The most repeated claims about AI, stated fairly and checked against the data: water and power, “it’s dumb”, “it’s autocomplete”, “it can’t discover anything”, and hype in the other direction.
A typical text prompt uses about 0.3 Wh of electricity and about 0.3 mL of water, a few drops. US golf facilities use about 30× more water than all US data centres. The real problems are local: particular grids and water supplies near big sites.
The ‘bottle’ comes from older estimates that covered 10–50 answers or assumed far more energy. Today's reports: about 0.3 mL per prompt.
Long-document, reasoning and agent requests can use 10–100× more than a short prompt.
Read the full check →AI-focused data centres used about 0.5% of the world's electricity in 2025. That is small but growing fast, and local strain is real.
“The environment” bundles carbon, water, air quality, land and supply chains. The evidence is strongest for local strain and weakest for global-scale harm today.
Read the full check →Google, OpenAI, Epoch AI and a Microsoft-led study agree on 0.24–0.34 Wh for a typical text prompt. Long, agent or video tasks use far more.
0.24 Whenergy for a median Gemini text prompt, all overheads counted (Google)
These are medians for short text prompts. Long documents, reasoning, agents, images and video use far more (see the entry on heavy workloads).
Read the full entry →The real problem is local: many new data centres are being built in places short of water. Nationally, golf courses use about 30× more.
17.4 bn galdirect water consumed by all US data centres in 2023 (LBNL)
Metrics differ: LBNL reports consumption, golf figures are water applied, and USGS figures are labelled as withdrawal or consumption.
Read the full entry →Billion US gallons per year. Log scale: each gridline is 10× the previous. Blue = data centres (all workloads, not just AI).
What this measures: How much water US data centres use in a year, next to other big water users.
Annual values derived from published daily or litre figures. Years and metrics differ (consumption vs water applied vs withdrawal; USGS 2015 is the latest national compilation in this form). National totals hide local water stress, which is where data-centre water use matters most.
| Item | Value | Note |
|---|---|---|
| US data centres, on-site (2023) | 17.4 bn gal | consumption |
| US data centres, via power plants (2023) | 211 bn gal | indirect consumption |
| US golf courses (2024) | 531 bn gal | irrigation applied |
| US livestock (2015) | 730 bn gal | withdrawal |
| US thermoelectric power plants (2015) | 1,573 bn gal | consumption |
| US irrigation (2015) | 26,718 bn gal | consumption |
Millilitres of water. Log scale: each gridline is 10× the previous. Blue = AI.
What this measures: How much water one AI text prompt uses, next to everyday activities and products.
Scopes differ, so compare like with like. The AI prompt, toilet and shower rows are direct use. Coffee, beef and jeans are supply-chain or life-cycle footprints (coffee and beef include rainwater taken up by crops and pasture); the fair AI comparison for those is the life-cycle row (Mistral, 45 mL). Long-document, reasoning and agentic requests use far more than a median text prompt.
| Item | Value | Note |
|---|---|---|
| One AI text prompt (Gemini median, on-site) | 0.26 mL | data-centre cooling |
| Same prompt incl. power-plant water | 1.35 mL | our calculation, US average |
| One AI response, full life cycle (Mistral Large 2) | 45 mL | incl. hardware manufacturing |
| 2024 “bottle per email” estimate (GPT-4) | 519 mL | assumed ~140 Wh per email |
| One toilet flush (US standard) | 6.06 L | direct use |
| One 10-minute shower | 94.6 L | direct use |
| One cup of coffee | 140 L | supply-chain footprint |
| One quarter-pound beef patty | 1.75 m³ | supply-chain footprint |
| One pair of jeans, cradle to grave | 3.78 m³ | life-cycle footprint |
Watt-hours. Log scale: each gridline is 10× the previous. Blue = AI.
What this measures: How much electricity different AI requests use, next to everyday activities.
AI figures are company disclosures (Google, OpenAI) or published models (Microsoft-led Joule study, Epoch AI, IEA); none is independently audited. Household figures assume typical UK appliances (Hannah Ritchie). The web-search figure is Google's from 2009.
| Item | Value | Note |
|---|---|---|
| One AI text prompt (Gemini median) | 0.24 Wh | Google, all overheads |
| One Google web search (2009 figure) | 0.3 Wh | old figure |
| Average ChatGPT query | 0.34 Wh | OpenAI, method unstated |
| Long reasoning-style AI query | 3.91 Wh | modelled median |
| LED bulb for 1 hour | 10 Wh | |
| Microwave for 1 minute | 16.7 Wh | |
| Charging a smartphone | 19.4 Wh | |
| Laptop for 1 hour | 20 Wh | |
| AI query over a 100,000-token document | 40 Wh | estimate, paid once per upload |
| Agentic AI task with reasoning | 50 Wh | IEA, GPU-only |
| TV for 1 hour | 60 Wh | |
| Streaming video for 1 hour | 77 Wh | 2019, incl. device |
| Boiling a full kettle | 100 Wh |
Models read text in word pieces, not letters, so letter-counting trips them up. The same kind of model now scores 96% on PhD-level science questions, where PhD experts score about 70%.
