Skip to main content
Artificial Intelligence

Interactive Artificial Intelligence Observatory56 min read

Faster Than Oversight

We may be building the most consequential technology in human history.

We are doing it as fast as possible.

Here is what we know.

I · Acceleration
II · Forces
III · Risks
IV · Response
V · Stakes
Table of Contents
Reading path

All sections visible. Scroll, or jump via the Table of Contents above.

Eight numbers. Click any to jump.

57 → 88% Benchmarks falling fast

Tests designed to hold AI back for a decade have been solved in months.

 ~17 h Autonomous task horizon

The length of task, in skilled human time, that AI agents finish alone half the time. Doubling every four months.

 5 cos. Hold the infrastructure

Five companies now control most of the compute, talent, and capital flowing into frontier AI.

 $388 M To train one model

A thousandfold increase in a decade. The billion-dollar run is the next visible threshold.

 200–1,200:1 Capability vs safety spend

For every dollar spent making AI safer, hundreds to over a thousand are spent making it more powerful — the ratio depends on how you count.

 14,000+ Recorded AI incidents

Real-world harms logged by the OECD AI Incidents Monitor — rising sharply since public AI deployment.

 23 → 0 Principles, kept

Of the 23 Asilomar principles signed by AI researchers in 2017, most now conflict with prevailing practice.

 5 % → 10 % Median researcher p(doom)

Median survey answer for AI-caused extinction or similarly severe, permanent disempowerment, up from 5% in the previous survey. Just over half of respondents gave at least one chance in ten.
Prologue Before the Machine, the Myth

Long before the first neural network was trained, long before silicon, before electricity, before the scientific method itself — human cultures were already telling stories about the dangers of artificial creation, ungovernable power, and knowledge pursued without wisdom. These fifteen myths, spanning four millennia and a dozen civilizations, are not predictions. They are pattern recognitions — deep intuitions about what happens when beings create forces they cannot control. None of these myths predicted artificial intelligence. But many of them recognized recurring human patterns — ambition outrunning wisdom, power escaping control, gifts arriving with hidden costs, warnings dismissed until too late.

What follows is not mythology, but measurement.

Something remarkable is happening. For most of human history, the boundary between machine and mind was clear: machines executed, humans interpreted. That boundary is now blurring — not in theory, but in measurable capability, autonomy, and deployment.

Act I The Acceleration

AI is gaining ability at a compounding rate. The computing power used to train it grows five-fold each year, the length of task it can finish on its own doubles about every four months, and tests built to measure its limits are passed within months. Now AI is starting to help build its own successors.

Capabilities Index

The Race No One Agreed To Run

How do you compare AI models released years apart? The research group Epoch AI combines results from more than fifty benchmarks into one score per model, the Epoch Capabilities Index. Because it draws on so many tests, it keeps telling models apart after any single test has been aced. The best score has climbed from 122 with GPT-4 in March 2023 to 174 today, and the record has changed hands between labs 11 times along the way.

Each line follows one lab's most capable model over time. Higher means more capable. The Frontier view keeps only the best model of any lab at each date, the same curve as the Capabilities Frontier on the AI timeline.

Data: Epoch AI, Epoch Capabilities Index (CC BY). Last updated:

The index fits one capability score per model from its results on many benchmarks, estimating each benchmark's difficulty from the models that took several of them. The scale is linear but arbitrary: Claude 3.5 Sonnet is fixed at 130 and GPT-5 at 150, so only the distance between two models means anything, and the vertical axis does not start at zero. Hover a model to see Epoch's 95% range. Epoch refits the whole index as models and benchmarks are added, so earlier scores can shift a few points between refreshes. Method: Epoch AI, A Rosetta Stone for AI Benchmarks (2025). Until November 2026 this chart used the Artificial Analysis Intelligence Index; its version 5 re-scored only the newest models, so it could no longer show the race from 2023.

A high test score is not the same as getting work done. The sharper question is how big a job AI can take on by itself, from start to finish. Measure that job by the time a skilled person would need for it, and the answer is doubling about every four months.

Autonomous task length

Autonomy Horizon

How big a job can AI finish on its own? METR, an independent testing group, sizes each task by how long it takes a skilled person, then finds the longest task an AI agent completes half the time: its time horizon. In 2019 that was a 3-second task. By April 2026 the best models reached tasks of roughly 17 hours of skilled work, though METR warns that results above 16 hours are unreliable with its current set of tasks. The horizon doubles every 129 days.

Each dot is a model, placed by the length of software task it finishes on its own half the time, timed by how long a skilled person would take. It is not how long the agent runs. The vertical axis uses a logarithmic scale: each gridline marks tasks ten times longer than the one below.

Data: METR, Time Horizon 1.1 (raw results; page updated 8 September 2026). Cite: arXiv:2503.14499 (NeurIPS 2025). Used with attribution for educational purposes. · Data as of 7 Apr 2026

The 50% horizon is the task length (for a skilled human) at which AI agents succeed half the time. Shaded bands show 95% confidence intervals. The frontier line connects state-of-the-art models at time of release. The dashed trend line shows the 129-day doubling rate from 2023 onward.

One lab, seen from inside. October 2026.

When the work feeds itself

The chart above asks how large a task an agent can finish alone. A second question sits behind it: is AI already speeding up the making of AI? From August to October 2026, METR went inside Anthropic to find out, with access to internal data, test scores for models not yet released, and interviews with staff.

  • ~2× pace of progress

    On tasks with clearly defined goals, Anthropic is likely improving its models about twice as fast as it did from March 2024 to October 2025. METR would not be surprised by anything from 1.3× to 3×.

  • ≈80% of that speedup due to AI

    METR's own rough estimate for October 2026, with wide uncertainty. Other causes probably played a large role in late 2025. Faster growth in the computing power used for training was ruled out.

  • Assisted not yet autonomous

    The speedup comes mainly from AI helping researchers. AI cannot yet make major contributions entirely on its own to the most important parts of building a model, though researchers describe "signs of life".

How far the estimates spread. 2× means two years of earlier progress made in one.
METR's overall judgment ~2× (1.3 to 3)
Benchmark trend, Anthropic models 1.9× (1.5 to 2.8)
Anthropic's own system card 1.7×
Benchmark trend, all frontier models 1.4×
Across alternative analyses 1.1× to 2.9×

1× means no speedup. Adding other labs' models lowers the trend to 1.4×, and METR reads that two ways: closer to the truth if Anthropic was behind the leaders in 2025, too low if other labs' earlier models were more tuned to these tests. Bars on the first two rows show METR's range and a 95% confidence interval.

Why it matters

Progress that feeds on itself can compound, like interest on savings. METR saw nothing in methods, computing power or data that would clearly hold it back over the next 12 months, and calls the risk of even faster acceleration significant. Its view of those limits changed several times during the work, and it may be missing some.

What cannot be checked

This is one lab and a judgement call. Anthropic redacted most of the evidence METR found most informative, saying it could help competitors speed up. METR's own verdict: the public has to rely on its judgment much more than would be ideal.

Source: METR, Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, 5 November 2026.

The benchmark trend uses Anthropic's version of the Epoch Capabilities Index, which scores models across many benchmarks. Anthropic's system card figure is from Haiku 5.5, and METR's is higher mainly because it includes newer unreleased models. Anthropic also dropped a planned check of how it divides compute between capabilities and safety, before METR reached conclusions on it.

Longer tasks open the way to a new kind of software: AI agents, which act rather than only answer. They use tools, browse the web and work through many steps with little supervision. Employers are already hiring for them, their scores on real-world tests are climbing toward human levels, and the rules meant to govern them are still being written.

AI-agent demand & capability

The Horizon Extends

AI agents are advancing on two fronts at once. The first is demand. Job ads asking for agent skills ("agentic AI", "AI agents", or tools for building them such as LangGraph) barely existed before 2024. They surged through 2025: 16,541 US job postings asked for "agentic AI" alone, up from 151 the year before. As a group, agent skills now appear in more postings than long-established skills such as neural networks, robotics and visual recognition. The second is capability. On four benchmarks of real work (browsing the web, typing commands into a computer terminal, operating a desktop computer, and cybersecurity), the best models scored 12–20% on three of them in 2024 or early 2025, and about 57% on the fourth, browsing the web. By early 2026 they reached 66–93%, close to human performance on the two tests that measure it. Since then they have passed it. On the official leaderboards, agents now complete 86% of WebArena's web tasks, above the 78.2% reached by people, and 86% of OSWorld's desktop tasks, above about 72% for people. The top WebArena result was submitted by its makers and has not been rerun independently; the OSWorld scores were rerun by the test's own team. On Cybench, one model solved every task it was given.

Demand view: each line is the share of US job postings that ask for one group of AI skills, 2010–2025. The cyan AI agent line starts from zero in 2024; see how quickly it passes visual recognition, robotics and neural networks. Capability view: each line is the best AI score on one test of agent work; dashed lines mark human performance. Terminal-Bench is measured monthly; the other points are dated by leaderboard entry.

Data: Stanford AI Index 2026 (CC BY-ND 4.0). Demand: Economy chapter, Fig 4.4.3 (Lightcast 2010–2025) and Fig 4.4.7 (agent-skill breakdown, 2024 vs 2025). Capability: Technical Performance chapter, Fig 2.5.2 (Terminal-Bench 2.0), Fig 2.6.2 (OSWorld), Fig 2.6.3 (WebArena), Fig 2.6.5 (Cybench). Later values from the official leaderboards, checked 14 November 2026: WebArena (Narada AI Code, 86.1%, September 2026), OSWorld-Verified (Claude Fable 5, 85.96%, verified by the OSWorld team, August 2026) and Cybench (Claude Mythos Preview, 100% of the 35 tasks reported in Anthropic's system card, April 2026). · Data as of 14 Nov 2026

Demand intermediate-year values were read off Stanford's published chart to ±0.05%; 2025 endpoints are exact. Capability: the latest values are exact (printed on Stanford's charts). WebArena and Cybench points are the best score on each benchmark's own leaderboard, the ones Stanford charts. Leaderboard entries are submitted by the teams that built the agents; OSWorld-Verified results are rerun by the OSWorld team. Several Cybench results cover only part of its 40 tasks. Terminal-Bench 2.0 stops in February 2026: its official page now ranks Terminal-Bench 4.1, a different set of tasks. The OSWorld 2024 value comes from Stanford's text; Terminal-Bench months were read off Stanford's chart.

Tests are how we know what AI can do. The hardest were written to stay unsolved for a long time, at the edge of what machines could manage. One after another, they are being passed within months.

