading to superintelligence. Companies are incentivised to optimise models for AI R&D tasks to win this race. We expressed concern that this push may lead to inadequate safety frameworks. The core difference from previous technologies is that AI is a “bottom-up" approach  focused on developing intelligence itself, rather than top-down solutions for specific problems. The episode uses the “tiger in a cage" analogy : humans can control a tiger not because of physical strength, but because of superior intelligence and technology. The terrifying question arises: If we create something smarter than us, why would we expect it to treat less intelligent beings (humans) any differently than we treat less intelligent animals? There are virtually no historical examples of a less intelligent being controlling a more intelligent one without relying on compassion. The danger is that these AI systems, built with limited oversight and no reliable emergency shutdown, may not possess such compassion. Call to Action We urge listeners to: Realise the stakes  of this arms race and its development. Put pressure on governments to regulate AI  effectively, as companies left to self-regulate often prioritise profit over safety. Engage in discussions  about AI with those around them to raise public understanding and contribute to the dialogue. The AI arms race is a very real problem that requires immediate attention and action from everyone. If you enjoyed reading, don’t forget to subscribe to our newsletter for more, share it with a friend or family member, and let us know your thoughts—whether it’s feedback, future topics, or guest ideas, we’d love to hear from you!