AI reads words in chunks, not letters, so counting letters is genuinely hard for it. It's a spelling blind spot, not proof it can't reason.
We found no dedicated benchmark of current frontier models on letter counting alone.
Read the full check →Best independently measured score, by model release date. Multiple choice, 4 options.
What this measures: 198 hard biology, chemistry and physics questions written by PhDs. Experts in the field get about 70% right; smart non-experts with Google get 34%; guessing gets 25%.
Saturated: Top scores (about 96%) are near the test's ceiling, where errors in the answer key cap what's achievable, so differences between today's best models are mostly noise. Now watch: Humanity's Last Exam.
Independent runs by Epoch AI, plotted at each model's release date (many runs happened later). The expert baseline comes from a study OpenAI ran with recruited PhDs; the original paper reports 65% on the full question set. Questions have been public since 2023, so training-data contamination can't be ruled out, and scores above ~95% may be near the benchmark's error floor.
| Date | Label | Value |
|---|---|---|
| 14 Mar 2023 | GPT-4 (0314) | 36% |
| 6 Nov 2023 | GPT-4 Turbo | 42% |
| 29 Feb 2024 | Claude 3 Opus | 47% |
| 13 May 2024 | GPT-4o | 49% |
| 20 Jun 2024 | Claude 3.5 Sonnet | 54% |
| 12 Sept 2024 | o1-mini (high) | 62% |
| 17 Dec 2024 | o1 (high) | 77% |
| 31 Jan 2025 | o3-mini (high) | 77% |
| 24 Feb 2025 | Claude 3.7 Sonnet (thinking) | 79% |
| 25 Mar 2025 | Gemini 2.5 Pro | 84% |
| 5 Jun 2025 | Gemini 2.5 Pro (06-05) | 85% |
| 9 Jul 2025 | Grok 4 | 87% |
| 13 Nov 2025 | GPT-5.1 (high) | 88% |
| 18 Nov 2025 | Gemini 3 Pro | 93% |
| 19 Feb 2026 | Gemini 3.1 Pro (high) | 94% |
| 5 Mar 2026 | GPT-5.4 Pro | 95% |
| 13 Aug 2026 | Gemini 3.7 Flash (high) | 95% |
| 2 Sept 2026 | Gemini 3.8 Flash (high) | 95% |
| 3 Sept 2026 | GPT-6 Astra (max) | 96% |
| 30 Sept 2026 | Gemini 4 Argon (high) | 96% |
| PhD experts in the field (69.7%) | 70% | |
| Skilled non-experts with web access (34%) | 34% |
Best monthly score on TrackingAI's verbal IQ-style quiz that has never been published online. Human average = 100.
What this measures: A short IQ-style pattern quiz that has never been published online, so models can't have memorised it. 100 is the human average.
Monthly maxima, dated to the end of each month. IQ norms are designed for humans and aren't validated for AI; this is a small matrix-style quiz, not a measure of general intelligence. On the public Mensa Norway test (which may be in training data) scores are higher, about 150.
| Date | Label | Value |
|---|---|---|
| 31 May 2024 | Claude 3 | 85.5 points |
| 30 Nov 2024 | o1 | 96.5 points |
| 30 Nov 2025 | Grok 4 | 123 points |
| 31 May 2026 | GPT-5 Pro | 127 points |
| 31 Oct 2026 | GPT Astra Ultra (TrackingAI label) | 136 points |
| Human average (100) | 100 points |
Models are trained to predict the next word, but researchers who look inside them find planning ahead and multi-step reasoning. They also find that a model's own account of its reasoning often doesn't match what it actually did.
AI is trained to predict the next word, but studies of its insides show it plans ahead and reasons in steps. Its explanations can mislead.
This is one company studying its own models (and this site is partly built with AI models from the same company).
Read the full check →In May 2026 an OpenAI model disproved an 80-year-old Erdős conjecture, and nine mathematicians checked it. AI has now solved 17 of 68 significant open Erdős problems, 3 of them under the strictest test rules.
AI disproved an 80-year-old conjecture and designed antibiotics that worked in mice. Real finds come where answers can be checked.
Most of these results came from systems built around a model (search, automatic checkers, thousands of attempts), not a chatbot answering once.
Read the full check →Big news
An unreleased OpenAI model found a counterexample few experts expected. Mathematicians checked it the same day; one called it AI's best yet.
n^1.014pairs exactly one unit apart among n points, more than Erdős predicted
The model was unreleased; compute and any human involvement were not disclosed in the sources we could read.
Read the full entry →Share of 68 significant open Erdős problems a model solves with a machine-checked proof, in one attempt with a ~$300 budget.
What this measures: Unlike exam-style tests, nobody knew the answers to these problems when the test was made. A model only scores if the Lean proof checker accepts its proof or disproof.