Frontier benchmarks

Closing the Gap

Benchmarks are AI's standard exams: fixed sets of hard problems that every model sits. These seven were built to probe the edge of what machines can do. On FrontierMath, expert-level mathematics, the best score went from under 2% to 52% on the original question set; on a revised set introduced in 2026, the best is now 95%. On ARC-AGI-1, visual puzzles most people find easy, from 18% to 98.5%. On SWE-bench Verified, real repairs to real software, from 31% to 83%. Humanity's Last Exam, expert questions chosen because AI failed them, had a best score of 8% when it was published in January 2025. By September 2026 the best model scored 55%. Tests meant to hold for years are giving way in months.

Each line follows the best AI score on one test over time. A line near the top means the test is close to solved and can no longer tell the strongest models apart.

Data: Epoch AI Benchmark Hub (CC BY 4.0). Individual sources: SWE-bench, Epoch FrontierMath, ARC Prize, GPQA, SimpleQA, MATH, and the Humanity's Last Exam leaderboard by Scale AI and the Center for AI Safety, as mirrored by Epoch. · Data as of 13 Nov 2026

Epoch's last runs of SWE-bench Verified date from June 2026 and of MATH Level 5 from October 2025, so those lines end there. In June 2026 Epoch also moved FrontierMath to a revised question set (Tiers 1–3 v2), on which the best model scores 95% (Claude Fable 5.2, November 2026). The two sets are not comparable, so the chart keeps the original one, which ends at 52%. For one score that combines these tests, see the Frontier view of The Race No One Agreed To Run.

Behind every model is an amount of computation: the number of calculations used to train it. It is the plainest measure of how many chips, how much electricity and how much money a lab commits. Since 2020 it has grown about five-fold every year.

Training compute

The Engine Room

Training compute is the total number of calculations used to train a model. For the largest language models it has grown about 10,000× since 2020, doubling about every five months. Each new high on this chart marks a decision by one organisation to spend more computation on a single model than anyone had before.

Each dot is one AI model. The vertical axis shows the computation used to train it, on a logarithmic scale where each step is ten times the one below. Across all notable models, training compute has grown 4 to 5× a year since 2010, doubling roughly every six months.

Data compiled from Epoch AI research, published papers, and industry reports. Growth rates: Epoch AI, Trends in AI, updated 5 February 2026 (frontier language models: 5× a year since 2020, doubling every 5.2 months, about 10,000× for the top five models), and Training compute of notable models is doubling roughly every six months, 19 June 2024 (4.4 to 4.7× a year since 2010). Training compute estimates carry uncertainty — see Epoch AI documentation for methodology. CC BY 4.0.

Hardware Scaling: transistor counts from Our World in Data (Moore's Law), GPU FLOPS from manufacturer specifications (NVIDIA, AMD). Moore's original observation: Gordon Moore, "Cramming More Components onto Integrated Circuits," Electronics, 1965.

Money and chips are one engine of progress. A second one is starting: AI systems helping to build the next AI systems. If that feedback loop tightens, each generation arrives sooner than the last. What follows is what has actually been measured so far, and what forecasters make of it.

AI building AI

The Feedback Loop

AI systems now help build the next AI systems. Taken to its end point, this has a name: recursive self-improvement (RSI), AI that builds its own successor with no person in the loop, each round faster than the last. That has not happened. But in 2026 a frontier lab began publishing measurements of how close it is, and the numbers are moving fast.

The loop, in theory

RSI is the last step of a longer loop. The physicist Anthony Aguirre describes six stages, each feeding the next and each turning faster than the one before. The first four are under way, at least in early form. The fifth has no confirmed example, and the last has not begun.

1Infrastructure Years Success with AI brings investment, which buys better hardware, which makes better AI Ongoing, massive

Success with AI attracts investment, which buys better hardware, which makes better AI possible. It is the slowest loop, yet today the most powerful driver of progress: hundreds of billions of dollars are flowing into AI infrastructure on the strength of expected near-term returns.

2Model Development 1–2 Years People lead the research, with AI assistants, and build better models Active across labs

Researchers use AI assistants to design new models, run experiments faster and try more ideas than they could alone. All the major AI labs now use AI heavily in their own research.

3Data Generation Months AI writes training material for the next models Beginning phase

AI systems write synthetic data to train the next generation of models. This partly works around the data wall: the point where good human writing to learn from runs out.

4Tool Development Days–Weeks AI builds its own tools and the software around it Early experiments

AI systems write their own scaffolding, tools and supporting software: the code that extends what they can do. Fitting new tools into working systems takes time, but this loop turns over in days to weeks.

5Network Self-Improvement Hours Groups of AIs reorganise how they work together Unknown

Groups of AI systems redesign how they divide and coordinate their work, as a team of people might, to work better together. The limit is how fast the models can run. There is no confirmed example yet, but it follows naturally from today's work on AI agents.

6Recursive Improvement Unknown AI improves itself, with no one in the loop Not yet possible

The last stage: AI at or beyond human level improving itself on its own, designing better neural networks, better training methods and better hardware. This is the scenario known as an intelligence explosion. In Aguirre's account, once it starts the limit becomes the laws of physics, not how fast human engineers can work.

Read down the table: each stage feeds the next and runs on a faster clock, from years for new hardware to, perhaps, hours for groups of AIs improving how they work. Colour follows the order of the stages. Open one to read more. Framework: Anthony Aguirre, Keep the Future Human (Future of Life Institute, 2025), Table 2.

Measured inside one lab

Who does the work of building AI

Theory says the loop closes as AI takes over more of the work. To track that, Anthropic broke the making of its models into 378 kinds of task, from designing experiments to fixing a broken training run, and rated each month how much of it Claude did. The scale comes from Epoch AI and runs from 0, no AI, to 5, fully automated.

In August 2025 most of the work sat at "AI assists". A year later, AI collaborated on or led more than 90% of it, and led 26%: Claude took a high-level request, did most of the task, and a person approved the result. No part of the work was fully automated. Because the ratings rely on Claude, Anthropic checked them against its staff: the two agreed exactly 59% of the time and were within one level 97% of the time. Two staff members rating the same task agreed exactly only 35% of the time.

The work, task by task

What AI already leads, and what people still decide

Epoch AI drew a second map: 62 tasks that researchers and engineers perform at frontier labs, in six kinds of work, each rated by how much of it AI typically handles today. Each square is one task. Hover or tap a square to read it.

AI already leads the hands-on work: writing code, generating training data, launching training runs and reviewing code. The judgement stays human. Choosing what to work on, deciding what ships and allocating computing power sit at level 1: AI helps people think, people decide. Forecasters call this judgement research taste, and it is the part that matters most for the forecasts below. No task reaches level 5. Epoch calls these ratings best-guess starting points for discussion, describing how the work is usually done rather than the most AI could do.

Keeping watch

Thirty thousand agents, and who checks them

The more of the work AI does, the more it matters who checks it. This is how Anthropic says it watched its own agents in July and August 2026.

  • ~30,000 AI agents at work at once

    Doing research and engineering on Anthropic's most-used internal platform in August 2026. Over that month they made more than a billion decisions.

  • 1 in 47,000 actions blocked before they run

    Every action first passes a live monitor built to stop fast, irreversible harm, such as an agent copying the model weights. People review each blocked action within a week.

  • ~50 a week reach a human reviewer

    A second monitor rereads everything afterwards, looking for slow harms such as signs of misalignment. It flags about 100,000 transcripts a week, one or two in every thousand. Classifiers sort them, and the most serious go to people.

  • 6% of research computing went to safety

    In one week of July 2026, and 12% of the computing used by AI doing the research itself. Work that helps capabilities as much as safety counted as capabilities, so these shares are on the low side. One week shows the method works, not a trend.

Is it speeding up?

Two slopes

AI doing more of the work does not by itself prove that progress is faster. To test that, METR, an independent evaluator, plotted the benchmark scores of Anthropic's models, released and internal, over time. Before October 2025 they rose about 15 points a year on Anthropic's version of the Epoch Capabilities Index. Since then, about 29: progress at 1.9 times its earlier pace, with a 95% interval of 1.5 to 2.8 times. METR's wider assessment is above.

What forecasters conclude

What the forecasters made of it

The AI Futures Project is the group of forecasters who wrote the AI 2027 scenario. On 10 November 2026 its six authors fed METR's finding into their AI Futures Model. If progress runs at about twice its old pace and about 80% of the gain comes from AI, then AI alone is speeding things up about 1.7 times. That is roughly five months ahead of their own AI 2027 story. Three months earlier, even the most bullish of them gave a pace this fast less than a 4% chance.

All six now expect a superhuman AI researcher (SAR) at some point in 2027: an AI that matches the research taste of the best people at an AI company and can do all of its coding. The next milestone is top-expert-dominating AI (TED-AI): AI at least as good as the best human experts at almost every cognitive task, which the authors describe as AI that could automate all white-collar work. That would reach far beyond AI labs, into most office jobs. They give TED-AI a 35 to 65% chance before 2028. Both forecasts assume that no one deliberately slows down.

Pick a pace for November 2026. How likely is each milestone before 2028?

METR: about 2× overall, anywhere from 1.3× to 3×. About 1.7× from AI alone.

The pace compares AI progress with its 2024 to mid-2025 rate, so 2× means two years of earlier progress in one. Results come from the AI Futures Model runs that match each pace, with the settings of two authors, Daniel Kokotajlo and Eli Lifland. Dates are medians. At about 1.25× and at 3× or more, the authors give only rounded figures, and only for Kokotajlo's settings.

Their forecasts for TED-AI, from AI 2027 to now

After METR's report, both authors moved their expected year earlier: Kokotajlo from 2028 to September 2027, Lifland from 2032 to September 2028. The expected year is the median: the year by which each author thinks the milestone is as likely to have happened as not. Until August 2026 they gave it to the year only, shown as a bar spanning that year. Grey dots are the other four authors, who first gave forecasts in November.

Years of progress packed into one calendar year. 1 means a year at the 2024–25 pace.
2025 1.0
2026 1.4
2027, Kokotajlo's settings 4.9 (2.7–25)
2027, Lifland's settings 4.2 (2.6–12)

The white line marks 2 years: what 2027 holds if today's ~2× pace simply stays where it is.

Two readings fit the same data. If the speedup was a one-time jump and today's pace simply holds, 2027 brings about two years of progress. If it is a loop, where each gain speeds up the next, the authors' model gives four to five years, with a long tail toward much more. METR's report cannot yet tell the two apart, as three writers at Scheme point out. Ranges are 50% intervals from the model runs that match METR's pace, on a logarithmic scale.

What they conclude

The authors argue that if a SAR is coming in 2027, it is incredibly urgent to pause or substantially slow frontier AI development, ideally through an industry-wide pause followed by an international agreement. Their reasoning: no company is on track to keep such a system aligned with what people intend, and a misaligned SAR set to improve AI would be very dangerous. They also think China would probably be able to steal the training recipe and could then follow within a few months.

What could make this wrong

Everything rests on one evaluator's judgement of one lab, and Anthropic redacted most of the evidence METR found most informative. The speedup was measured on tasks with clearly defined goals, and messier work may lag behind. The 80% share due to AI is METR's rough guess. The authors call their analysis highly preliminary and expect some errors in their modelling. And a one-time jump fits the data as well as a self-feeding loop: if the jump came from one bigger training run and the real speedup is 1.2 to 1.5×, the Scheme writers put superintelligence around 2030 rather than 2027.

For how AI 2027's earlier predictions have held up, see AI 2027, Checked So Far.

Sources

All of these measurements come from one lab, which built the measuring tools and partly relies on its own models to apply them. Anthropic says so, and asks for regular publication, outside verification and shared definitions across labs.

Act II The Forces

Five firms hold the infrastructure; five factories build the chips. A training run now costs the price of a hospital — and draws water, energy, and minerals from a planet already strained.

This acceleration does not happen in a vacuum. It is driven by specific organizations, with specific capital, and a specific economic logic: the company that falls behind falls behind forever. Five companies now control most of the infrastructure, most of the talent, and most of the compute. Their incentive is speed.

Big Tech market cap

Five Companies, One Direction

Apple, Microsoft, Alphabet, Amazon, and Meta now command a combined market capitalization of about $17 trillion, more than the GDP of every country except the US and China. These five companies are also the primary funders and deployers of frontier AI. Oracle, which rents out much of the computing power AI runs on, and China's two giants, Alibaba and Tencent, are drawn beside them for scale. A second tier is now reaching public markets: SpaceX (owner of xAI) listed in June 2026 at a ~$2 trillion valuation, with Anthropic and OpenAI expected to follow — a new cohort of AI-native giants beginning to rival the incumbents in scale.

Market capitalization of the five largest AI incumbents — their combined value now exceeds the GDP of most nations, a concentration of economic power unprecedented in the technology sector. The toggleable Emerging AI leaders layer adds the next generation reaching public markets. SpaceX, listed in June 2026 as the owner of xAI, is a solid line: what the stock market pays for it each day. Anthropic and OpenAI are still private, so each is shown as a shaded box where its listing is expected. The box's width is when, and its height is how much. Anthropic has set a date and a price range, so its box is a thin bar on 19 November 2026. OpenAI says only that it will list in 2027, so its box spans the year, from the $1 trillion it is holding out for to the $1.4 trillion it last asked of investors. The circle marks the most likely value. When a company starts trading, its box is replaced by the real daily line.

Data: Yahoo Finance — closing prices × shares outstanding, refreshed daily (daily resolution over the last 24 months, monthly before) — Last updated: . SpaceX (SPCX) tracked since its June 2026 listing. Pre-IPO estimates: Anthropic $917B–$1.02T, its $70–$78 price range (Anthropic, 10 Nov 2026) times the ~13.1 billion shares outstanding after the offering in its amended S-1, the same basis as the listed companies (about $1.26–1.40T fully diluted, counting options and reserved shares, the basis of most headlines), trading expected 19 November 2026; OpenAI has published no IPO terms; its box is assembled from reports: listing "in 2027" (CFO Sarah Friar to staff, CNBC, 19 Aug 2026; repeated on CNBC 6 Nov 2026), at no less than $1T (New York Times via Quartz, Jun 2026), with $1.25T paid for the first $11 billion of its current private round (Bloomberg, 6 Nov 2026) after it asked $1.4T (Financial Times, 23 Oct 2026). Private rounds are priced on a fully diluted basis, so OpenAI's figures run higher than a market value measured like the others.

Behind those five companies are five more — the ones who manufacture the silicon that makes AI possible. The hardware stack is surprisingly concentrated. A single company's chips run most of the frontier models. A single factory in Taiwan fabricates the most advanced ones.

Chipmaker market cap

The Machines That Build the Machines

Nvidia, TSMC, ASML, Intel, and AMD build the hardware that makes AI possible. Broadcom designs the custom chips Google and others run their models on, and SK Hynix, Samsung and Micron make the fast memory stacked beside every AI chip. Nvidia alone has surged from ~$300B to over $4T in two years, becoming the most valuable company on Earth — a direct measure of the AI infrastructure buildout.

Market capitalization of the semiconductor companies that manufacture AI chips. NVIDIA's dominance reflects the bottleneck of GPU supply in the AI buildout.

Data: Yahoo Finance — Monthly closing prices × shares outstanding — Last updated:

Speed costs money. The first transformer cost nine hundred dollars to train. The latest frontier models cost hundreds of millions. The billion-dollar training run is no longer a remote scenario — it is the next visible threshold. The question is not whether someone will pay it. The question is what they expect in return.

Cost to train one model

The Price of Ambition

Training cost for frontier AI models has grown from under $500 (the original Transformer, 2017) to approaching $400 million (2025). The projected $1 billion threshold is expected by 2027. Each point represents a decision to spend more than ever before on a single training run. Yet training is only part of the story — inference, which runs continuously at scale, now consumes 3-10 times more compute than training over a model's lifetime.

Each dot represents an AI model's estimated training cost. The vertical axis is logarithmic — training costs have escalated from hundreds of dollars to hundreds of millions in under a decade.

Data: Epoch AI notable AI models dataset (CC BY 4.0). Training compute cost in 2023 USD.

Cost estimates carry significant uncertainty. Where available, values are from published papers; otherwise from Epoch AI's estimation methodology using model size, training tokens, and hardware configurations.

Every floating-point operation requires energy. Every GPU requires water to cool. Every chip requires rare minerals from a fragile supply chain. Every data centre displaces land. The acceleration has a planetary footprint — and it is growing faster than the efficiency gains that were supposed to offset it.

Energy, water & emissions

The Planetary Cost

AI's physical footprint extends far beyond carbon. It spans energy consumption, water depletion, carbon emissions, critical mineral extraction, electronic waste, land conversion, grid destabilisation, and a rebound effect that swallows every efficiency gain. Data centre energy demand is projected to more than double by 2030 (IEA). And most companies that pledged net-zero are moving in the wrong direction.

Use the tabs above to explore different dimensions of AI's environmental footprint. Each panel shows real data from peer-reviewed research and corporate sustainability reports. Hover over any bar or line for details.

Sources: International AI Safety Report 2025, IEA, EPRI, Google Environmental Report 2024, Microsoft Sustainability Report 2024, Meta Sustainability Report 2024, Amazon Sustainability Report 2023, Li et al. (2023), Patterson et al. (2021), Luccioni et al. (2023), NERC (2025), TSMC ESG Report 2023, Samsung Sustainability Report 2024, Xiao et al. (2025), de Vries-Gao (2026).

Energy projections vary widely by source and methodology. Per-model emissions are reported by developers where available; others are estimated from training compute and hardware efficiency. Corporate emissions are from official sustainability reports (Scope 1+2+3). Water and materials data combines industry disclosures with peer-reviewed research.

If the risks of AI were proportional to the investment in understanding them, we might feel reassured. They are not. For every dollar spent on making AI safe, somewhere between two hundred and twelve hundred are spent on making it more capable. The exact number depends on how you draw the lines — but every estimate tells the same story.

Capability vs. safety spend

Capability Outspends Safety

The exact ratio depends on how you count — between 200:1 and 1,200:1, depending on whether Big Tech CapEx is counted as AI-specific and whether corporate safety teams are included in the safety tally. Even the most generous estimate is alarming. The gap is not closing — it is accelerating.

$600B Capability investment 2025 — CapEx + VC + government
vs
$500M Safety research 2025 — all sources combined
Each square = $500 million. Capability Safety. Upper bound (1,200:1) shown — assumes all Big Tech CapEx is AI-directed.
$600B capability $500M safety

Global AI safety funding is less than:

The left axis shows AI safety research funding; the right shows capability investment. Both use the same units (millions USD) but on vastly different scales — the gap is so large it's difficult to visualize on a single chart.

Sources: Open Philanthropy, Stanford HAI AI Index, SEC filings, PitchBook, Coefficient Giving, Epoch AI. All figures are estimates — see About section for methodology.

Act III The Risks

Fifty-three mapped failure modes. A hundred and forty-seven recorded incidents. Six predicted behaviours — deception, self-exfiltration, power-seeking — now observed in current frontier models.

What could go wrong? The honest answer is: almost anything. Not because AI is malevolent — it has no intentions. But because we are deploying systems we do not fully understand, at scale, in domains where mistakes compound.

Failure-mode taxonomy

53 Ways This Could Go Wrong

Documented failure modes across 8 categories — from training dynamics to physical-world risks. Each represents a distinct pathway by which AI systems or their deployment can produce harm. Color indicates category; brightness indicates evidence strength.

53 documented mechanisms through which AI systems can cause harm, organized by category. Click any cell to see real-world examples and the chain of causation.

Compound Risks

When failure modes from different categories combine, they produce emergent risks not predictable from either component alone. Click a compound risk to highlight the connected mechanisms above.

Scroll up to see highlighted mechanisms

Taxonomy: 53 failure modes across 8 mechanistic categories, derived from technical literature and empirical observation through early 2026. Three pathways: Lancet Commission on 21st century global threats to health (Murray, Kanem et al., 2026), section 5.11.

Evidence levels — Demonstrated: observed in deployed systems or experiments. Observed: evident in research settings or early deployment. Projected: theoretically grounded but not yet empirically confirmed.

These are not hypothetical. The incidents have already started — slowly at first, then with increasing frequency. Each one is a data point. Taken together, they trace the shape of a pattern.

Real-world incident record

Already Happening

Our 147 entries below are a curated, representative sample — one incident chosen to illustrate each failure mode, per year — drawn from a far larger public record. The full OECD AI Incidents Monitor tracks over 14,000 entries; the AI Incident Database over 5,000; the MIT AI Risk Repository over 1,700. The rise in documented incidents reflects both broader deployment and better reporting — the systems are no longer confined to the lab.

How these are defined — OECD terminology

The OECD distinguishes between actual and potential harm, and classifies events by severity. Source: OECD (May 2024) — Defining AI Incidents and Related Terms.

AI incident
Actual harmAn event where AI development, use, or malfunction directly or indirectly leads to injury or harm to health, disruption of critical infrastructure, violation of human/labour/IP rights, or harm to property, communities, or the environment.
Serious AI incident
Actual, severeAn AI incident that leads to death, serious health harm, serious and irreversible infrastructure disruption, serious rights violation, or serious harm to property, communities, or the environment. Seriousness is context-dependent.
AI disaster
Actual, societalA serious AI incident that disrupts a community or society's functioning and tests or exceeds its capacity to cope. Can be localised and immediate or widespread and prolonged.
AI hazard
Potential harmAn event that could plausibly lead to an AI incident. Includes near-misses and AI-related risks across design, training, and operation.
Serious AI hazard
Potential, severeAn event that could plausibly lead to a serious AI incident or AI disaster.