Plotted at model release dates (the benchmark launched 1 September 2026, so earlier models were evaluated retroactively). Outside the official one-attempt rules, 17 of the 68 problems have now been solved by some means, including 12 credited to OpenAI's 6 October research release. Contamination risk grows as solutions are published.
| Date | Label | Value |
|---|---|---|
| 23 Apr 2026 | GPT-5.5 | 0% |
| 3 Sept 2026 | GPT-6 Astra (pre-release) | 2.9% |
| 29 Sept 2026 | GPT-6.1 Sol | 4.4% |
The length of software tasks AI agents can finish went from about 4 minutes in 2023 to about 12 hours for public models in early 2026. Returns from scaling up pretraining alone did slow.
AI kept improving fast through 2026, from 4-minute tasks in 2023 to 12-hour ones. Just making models bigger did slow, and real use lags.
Time horizons are measured mostly on software tasks, and METR's task set changed in 2026.
Read the full check →Models do get worse if trained only on their own output, again and again. Keeping real data in the mix avoids it, and AI kept improving.
Shumailov has argued that keeping real data reduces the effect at increasing cost rather than removing it.
Read the full check →METR found AI roughly doubling Anthropic's pace, a level the scenario reaches in March 2027. Its lead author says the team is ‘spooked’.
~2×AI-driven speed-up of Anthropic's research, Oct 2026 (METR)
METR's figure covers one company, rests partly on interviews and redacted evidence, and could be as low as 1.3×. Its "2× faster progress" and the scenario's research multiplier are not the same measure; matching them is Greenblatt's reading, not METR's.
Read the full entry →How fast each measure moved compared with the AI 2027 scenario, as graded by its authors. 1.0 = exactly on the scenario's pace.
What this measures: AI 2027 is a detailed 2025 forecast in which AI that out-codes the best humans arrives inside a lab by March 2027. A bar longer than the dashed line means reality is moving faster than the scenario on that measure.
These grades stop in August 2026. Since then, METR's audit (5 November) found AI roughly doubling Anthropic's research pace, a level the scenario reaches around March 2027, so the “AI speeding up AI research” bar is out of date. GPT-6 Astra and Claude Fable 5.2 are not graded yet.
Paces are the AI Futures Project's own grading (Feb and Aug 2026) and are sensitive to how each metric is measured. They are the scenario authors' estimates, not independent ones.
| Item | Value | Note |
|---|---|---|
| Global AI data-centre spending | 1.76× | ~$850B/yr vs ~$510B predicted (Aug 2026) |
| Leading-lab revenue | 1.43× | Anthropic $47B annualised (May 2026) |
| AI chips in the world | 1.14× | ~24.5M H100-equivalents vs ~22M predicted |
| Length of tasks AI can do (METR) | 1.05× | Claude Mythos Preview, Apr 2026 |
| Using a computer (OSWorld) | 0.84× | 60.8% vs 65% predicted, Aug 2025 |
| Top lab valuation | 0.7× | $852B vs $1.3T predicted (Mar 2026) |
| Largest training run | 0.58× | authors' estimate |
| AI speeding up AI research | 0.31× | authors' Aug 2026 estimate; METR's Nov 2026 audit points higher |
| Coding benchmark (SWE-bench) | 0.18× | 74.5% vs 85% predicted for mid-2025 |
| Public concern about AI | 0.1× | Gallup, 4-month average to Jul 2026 |
The famous “95% of AI pilots fail” came from 52 interviews and was labelled preliminary. Measured effects are mixed but moving up: in 2025 METR found developers 19% slower with AI; with newer tools, its 2026 follow-up estimated 18% less time.
The 95% rests on 52 interviews and a 153-person survey with a strict test of success. It points to a real problem, not a measured rate.
The report does describe a real problem: custom AI tools built for companies often fail to reach everyday use.
Read the full check →With late-2025 AI tools, returning developers took 18% less time and new recruits 4% less. Neither is certain; METR expects more.
−18%task time with AI, returning developers (range −38% to +9%)
Both confidence intervals include zero, so the trial alone does not prove a speed-up.
Read the full entry →Google's 75% counts AI-suggested code that engineers approved. Independent studies, using stricter definitions, find much lower shares.
25% → 75%Google's reported AI share of new code, Oct 2024 → Apr 2026
Google and Microsoft don't publish a method; "AI-generated" includes accepted autocomplete suggestions and is hard to fact-check.
Read the full entry →In 2025, 16 skilled developers took 19% longer with AI while feeling 20% faster. With late-2025 tools, a follow-up found 18% less time.
+19% → −18%task time with AI: early-2025 tools (2025) vs late-2025 tools (2026)
Both 2026 follow-up intervals include zero, so the newer trial alone does not prove a speed-up. METR is redesigning the study.
Read the full entry →Myths run in both directions. About 60% of the robots in Beijing's 2026 half-marathon were remote-controlled, and AI has not solved the version of Navier–Stokes most mathematicians care about.
Some do, but often a person steers them remotely without saying so. About 60% of robots in a 2026 Beijing race were remote-controlled.
Teleoperation is not deception when it is disclosed. 1X says remote sessions also teach NEO new tasks.
Read the full check →OpenAI's proof needs a designed outside force pushing the fluid, which the prize rules allow. A fluid left alone is still an open question.
The Clay problem statement genuinely allows forcing, so calling it a loophole is a value judgment. Clay has said its evaluation will be deliberately slow, and no prize has been awarded.
Read the full check →