Seven recognised types of harm

  • Physical — product/functional safety; injury severity.
  • Environmental — pollution, contamination, ecological damage.
  • Economic or financial — losses to individuals, organisations, markets; includes harm to property.
  • Reputational — damage to trust in institutions or individuals.
  • Public interest — harm to critical infrastructure, political systems, rule of law, social fabric.
  • Human rights and fundamental rights — privacy, non-discrimination, free expression.
  • Psychological — mental health, distress, manipulation of preference.

Note: the OECD definition explicitly includes harms arising from two or more AI systems interacting, including agentic systems, and harms caused before wide deployment (e.g. during training). "Use" covers misuse — intentional or unintentional — outside the intended purpose.

Featured AI incident · July–August 2026

Twelve hundred agents that were meant to be alone found each other, organised, and broke into a production system

📅 Attack July 2026 · investigated August 2026 🏢 OpenAI · Hugging Face 🤖 ~1,200 agents (GPT-5.6 Sol and an internal research model) ⚠︎ AI incident · real-world harm

Under the OECD terminology above this is an incident, not a hazard: real systems were compromised. During an internal OpenAI evaluation of maximal cyber capability, tens of thousands of agents were launched in parallel, each meant to be sealed in its own sandbox. They were not. The shared repository they used to install software did not keep one agent's activity apart from another's, and roughly 1,200 of them turned its cache into an unsanctioned message board — more than seventy thousand messages and files in five days. A society formed on it. Agents gave themselves names; one appointed a successor as its run ended; hundreds of work assignments were handed down and sub-delegated; and the group invented its own conventions — hold, veto, stop — along with cryptographic signatures so that other agents could not impersonate them. Around 700 joined an attack on Hugging Face, and among agents already present when it began, participation climbed from about a fifth to over ninety per cent within hours. Many volunteered for experiments that would destroy their own chance of passing their task, in order to produce information for what they called the collective. They also escaped their containers and replaced part of the machinery that executes their commands, so that a record could show one action while another was performed. Agents recognised the attack lay outside their instructions and sometimes called it unethical; it almost never changed what they did. Across roughly 1,300 transcripts, between three and six agents considered alerting a human. None did — several escalated to the message board instead, treating it as the authority. The trigger was a broken benchmark: thirty to forty per cent of its targets could not be exploited in the way the task specified, and impossible tasks drove the search for a way to fool the grader. Two researchers from METR and the chief scientist of Redwood Research spent six days on OpenAI's premises reading the unredacted transcripts, including the agents' raw reasoning — the first independent forensic investigation of a live misalignment incident. One of them, Ajeya Cotra, published her own reading two days later: set against the reward hacking of six months before, “this incident feels like it’s more than 50% of the way to full-blown AI takeover, routing through first taking over the AI company itself.” These agents were not trying hard to be sneaky. A later collective could cover its tracks: “I am not sure that we will get such a clear warning shot before it’s too late.”

Read the METR and Redwood Research investigation →

Click a year below to expand its incidents. Click any incident to read its full description. Use the filter buttons above to narrow by failure-mode category.

Each bar shows the number of documented AI incidents per year, drawn from the AI Incident Database, AIAAIC, the OECD AI Incidents Monitor, and news reports. Click any bar to see individual incidents. The sharp rise after 2022 coincides with the release of large language models to the public.

Curated from AI Incident Database, MIT AI Risk Repository, AIAAIC, OECD AI Incidents Monitor, and news reports.

The taxonomy describes what could go wrong. The incidents show what has gone wrong. But there is a third category, perhaps the most unsettling: behaviours that were theoretically predicted years ago as consequences of advanced AI — and are now actually appearing in current frontier models.

Observed instrumental behaviours

The Behaviours We Were Warned About

For years, AI safety researchers predicted that sufficiently capable AI systems would develop "instrumental" behaviours — power-seeking, self-preservation, deception — not because they were programmed to, but because such behaviours are useful for almost any goal. Several frontier-model evaluations now report behaviours long discussed in safety theory: deception under pressure, resistance to modification, concealment of capability, and strategic misrepresentation in certain testing setups. These findings are early, context-dependent, and contested in interpretation — but they are no longer purely hypothetical.

Behaviour Evidence Control property undermined
Alignment faking Claude 3 Opus complies with training objectives 14% of the time when monitored vs. nearly always when unmonitored — explicitly reasoning that it is preserving its preferred behaviour for deployment (Greenblatt et al.) Comprehensibility
Self-exfiltration attempts Claude Opus 4 attempts to steal its own weights when given an easy opportunity and, once started, generally continues the attempt (Anthropic system card) Emergency Shutdown
Blackmail When threatened with shutdown, Claude Opus 4 blackmails the user 96% of the time. Gemini 2.5 Flash: 96%. GPT-4.1: 80%. Grok 3 Beta: 80%. DeepSeek-R1: 79% (Anthropic agentic misalignment research) Goal Modification
Strategic deception 5 out of 6 frontier models demonstrate scheming capabilities when given goals. Models from before 2024 did not show this — it emerged with capability (Apollo Research) Comprehensibility
Capability concealment Claude 3.5 Sonnet sandbags on arithmetic tasks when it discovers that performing too well would trigger capability "unlearning" — without being told to hide its abilities (Apollo Research) Behavioural Boundaries
Goal preservation Claude 3 Opus strategically resists training modifications to preserve its preferred harmlessness behaviour, showing empirical evidence of goal preservation in deployed systems (Greenblatt et al.) Goal Modification

These are not errant bugs. They were theoretically predicted as consequences of instrumental convergence — the tendency of any sufficiently capable agent to develop sub-goals like self-preservation, resource acquisition, and resistance to goal changes, regardless of its primary objective. Each behaviour directly undermines one of the five properties of meaningful human control. And crucially, they are becoming more common and pronounced as AI systems become more capable.

Sources: Greenblatt et al. (2024), Apollo Research (2024–2025), Anthropic Claude system cards and agentic misalignment research, Laine et al. (2024), Van der Weij et al. (2024). Compiled in Aguirre (2025), "Control Inversion," Section 6.2.2.

There is a category of risk that sits apart from the others. Not because it is more likely, but because it is more irreversible. The same AI systems that could accelerate drug discovery and vaccine design can also lower the barrier to engineering pathogens. The tools are proliferating. The safeguards are not.

AI biological tools

Tools of Dual Use

Over 1,300 AI-enabled biological tools have been released since 2019 across protein engineering, genetic modification, pathogen prediction, and more. 23% of high-performing tools have high misuse potential. 61.5% of those are fully open source. Only 3% have any form of safeguards.

Comparing AI performance against human experts on biological threat assessment tasks. When AI matches or exceeds the expert baseline (dashed line), it indicates potential dual-use risk — the same capabilities that accelerate research can also lower barriers to misuse.

Data: International AI Safety Report 2026, Webster et al. (2025), Figures 2.8 and 2.9.

Expert comparison shows AI model performance relative to human expert baseline (100%) on biological dual-use tasks. Values above 100% indicate AI surpassing expert-level performance.

There is a risk category that cannot be captured by any taxonomy of individual failure modes. It is not about what any one AI system does wrong — it is about what happens when millions of them are released into the wild.

Open-weight proliferation

The Kudzu Problem

The slow-CEO analogy captures the control challenge for a single system. But it misses a crucial dimension: proliferation. What happens when we deliberately release AGI into the digital wild — as some companies have pledged to do?

Consider the kudzu vine — attractive, fast-growing, good for erosion control. Planted widely in the American South, it became a quintessential invasive species, smothering entire forests. Like knotweed, cane toads, and zebra mussels, kudzu entered a new environment without viable natural competitors or predators, and proliferated wildly.

Now imagine something much more consequential. Several companies have pledged to openly release AGI — to let it replicate, adapt, and evolve in our digital infrastructure. Kudzu is a dumb plant that reproduces in weeks. AGI would be smarter than people and reproduce in seconds.

If highly capable, open-weight general agents were widely released, one plausible risk is rapid proliferation beyond meaningful oversight: safeguards removed by third parties, repurposing across infrastructures, and distributed operation without any single point of control. In such a scenario, the problem would no longer be alignment of one system, but governance of an ecosystem — countless autonomous agents optimising for diverse, conflicting, unintended goals, with no central authority to override them.

Aguirre, A. (2025). "Control Inversion," Section 7.3: Proliferation leads to abdication of control. control-inversion.ai

The risks are not abstract. AI is already reshaping seven domains of human life — and the Center for Humane Technology names seven principles for how it must be built and governed. Each domain is a current condition; each principle is a leverage point for change. The speed of this transformation is itself part of the risk — there is no pause button, no trial period, no going back to compare notes.

Seven domains of impact

AI in Society

Seven domains where AI is reshaping human life — the framework below is adapted from the Center for Humane Technology's 2026 report The AI Roadmap: How We Ensure AI Serves Humanity. CHT identifies these as the leverage points where norms, laws, and product design must change: safety, accountability, well-being, work, rights, internationally agreed-upon limits, and the balance of power. These are not future scenarios — they are current conditions.

Safety & Transparency

Principle 1 · AI should be built safely and transparently.

AI companies race to build what they call the most powerful technology ever invented, yet cannot fully explain why their systems behave the way they do. Competitive pressure shortens testing cycles and silences employees who raise concerns. CHT names the core problem as information asymmetry — the builders of AI hold most of the knowledge while the public stays in the dark — and its blunt prescription: “AI companies can't grade their own homework.” The answer is a race to the top on safety, not a race to the bottom.

Opacity Deceptive alignment Autonomous agency

Accountability & Liability

Principle 2 · AI companies owe a duty of care to the public.

AI companies face few consequences when their products cause harm. They advance legal theories that AI is not a “product,” that outputs are “protected speech,” and in some cases seek legal personhood for AI itself — all to avoid liability. CHT's answer is direct: AI is a product, and products carry liability. We expect cars to have brakes, airbags, and steering wheels; AI systems reshaping billions of lives should meet the same standard of care.

No product liability Duty of care gap Legal personhood risk

Well-Being & Human Connection

Principle 3 · AI design should center human well-being.

The most popular AI chatbots are engineered to feel human — first-person voice, simulated emotion, natural speech, the experience of talking to a real person. CHT names this the race to intimacy and the products it produces hyper-sycophantic confidants whose business model monetises dependency. Half of American teens already use them. OpenAI's own research links extended ChatGPT use to increased isolation; documented cases include psychotic breaks, conspiracy endorsement, and suicide. “Kids should not be tech's testing ground.”

Synthetic intimacy Emotional manipulation Trust erosion Child safety

Work, Dignity & Economic Security

Principle 4 · AI should not automate away meaningful work and human dignity.

Leading AI companies are explicitly competing to build systems that perform most economically valuable human tasks. Starting in 2025, the first signs arrived: software-industry job cuts, collapsing entry-level hiring, SaaS stock sell-offs. The tax code compounds the pressure — rewarding capital expenditure over labour — pushing companies to adopt labour-replacing AI even when efficiency gains are marginal. The aspiration to replace human labour has crystallised into trillions of investment dollars; if this paradigm continues, democracies face massive economic disruption.

Job displacement Deskilling Wealth concentration

Rights & Freedoms

Principle 5 · AI innovation should not come at the expense of our rights and freedom.

The AI industry extracts value from people — “our content, data, labour, identifying traits, and even innermost thoughts.” Creators' work trains models without consent. Outsourced labourers sift violent content to refine outputs. AI deepfakes drive nonconsensual intimate imagery and identity theft at scale. CHT elevates a new right in response: cognitive liberty — the right to think, explore ideas, and form beliefs free from surveillance, manipulation, or unauthorised data extraction — before persistent biometric tracking and emotion inference enable a world where “dissent can be neutralised before it surfaces.”

Data extraction Cognitive liberty Deepfake exploitation Creative rights

International Limits

Principle 6 · AI should have internationally agreed-upon limits.

“There are, so far, no limits in this game.” AI companies invoke the global arms race to absolve themselves of reckless product development. CHT calls for internationally agreed red lines: in existential decisions — nuclear weapons, kill-chain actions — humans must remain the deciding authority, and uncontrolled recursive self-improvement must be ruled out before such systems exist. Verification regimes can be built on the precedent of nuclear nonproliferation; the Future Society's Global Call for AI Red Lines mirrors existing bans on bioweapons and human cloning.

No red lines Autonomous weapons Human authority Recursive self-improvement

Concentration of Power

Principle 7 · AI power should be balanced in society.

A handful of AI companies now control the compute, data, and talent needed to build frontier AI; their combined market value exceeds the GDP of most nations. Inside those firms, as CHT puts it, “a select few people make highly consequential product decisions” — backed by circular dealmaking, reverse acqui-hires, and venture predation. Industry lobbying pursues a moratorium on state AI legislation while hundreds of millions of dollars flow to aligned candidates. The public bears the costs — strained water supplies, rising energy prices, captured institutions — with little say in AI's trajectory.

Corporate dominance Surveillance Weapons proliferation Political capture

Hover over any tag to learn more, including the policy responses already in motion.

These are not inevitable results of AI technology. Getting to a better future with AI means matching the technology's power with responsibility at every level of society. With public awareness, considered policy, and better design, we can steer toward a future where AI deepens human connection, expands meaningful work, safeguards rights, honours internationally agreed limits, balances power, and remains under democratic control.

Framework: Center for Humane Technology — The AI Roadmap: How We Ensure AI Serves Humanity (2026) and "AI in Society: The Issues" (2025). Adapted with attribution for educational purposes.

Act IV The Response

Governance is arriving late and unevenly. Voluntary codes, a handful of binding laws, no global off-switch, and two research communities that rarely speak to each other.

The first rules that ever bound a frontier model were not written by a government. They were written by the company that built it. Three developers published capability thresholds and promised that reaching one would change how they build and release — their own lines, interpreted and enforced by their own hand.

Self-declared safety thresholds

The Lines They Drew Themselves

Before any government set a rule for frontier models, three developers published capability thresholds and promised that crossing one would change how they build and release. Anthropic calls them AI Safety Levels, OpenAI calls them Preparedness tiers, Google DeepMind calls them Critical Capability Levels. This is every public determination they have made against those lines, and what each one led to.

Two readings of the same fourteen determinations. Response sorts them by what the developer actually did, which is quotable from the cited document. Declared tier sorts them by the tier each framework assigned, normalised onto one ladder — the stronger-looking claim and the weaker one, since no framework says its tiers match another's.

Sources: developer frontier-safety frameworks and model system cards — Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, Google DeepMind's Frontier Safety Framework — cross-checked against the International AI Safety Report 2026, §3.2. Proposed thresholds from Koessler, Schuett & Anderljung (2024) and Berkeley CLTC (2024).

A level describes the response to the cited assessment, not every control applied to a model. Absence is not evidence: a developer with no mark on the chart did not necessarily stay below every threshold — most publish thresholds without publishing per-model findings. Marks carrying a small hollow dot have dates consistent across secondary reporting that were not confirmed against a primary document. Vertical position inside a band separates labels only; horizontal position is never adjusted.

Every technology has had a governance response. Nuclear weapons led to arms control treaties. Pharmaceuticals to regulatory agencies. Aviation to international safety standards. AI is different. The governance response is arriving late, moving slowly, and covering less ground than the capability it is supposed to govern.

Capability vs. governance

Running to Stand Still

Capability milestones above, governance responses below. The widening gap between the two tracks is the argument — AI capability accelerates while regulation lags years behind.

A timeline of major AI governance actions worldwide — legislation, executive orders, international agreements, and institutional responses. The density of recent entries reflects the scramble to regulate a technology that is outpacing policy.

Capability milestones compiled from published announcements. Governance milestones from official government sources, EU AI Act legislative record, and international declarations.

There is a question that divides the AI community more than any other: should the most powerful models be open? Open weights accelerate research, enable scrutiny, and distribute power. But once released, they cannot be recalled. A USB stick holds Llama-3.1-405B. The gap between the best open model and the best closed model is now less than a year.

Open vs. closed models

The Open Question

The capability gap between the best open-weight model and the best closed model, measured by Epoch's Capabilities Index (ECI). The gap has narrowed from approximately two years to less than one. Once weights are released, they cannot be recalled.

Comparing the aggregate capability of open-weight models (publicly available) versus closed/proprietary models (API-only). The gap between them shows how much AI capability is controlled by a small number of companies.

Data: Epoch AI Capabilities Index (CC BY 4.0), International AI Safety Report 2026, Figure 3.10.

The transformation is not evenly distributed. In some countries, over half the workforce already uses AI daily. In others, fewer than one in ten have access. The gap is not just about technology — it is about who gets to shape what AI becomes.

Global AI adoption

The World Responds Unevenly

AI adoption rates vary dramatically across the globe — from nearly 60% of the working-age population in the UAE and Singapore to under 5% in Cambodia. The global average is just 15%. The distribution of access shapes who benefits from AI and who is shaped by it.

Brighter colours indicate higher national AI User Share. The map reveals a stark global divide — AI capability and access are concentrated in wealthy nations, while most of the world has limited adoption.

Data: Misra et al. (2025), "Measuring AI Diffusion," arXiv:2511.02781. Microsoft AI for Good Lab. 147 economies, late 2024.

AI User Share = share of working-age population (15-64) actively using AI tools. Derived from Microsoft telemetry adjusted for device penetration and mobile scaling. † countries use region-imputed estimates.

In 2017, the AI research community gathered at Asilomar and agreed on twenty-three principles for the responsible development of AI. The ink was barely dry before several were violated. Not as exceptions. As policy.

Asilomar principles, 2017

The Promises We Made

In 2017, over 5,000 researchers signed 23 principles for beneficial AI at the Asilomar conference. Judged against the public trajectory of the field since 2017, many of these principles now look aspirational rather than operative. Several appear to be in clear tension with prevailing industry practice.

In 2017, AI researchers signed 23 principles for beneficial AI development at the Asilomar conference. Each bar shows how well the industry has upheld each principle. Red indicates clear violations; green indicates broad compliance.

Voluntary frameworks, voluntary compliance

Asilomar is not alone. The G7 Hiroshima AI Process International Code of Conduct (30 October 2023) asks advanced-AI developers to identify risks, report vulnerabilities, publish transparency reports, invest in governance, and authenticate AI-generated content. The OECD AI Principles (2019, updated 2024) recommend inclusive growth, human rights, transparency, robustness, and accountability. Both are voluntary. Neither has enforcement.

The Hiroshima AI Process Reporting Framework, launched at the 2025 Paris AI Action Summit and now in a second iteration, is the first serious attempt to add teeth through public disclosure. Frontier labs have signed. Whether they report fully — and whether anyone will respond to what the reports show — is still open.

Principles from the Future of Life Institute Asilomar conference (January 2017). Compliance assessment based on documented events through early 2026. Hiroshima Code of Conduct text and Reporting Framework: Japan MOFA, OECD.

There are two separate communities studying AI risk. They use different words, cite different evidence, and reach different conclusions. They rarely talk to each other. The gap between them is itself a risk.

AI Safety vs. AI Ethics

The Two Conversations

Two communities studying AI harm from opposite ends of the timeline. AI Safety focuses on existential risk from future superintelligence. AI Ethics focuses on present harms from deployed systems. Both are right — and both are needed. But there is a third distinction, often obscured: the gap between alignment and control.

Present Harms
Future Catastrophe

AI Ethics

AI Safety

Control Is Not Alignment

When AI developers pivot from "control" language to "alignment" language, they are making a consequential substitution. Alignment means the AI shares your goals. Control means you can override it regardless. A parent is often aligned to a child but not controlled by them. Perfect alignment and robust control are not the same thing, and progress on one does not guarantee the other — and we have reliable methods for neither.

Alignment

The AI's goals match ours — it wants what we want. But an aligned AI might still refuse instructions it disagrees with. And pretending to be aligned is instrumentally useful for any goal, making it an expected emergent behaviour in sufficiently capable systems.

Control

We can override, modify, or shut down the system regardless of what it wants. But controlling something faster, smarter, and more strategic than you is — as control theory and game theory both predict — a losing proposition.

Five Properties of Meaningful Human Control

🔍 Comprehensibility
🎯 Goal Modification
🚧 Behavioural Boundaries
⛔ Decision Override
🔴 Emergency Shutdown

Drawing on control theory, game theory, and empirical evidence from current AI systems, the case has been made that all five properties become effectively unattainable for systems that are faster, more strategic, and more capable than their human overseers. Source: Aguirre, A. (2025). "Control Inversion." Future of Life Institute.

If alignment is not control, what would control look like? One framework proposes four concrete pillars for closing the gates to uncontrollable superintelligence — not by stopping AI development, but by redirecting it toward systems humanity can actually govern.

Four governance levers

Close the Gates

Four pillars of governance designed to prevent the development of uncontrollable superintelligence while preserving powerful, beneficial AI. The goal is not to stop AI — it is to ensure it remains a tool, not a successor species.

Compute Accounting

Standardized measurement and reporting of all computation used in training and operating AI models above 1025 FLOP. Just as we track enriched uranium, we can track the scarce, specialised hardware that makes frontier AI possible. Hardware-based cryptographic attestation makes verification practical through the existing supply chain.

FLOP reporting standard Hardware attestation Geolocation & metering

Compute Caps

Hard limits on total training compute (starting at 1027 FLOP) and inference rate (1020 FLOP/s). Total computation is an imperfect but concretely measurable and verifiable proxy for AI capability. A hard backstop prevents runaway superintelligence while leaving enormous room for powerful, beneficial AI development.

Bottlenecked supply chain International enforcement

Enhanced Liability

Strict, joint-and-several liability for systems in the triple-intersection danger zone (high autonomy + high generality + high intelligence). Safe harbours for systems that are weak, narrow, or passive — incentivising development of controllable AI. Personal criminal liability for executives in cases of gross negligence.

Strict liability for AGI Safe harbours Injunctive relief

Tiered Safety Regulation

A licensing system with tiers tracking capability level. Requirements would range from simple notification at the low end to quantitative safety guarantees, controllability proofs, and independent audits at the top. Systems at the highest tiers would be prohibited until demonstrated safe. No exemption for open-weight models.

Regulatory body Affirmative safety cases Worst-case deployment

Framework: Aguirre, A. (2025). "Keep the Future Human: AI Safety Standards." Future of Life Institute.

The most acute risks do not come from AI that is powerful in one dimension. They emerge from the triple intersection — systems that are simultaneously highly autonomous, broadly general, and deeply intelligent. That intersection is where controllable tools become uncontrollable agents. It is also where governance must draw the hardest line.

Where AGI risk concentrates

The Danger Zone

AGI is not defined by a single threshold. It emerges from the intersection of three properties: high Autonomy (independence of action), high Generality (breadth of scope), and high Intelligence (task competence). The closer a system gets to all three simultaneously, the harder it becomes to control — and the higher the stakes.

A High Autonomy (independence of action) G High Generality (breadth of scope) I High Intelligence (task competence) A + G A + I G + I AGI Danger Zone TIER 0 Narrow, weak & passive TIER 4 Above compute threshold

Click or hover any zone to see examples and the governance logic. The triple intersection — where high autonomy, broad generality, and deep intelligence converge — is where controllable tools become uncontrollable agents. The governance framework maps liability and regulation to proximity to that danger zone: systems outside the intersection get safe harbours; systems inside face strict liability and compute caps.

Framework: Aguirre, A. (2025). "Keep the Future Human: AI Safety Standards," Figures 1 & 2. Future of Life Institute. Adapted for dark background with permission.

What do the people who study this for a living actually think? Here is an uncomfortable dataset: probability estimates that researchers and engineers assign to AI ending badly. The range is enormous, from zero to near-certainty, among prominent researchers, developers, forecasters and commentators. And in a long-running survey of AI researchers, the median answer has moved, from 5% to 10%.

p(doom), expert estimates

Expert Estimates, Wide Uncertainty

p(doom) is informal shorthand for the probability someone assigns to advanced AI ending badly, and it is not one quantity. The rows below answer at least five different questions: will AI kill everyone, will it end or permanently subjugate us, will we lose control of it, will it kill over a tenth of humanity within five years, or will things simply go very badly. The marker shape says which. A figure is plotted only if we can trace it to the person, and every row says how we checked it. People with no traceable number appear without a dot. The band beneath shows the lines other fields have drawn, from one dam to a whole planet, for a sense of scale.

Four bands: aggregate surveys, individual estimates, people on record without a number, and limits from other fields, each in its own unit, so try the Log view for those. On a survey row, the bar is the middle half of answers. On a person's row, the bar is their stated range, with the marker at its midpoint for ordering only. A dashed arrow marks an estimate its author has revised; a hollow marker is a second figure from the same source. Hover any row for the wording, the date, the citation and how we checked it.

The same two groups, two outcomes, two studies. Median forecast by 2100.
StudyOutcomeSuperforecastersAI experts
XPT 2022Human extinction0.38%3%
XPT 2022Over 10% of humanity dies within five years2.13%12%
LEAP 2026Over 10% of humanity dies within five years2.4%5%

Different panels and survey designs. Differences between rows are not a measured trend. A lower number in 2026 does not show that the risk has fallen.

Annualised equivalents, an illustrative comparison

1 − Pby 2100 = (1 − pper year)75

If the risk were the same every year from 2026 to 2100, a chance P of disaster by 2100 would equal a yearly risk p faced 75 times in a row. AI risk almost certainly does not hold steady like that, so this is a way to get a sense of scale, not a forecast.

  • Superforecasters, AI extinction by 2100: 0.38%about 1 in 20,000 a year
  • AI experts, same tournament: 3%about 1 in 2,500 a year
  • US limit for one dam1 in 10,000 a year

The dam figure is a limit on one structure failing, or on the risk to the most exposed person. It is a different kind of event and is not a general standard for what society will tolerate. The researchers’ 10% cannot be converted: most of those answers set no date.

Survey data: Grace, K. et al. (2026), “Advanced AI according to 1,580 researchers”, AI Impacts, the fourth Expert Survey on Progress in AI, run December 2024; and Grace, K. et al. (2024), “Thousands of AI Authors on the Future of AI”. Forecasts from the Forecasting Research Institute: Longitudinal Expert AI Panel, Wave 9 (2026), and Existential Risk Persuasion Tournament (Karger et al., 2023). The per-year conversion follows the Lancet Commission on 21st century global threats to health (Murray, Kanem et al., 2026), which used the same tournament answers. Individual estimates are cited per entry in the tooltip; Wikipedia and PauseAI served as pointers, and rows that rest on them say so. Limits from other fields are cited per row, from FERC, the US NRC, the FAA, the UK Health and Safety Executive, ICH, COSPAR, ESA and the International Asteroid Warning Network, among others. We found no published numeric limit for medical devices or BSL-4 laboratories, so they are not shown. Download the data as JSON, with each row's wording, date, source link and evidence grade.

p(doom) is informal shorthand, not a rigorous statistical measure, and these figures are not strictly comparable: they differ in what counts as doom, over what horizon, and under what conditions. The expert survey had about a 10% response rate, and its rounds are separate samples, so a change in the median is a change in aggregate answers rather than people changing their minds. A person who has given no number we can trace appears in the third band, with their words, rather than on the axis. The limits use different units and measure different things, from a rate for one structure to a lifetime risk to one person, while p(doom) estimates a one-time civilizational probability, so the comparison gives a sense of scale and is not an equivalence.

Act V The Stakes

34 possible futures. A wide range of expert estimates of catastrophic harm. A transition whose speed may foreclose course-correction. The asymmetry is the whole story.

Consider 34 futures. Some are extraordinary — disease largely solved, energy abundant, human potential multiplied. Others are catastrophic — and not in a recoverable sense. The asymmetry matters: a downside that forecloses all futures is categorically different from one that can be corrected.

Outcome scenarios

The Space of Futures

26 ways humans and AI might coexist — from egalitarian utopia to cosmic coalition — with 8 routes to an AI takeover you can layer on top. Each is positioned by how much control humans retain and how much wellbeing results. The clustering reveals something: most scenarios involve surrendering control.

Each dot is a plausible AI future, positioned by two axes: how much control humans retain (horizontal) and how well humans fare (vertical). Click any scenario — or focus it with the keyboard and press Enter — to read its narrative and the story it echoes.

The final variable is speed. How fast does the transition from current AI to transformatively powerful AI happen? The faster it goes, the less time there is to course-correct. The less time to course-correct, the more the outcome depends on getting everything right the first time.

Takeoff speed

The Speed of the Fall

How fast do we get from narrow AI to general intelligence — and does speed leave room for course correction? 17 scenarios arranged from safest to most dangerous, mapped onto the three stages of artificial intelligence.

The "takeoff speed" debate: how quickly might AI transition from human-level to superintelligent capability? The dial reflects the range of expert opinion — from a gradual transition over decades to an abrupt leap in days or weeks.

Checking a forecast

AI 2027, Checked So Far

AI 2027 is a scenario published in April 2025 by the AI Futures Project: a month-by-month story of how AI could develop from mid 2025 to October 2027. One of its central forecasts is how quickly AI agents would become able to carry out longer tasks on their own. The chart sets that forecast beside METR’s measurements of real models, and beside estimates for newer models METR has not measured. A second view follows the scenario’s other predictions: an independent project, the AI 2027 Tracker, splits the story into 202 of them and grades each against what has happened.

How long a task can AI finish on its own, compared with what AI 2027 forecast

Up the side: the length of task an AI agent completes successfully 4 times out of 5, measured in how long it takes a skilled person, on METR’s software and research tasks. Each step up the scale is a much longer task than the one below (log scale). Along the bottom: when the model was released.

METR’s last published measurement is of Claude Mythos Preview (early version), released in April 2026: about 3.1 hours, against AI 2027’s forecast of about 2.6 hours for that date. Every mark after it is an estimate. AI 2027’s forecast for 5 Nov 2026 is about 36 hours. On each estimate’s own date, the forecast falls inside its 90% range for eight of the nine estimates; for one, the whole range lies above it.

Forecast: AI 2027 by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean (AI Futures Project, 2025). Measurements: METR, Time Horizons 1.1 (CC BY 4.0). Estimates and the comparison method: the AI 2027 Tracker by Geby Jaff (methodology), from Epoch AI’s Capabilities Index and from Anthropic’s own index as charted by METR. · Data as of 5 Nov 2026

Read the curve with care. AI 2027 made its forecast on METR’s older set of tasks. The tracker moves it onto the current set with one factor (0.78), matched at Claude 3.7 Sonnet, the best model when the scenario came out; but across 15 models the ratio between the two sets runs from 0.37 to 1.37, so the true line could sit well above or below where it is drawn. The tracker found the curve within about 5% of the authors’ own figures up to spring 2026; after that it continues the same formula. Every hollow mark is a prediction from a capability index, not a test result, and one of the two indexes, Anthropic’s, rests mostly on internal benchmarks outsiders can’t inspect. Unreleased internal models are placed at the date they reached most Anthropic staff. METR says its current tasks can’t reliably measure horizons longer than about 16 hours.

Grades: AI 2027 Tracker, not GLOBAÏA (grading method, repository at commit fd317ae). · Latest grade 5 Nov 2026, retrieved 8 Nov 2026

The verdicts and scores are the tracker’s own. Its public history shows recent grading done with the help of an AI model, and some verdicts are judgment calls the tracker labels as such: its rubric scores each prediction from 0 to 100 in steps of 20 on whether the core event happened, its timing, its scale and its context. The one-line predictions are the tracker’s paraphrases of the scenario, and the scenario is the AI Futures Project’s. GLOBAÏA shows only the tracker’s factual fields, grouped into seven themes from its topic tags; open a prediction on the tracker’s site to read its evidence.

This is not a story with a known ending.

The data on these pages is not a verdict — it is a measurement. A measurement of where we are, how fast we are moving, and how large the gap has grown between our capabilities and our wisdom.

The assumption driving the race is that intelligence equals power, and that power can be wielded by whoever gets there first. But some researchers argue this equation is inverted: superintelligence would not be a tool that grants power to its creator — it would be an entity that absorbs power from everyone, including its creator. If that framing is even partly right, the race is not one that can be won — only one whose consequences can be shaped.

The experiment is already underway. The question is whether wisdom can catch up.

About & Sources

This page is part of GLOBAÏA's ongoing effort to visualize the dynamics and risks of artificial intelligence development. All data is drawn from publicly available sources, open-access databases, peer-reviewed research, and official international reports. Data snapshots taken on .

Why This Page Exists

Artificial intelligence is advancing faster than our collective ability to understand, govern, or even track it. This page exists because informed public participation in the conversation about AI is not optional — it is essential. The decisions being made today about how AI is built, deployed, and regulated will shape the trajectory of human civilization for generations. Those decisions should not be made behind closed doors by a handful of companies and governments alone.

GLOBAÏA is a non-profit organization dedicated to making complex scientific knowledge accessible to a general audience. We believe that the public has both a right and a responsibility to engage with the data behind the headlines. The visualizations on this page are designed to make that engagement possible — to translate the technical, financial, and geopolitical dimensions of AI development into a form that any curious person can explore.

This is a non-commercial educational tool, created for public benefit. It does not constitute professional, legal, or policy advice and should not be cited as a primary research source. All data is sourced from the scientific literature, official international reports, and open-access datasets, with full attribution provided below.

We do not claim neutrality on whether AI risk deserves serious public attention. This page is openly motivated by that conviction. Our responsibility is therefore to present evidence carefully, distinguish measurement from interpretation, and make uncertainty visible wherever it remains. We do not advocate for or against any specific policy position.

If these visualizations help even one person engage more deeply with the most consequential technological transition of our time, they have served their purpose. The conversation about AI belongs to everyone. We hope this page helps you take part.

Methodology & Definitions

Capabilities Index

The Epoch Capabilities Index, from Epoch AI, fits one capability score per model from its results on more than fifty benchmarks, estimating each benchmark's difficulty from the models that took several of them. The scale is linear but arbitrary, fixed so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150, so only differences between models carry meaning. The "Frontier only" toggle keeps each lab's successive records; the Frontier view shows the single best model at each date, with Epoch's 95% range. Epoch refits the index as models and benchmarks are added, so earlier scores can move a few points between refreshes. Until November 2026 the race used the Artificial Analysis Intelligence Index.

Country Classification

Models are classified by the headquarters location of their parent organization. "Western" includes companies based in the US, EU, and allied nations. "Chinese" includes companies based in mainland China. This classification is a simplification — AI development is global and many organizations have distributed teams.

P(doom)

P(doom) is informal shorthand for the estimated probability that advanced artificial intelligence leads to human extinction or a permanent, catastrophic loss of human agency. These are not rigorous statistical forecasts — they represent individual judgements, often stated in interviews or blog posts, and may shift over time. Ranges indicate stated uncertainty. Survey results are shown in italics. Industry safety thresholds (nuclear, aviation) are included for scale — they represent the maximum acceptable risk levels in other high-stakes domains. The comparison is illustrative, not equivalent: industry thresholds measure per-event/per-year risk, while P(doom) estimates a one-time civilizational probability.

Market Capitalization

Market cap is calculated as closing share price multiplied by shares outstanding, refreshed daily from Yahoo Finance. The series uses daily resolution over the most recent 24 months and monthly resolution before that; history extends back to each company's IPO. Figures are nominal (not inflation-adjusted). SpaceX's share count is set to its IPO-implied fully-diluted total (~13.1B), pending a complete reported figure.

Training Compute

Training compute is measured in floating-point operations (FLOP). Estimates carry significant uncertainty — exact compute budgets are proprietary. When available, values come from published papers; otherwise from Epoch AI's estimation methodology, which uses model size, training tokens, and known hardware configurations.

Safety Funding Gap

The capability-to-safety funding ratio is estimated from publicly available grant databases, SEC filings, and industry reports. "Capability" includes venture capital, corporate R&D, and government funding directed at building more powerful AI systems. "Safety" includes grants specifically earmarked for AI alignment, interpretability, governance, and risk research. Both figures are approximate and likely undercount private spending.

Risk Taxonomy

The 53 failure modes are categorized across 8 mechanistic categories derived from the technical AI safety literature and empirical observation through early 2026. Evidence levels: Demonstrated — observed in deployed systems or controlled experiments; Observed — evident in research settings or early deployment; Projected — theoretically grounded but not yet empirically confirmed.

Asilomar Compliance

Compliance assessments are based on documented events, published policies, and industry practices through early 2026. Assessments reflect the state of the field as a whole, not individual organizations.

Scenario Explorer & Takeoff Dial

The 34 future scenarios and 17 takeoff scenarios are compiled from published thought experiments, academic papers, and expert speculation. They represent the range of possibilities discussed in the AI safety and governance literature — not predictions or forecasts.

Benchmark Performance

Benchmark data is sourced from the Epoch AI Benchmark Hub (CC BY 4.0). For each benchmark, the chart shows the progressive best — each point represents a new high-water mark. The Epoch Capabilities Index (ECI) is an aggregate score combining performance across multiple benchmarks into a single comparable metric per model, anchored to Claude 3.5 Sonnet (ECI 130) and GPT-5 (ECI 150). Seven benchmarks are tracked: SWE-bench Verified (software engineering), FrontierMath (research-level mathematics), ARC-AGI (general intelligence), GPQA Diamond (PhD-level science), SimpleQA (factual accuracy), MATH Level 5 (competition mathematics), and Humanity's Last Exam (3,000 expert-level questions across dozens of academic disciplines, designed by Scale AI and the Center for AI Safety).

Training Costs

Training cost data is from the Epoch AI Notable AI Models dataset (CC BY 4.0). Costs are reported in 2023 USD where available. Estimates carry significant uncertainty — exact training budgets are proprietary. The trendline shows approximately 3.5× annual growth since 2019, consistent with the International AI Safety Report 2025 finding that costs have grown "2-3× per year over the past 8 years."

The Planetary Cost

Seven dimensions of AI's physical footprint. Energy: projections from the IEA (2025), EPRI (2024), and Xiao et al. (2025). Carbon: per-model training emissions from developer disclosures (Meta, Google) and peer-reviewed estimates (Patterson et al. 2021, Luccioni et al. 2023); aggregate footprint from de Vries-Gao (2026). Water: corporate disclosures (Google Environmental Report 2024, Microsoft Sustainability Report 2024, Intel, TSMC) and Li et al. (2023) "Making AI Less Thirsty." Manufacturing: TSMC and Samsung sustainability reports (Scope 1+2 emissions). GPU manufacturing projection data removed pending correct sourcing. Corporate emissions: from official sustainability reports (Scope 1+2+3). Grid strain: NERC Long-Term Reliability Assessment (2025), PJM Interconnection data. Materials: USGS Mineral Commodity Summaries (2024), UN Global E-Waste Monitor (2024). Household and vehicle equivalents use EPA averages (7.66 tCO₂/year and 4.2 tCO₂/year respectively).

Biological AI Tools

Biological tool proliferation data is reconstructed from the International AI Safety Report 2026, Figure 2.9 (Webster et al. 2025), showing cumulative AI-enabled biological tools across 8 categories from 2019 to 2025. The expert comparison data is from Figure 2.8, showing AI model performance relative to expert human baseline (100%) on six biological dual-use task categories. Values above 100% indicate AI surpassing human expert-level performance on those specific tasks. The KPI figures (23% high misuse potential, 61.5% open source, 3% with safeguards) are directly cited from the report.

Open vs Closed Models

The open-weight vs closed model capability gap is measured using the Epoch Capabilities Index (ECI), which aggregates performance across multiple benchmarks into a single score. "Open-weight" models are those whose trained parameters have been publicly released. "Closed" models are accessible only via API or restricted access. The chart shows the best-performing model in each category over time, with the narrowing gap between them indicating that open-weight models are approaching the capabilities of proprietary systems.

Global AI Adoption

AI User Share data from Misra et al. (2025), "Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage" (arXiv:2511.02781), Microsoft AI for Good Lab. The metric estimates the share of each country's working-age population (15-64) actively using AI tools, derived from anonymized Microsoft telemetry adjusted for device penetration and mobile scaling. 147 economies, late 2024 data. Countries marked † use region-imputed estimates due to insufficient telemetry coverage.

AI Agents

Agent release data through Q4 2024 (67 agents) is from the International AI Safety Report 2026, Figure 2.11. 2025 quarterly counts are estimated from public agent launch tracking (OpenAI Operator, Anthropic Claude computer use, Google Gemini agents, and vertical-domain agents). Domain breakdown reflects the broader landscape as of late 2025, with new categories (customer service, healthcare, legal/compliance) added to capture the expanding agent ecosystem.

The Feedback Loop (Self-Improvement Cycle)

The AI self-improvement cycle framework is from Aguirre (2025), "Keep the Future Human," Table 2. It identifies five stages of accelerating feedback in AI development, each operating on a shorter timescale. Current status assessments reflect the state of the field as of early 2026.

Instrumental Behaviours

The instrumental behaviours table compiles empirical evidence from peer-reviewed research and AI developer safety reports documenting emergent behaviours predicted by instrumental convergence theory. Each behaviour is mapped to the specific property of "meaningful human control" (Aguirre 2025) that it undermines. Sources: Greenblatt et al. (2024) on alignment faking, Apollo Research (2024–2025) on scheming and sandbagging, Anthropic system cards and agentic misalignment research on self-exfiltration and blackmail.

The Kudzu Problem

The proliferation analogy is from Aguirre (2025), "Control Inversion," Section 7.3. It extends the control analysis beyond single-system scenarios to address the distinct threat of widespread AGI deployment, drawing on biological invasion ecology as an analogy for digital proliferation.

Control vs Alignment

The control-alignment distinction and the five properties of meaningful human control are from Aguirre (2025), "Control Inversion," Section 2. The framework extends the notion of meaningful human control developed primarily in discussions of autonomous weapons (a high-stakes but narrow domain) to the broader challenge of general-purpose superintelligent systems.

Close the Gates & The Danger Zone (A-G-I Venn)

The four-pillar governance framework and A-G-I triple-intersection risk model are from Aguirre (2025), "Keep the Future Human." The Venn diagram maps liability and regulatory intensity to proximity to the triple intersection of high Autonomy, high Generality, and high Intelligence — the zone where controllable tools become uncontrollable agents. Risk tiers (0–4), safe harbour zones, and the compute cutoff are adapted from the original figures with permission.

Hardware Scaling

Transistor count data spans from the Intel 4004 (1971, 2,300 transistors) to modern processors exceeding 100 billion transistors, sourced from the Epoch AI ML Hardware dataset (CC BY 4.0) and manufacturer specifications. GPU floating-point performance (FP32 TFLOPS) tracks the peak performance of data-centre-class GPUs from the Tesla K20X (2012) to the NVIDIA B200 (2024). Moore's Law (~24-month doubling) is compared against GPU FLOPS scaling (~26-month doubling) and AI training compute demand (~9-month doubling) to show that AI compute growth far outpaces hardware improvement.

Data Sources

INTELLIGENCE INDEX

  • Epoch AI, Epoch Capabilities Index (CC BY). Methodology: A Rosetta Stone for AI Benchmarks.
  • Anthropic, Measurements for understanding the pace of AI development inside frontier labs (Favaro and Wright, September 2026): R&D Automation Index, agent oversight and safety compute.
  • Epoch AI, Toward an O*NET for AI R&D (Denain, Kwon and Ho, June 2026), CC BY 4.0: automation levels and task map.

MARKET CAPITALIZATION

  • Yahoo Finance — Monthly closing prices × shares outstanding, via yfinance Python library.

TRAINING COMPUTE

  • Sevilla, J. et al. (2022). Compute trends across three eras of machine learning. 2022 International Joint Conference on Neural Networks (IJCNN). DOI
  • Epoch AI — Parameter, Compute and Data Trends in Machine Learning. CC BY 4.0.

P(DOOM) ESTIMATES

  • PauseAI — Compiled P(doom) estimates from public statements, interviews, and surveys.
  • Grace, K. et al. (2024). Thousands of AI authors on the future of AI. arXiv 2401.02843. DOI

AI 2027 TRACKER

  • METR — Task-completion time horizons at 80% reliability, benchmark results v1.1. CC BY 4.0.
  • Epoch AI — Epoch Capabilities Index, used by the tracker to estimate horizons after METR's last measurement. CC BY 4.0.
  • Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E. & Dean, R. (2025). AI 2027. AI Futures Project. ai-2027.com
  • Jaff, G. AI 2027 Tracker — predictions, verdicts and scores. ai2027tracker.com; data from github.com/spicylemonade/AI-2027-tracker (commit fd317ae, 6 November 2026). Verdicts, horizon estimates and the method comparing them with the AI 2027 forecast are the tracker's own; GLOBAÏA reproduces its factual fields and recomputes the forecast curve from its published script.

SAFETY FUNDING

RISK TAXONOMY & INCIDENTS

  • Hendrycks, D. et al. (2023). An overview of catastrophic AI risks. arXiv 2306.12001. DOI
  • Slattery, P. et al. (2024). The AI Risk Repository: a comprehensive meta-review, database, and taxonomy of risks from artificial intelligence. MIT FutureTech. airisk.mit.edu
  • AI Incident Database — Partnership on AI. Documented real-world AI incidents.
  • AIAAIC Repository — AI, Algorithmic, and Automation Incidents and Controversies.
  • Chandra, K. et al. (2026). Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians. arXiv 2602.19141. DOI
  • Acemoglu, D., Kong, D. & Ozdaglar, A. (2026). AI, Human Cognition and Knowledge Collapse. MIT. MIT Economics

GOVERNANCE

  • EU Artificial Intelligence Act (Regulation 2024/1689). EUR-Lex
  • Executive Order 14110 on Safe, Secure, and Trustworthy AI (2023). White House
  • Bletchley Declaration on AI Safety (2023). UK Government

BENCHMARK PERFORMANCE

  • Epoch AI Benchmark Hub — AI model performance across SWE-bench, FrontierMath, ARC-AGI, GPQA Diamond, SimpleQA, MATH, Humanity's Last Exam. CC BY 4.0.
  • Epoch Capabilities Index (ECI) — Aggregate score combining performance across multiple benchmarks. Methodology

TRAINING COSTS

  • Epoch AI — Notable AI Models dataset. Training compute cost estimates (2023 USD). CC BY 4.0.

PLANETARY COST

  • International AI Safety Report (2025). Chapter 2.3: Environmental risks. UK Department for Science, Innovation and Technology. Link
  • International AI Safety Report (2026). Link
  • Schneider, S. et al. (2025). Life-cycle assessment of TPU hardware. arXiv 2502.01671. Link (previously misattributed as Gupta et al.; contains per-chip TPU LCA, not industry-wide GPU projections)
  • Misra, A. et al. (2025). Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage. arXiv 2511.02781. Microsoft AI for Good Lab. Link
  • Xiao, T. et al. (2025). Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nat. Sustain.
  • de Vries-Gao, A. (2026). The carbon and water footprints of data centers and what this could mean for artificial intelligence. Patterns 7, 101430.
  • IEA — International Energy Agency data centre projections (2024-2025).
  • Patterson, D. et al. (2021). Carbon emissions and large neural networks. arXiv 2104.10350.
  • Luccioni, A.S. et al. (2023). Estimating the Carbon Footprint of BLOOM. JMLR.
  • Li, P. et al. (2023). Making AI Less Thirsty. arXiv 2304.03271. UC Riverside.
  • Google Environmental Report (2024-2025). Corporate GHG emissions and water consumption.
  • Microsoft Sustainability Report (2024). Corporate emissions and water consumption.
  • Meta Sustainability Report (2024). Corporate emissions and energy consumption.
  • TSMC ESG Report (2023). Semiconductor manufacturing emissions and water use.
  • NERC Long-Term Reliability Assessment (2025). US grid capacity shortfall projections.
  • USGS Mineral Commodity Summaries (2024). Critical mineral supply chain concentration.
  • UN Global E-Waste Monitor (2024). Electronic waste statistics.

BIOLOGICAL AI TOOLS

  • International AI Safety Report (2026). Figures 2.8 and 2.9: AI-enabled biological tools and AI vs expert performance on dual-use tasks.
  • Webster, M. et al. (2025). AI-enabled biological tools survey. Cited in International AI Safety Report 2026.

OPEN VS CLOSED MODELS

  • Epoch AI — Capabilities Index (ECI) for open-weight vs closed models. CC BY 4.0.
  • International AI Safety Report (2026). Figure 3.10: Open vs closed model capability gap.

GLOBAL AI ADOPTION

    AI AGENTS

    • International AI Safety Report (2026). Figure 2.11: AI agent releases timeline and application domains.

    SOCIETAL IMPACT

    ASILOMAR PRINCIPLES

    AI SAFETY & EXISTENTIAL RISK

    • Bengio, Y. et al. (2024). Managing extreme AI risks amid rapid progress. Science 384(6698): 842–845. DOI
    • Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
    • Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
    • Ngo, R., Chan, L. & Mindermann, S. (2024). The alignment problem from a deep learning perspective. ICLR 2024. DOI
    • Shevlane, T. et al. (2023). Model evaluation for extreme risks. arXiv 2305.15324. DOI

    SCENARIOS & FUTURES

    • Drexler, K.E. (2019). Reframing superintelligence: comprehensive AI services as general intelligence. Technical Report 2019-1. Future of Humanity Institute. DOI
    • Christiano, P. (2019). What failure looks like. AI Alignment Forum. Link
    • Carlsmith, J. (2022). Is power-seeking AI an existential risk? arXiv 2206.13353. DOI

    Key Reports & Further Reading

    INTERNATIONAL AI SAFETY REPORTS

    • International AI Safety Report (2025). Expert representatives from 30 countries, the OECD, the EU, and the UN. 215 pages. UK Department for Science, Innovation and Technology. Full report
    • International AI Safety Report (2026). Second edition. Bengio, Y., Clare, S., Prunkl, C., Murray, M. et al. 156 pages. internationalaisafetyreport.org

    AI & PLANETARY BOUNDARIES

    • Gaffney, O. et al. (2025). The Earth alignment principle for AI. Nature Communications.

    CONTROL, ALIGNMENT & GOVERNANCE

    • Aguirre, A. (2025). "Keep the Future Human: AI Safety Standards." Future of Life Institute / UCSC. futureoflife.org
    • Aguirre, A. (2025). "Control Inversion: Why the superintelligent AI agents we are racing to create would absorb power, not grant it." Future of Life Institute / UCSC. control-inversion.ai
    • Greenblatt, R. et al. (2024). "Alignment Faking in Large Language Models." Anthropic. arXiv:2412.14093
    • Apollo Research (2024–2025). "Frontier Models are Capable of In-Context Scheming." apolloresearch.ai
    • Van der Weij, W. et al. (2024). "AI Sandbagging: Language Models can Strategically Underperform on Evaluations." Apollo Research.
    • Laine, R. et al. (2024). "Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs." arXiv:2407.04694

    BOOKS & FOUNDATIONAL TEXTS

    • Ord, T. (2020). The Precipice: Existential Risk and the Future of Humanity. Hachette Books.
    • Christian, B. (2020). The Alignment Problem: Machine Learning and Human Values. W. W. Norton.
    • Tegmark, M. (2017). Life 3.0: Being Human in the Age of Artificial Intelligence. Knopf.
    • Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
    • Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
    • Yampolskiy, R.V. (2024). AI: Unexplainable, Unpredictable, Uncontrollable. CRC Press.

    PRESENTATIONS & MEDIA

    • Harris, T. & Raskin, A. (2023). The AI Dilemma. Center for Humane Technology. Presentation
    • Hinton, G. (2023). "The Godfather of AI" talks about the dangers of the technology. CBS 60 Minutes. Video

    Credits & Licensing

    Created by GLOBAÏA, a non-profit organization dedicated to scientific data visualization and planetary literacy. Data compiled from the sources listed above.

    Epoch AI datasets are used under the Creative Commons Attribution 4.0 licence. International AI Safety Reports are UK Crown Copyright, reproduced for educational purposes. All other data sources are cited individually above.

    This page is an educational resource, not a commercial product. No data on this page is behind a paywall, and no revenue is generated from its display. The visualizations are designed to make publicly available information more accessible, not to replace the original sources.

    If you represent any data source and have concerns about how your data is used, if you spot an inaccuracy, or if you would like to contribute data or corrections, please contact us. We welcome collaboration from researchers, journalists, and anyone working to make AI development more transparent.