We gather and link. We don't grade.

In short

• Every fact links to its source, with the date it was published or read.
• What was measured is kept apart from what was claimed. A maker's own score is labeled as the maker's, however good it looks.
• Stages are computed from the evidence by the written rules below. Nobody sets one by hand, and nobody can pay to change one.
• Quotes are word for word, from a page we fetched ourselves, with the person named and dated.
• AI helps find and summarize material, is labeled where it wrote something, and never sets a stage, a score or a quote.
• Anything that fails a check is left out, not guessed at.

Measured, claimed, forecast

Colour carries meaning across the whole site. Blue is what was measured: an independent test, a study with real people, a regulator's decision, a court record. Apricot is what was claimed or predicted: a maker's own numbers, a launch, a named person's forecast. Lilac is what forecasters and prediction markets currently expect. We never mix the three in one number.

Where facts come from

Independent tests

Research groups that test AI systems themselves and publish their methods, such as Epoch AI, METR, the UK AI Security Institute and the Remote Labor Index, plus benchmarks published in research papers with a human baseline. We show only data we are allowed to show.

Research

Papers from arXiv (labeled as preprints that haven't been peer reviewed), studies in people from PubMed and journals, and registered trials from ClinicalTrials.gov.

Makers and launches

Companies' own announcements, model cards and release notes. These are how we know what launched and what the maker says it can do, and they are always labeled as the company's own claims.

Regulators, courts and incidents

Government sources in the US, the EU, the UK and China: laws, rules, approvals (for example the FDA's list of AI-enabled medical devices) and enforcement actions. Court records and documented incidents come from the courts themselves, the AI Incident Database and named news outlets.

News

A fixed list of established news organizations and science publications. We link to what they report; we don't rewrite it as our own.

Forecasts and prices

Crowd forecasts are read from the platforms named beside them (for example Metaculus), with the date we read them. Share prices are TradingView's delayed prices, shown in TradingView's own widgets and never stored by us.

How stages work

Each task on Can AI do it yet? sits on a six-step ladder. A stage says how far the evidence has got. It does not say whether the AI is good, safe or worth using. The highest stage whose rule the evidence meets is the one shown, and each task page names the piece of evidence that placed it.

1. 1
   ClaimedA company we follow says, in a passage we checked word for word, that its AI can do the task.
2. 2
   LaunchedA product that does the task is available to the public. We link it and date it.
3. 3
   Independently testedSomeone other than the maker has tested it: an independent benchmark, a peer-reviewed paper, or a hands-on test by a trusted outlet.
4. 4
   Tested with real peopleA randomized trial or field study has tested it with real people doing the task.
5. 5
   Regulator-clearedA regulator has cleared or approved it for the task, where approval applies (for example the FDA for medical AI). Tasks no regulator covers are marked “No approval process exists”.
6. 6
   Widely used with a verified recordIt has been tested with real people or cleared by a regulator, at least 10% of US adults use it according to a published survey, and at least three independent studies in people exist.

Track record

Beside the stage, never folded into it, each task shows its track record: documented failures, court cases and regulator actions linked to it, each with its source. A task can be widely used and still have a long record of things going wrong. We show both.

Quotes and predictions

A quote appears only if we fetched the page ourselves and the words appear on it exactly; the person is named, with their role, where they said it and when. We quote briefly and never change the wording. Predictions on Superintelligence watch are quoted the same way, with the date they were made, so they can be checked against what happened.

The role of AI

We use Claude, an AI model made by Anthropic, to find candidate material, point to the passages that matter on pages we fetched ourselves, and write plain-English summaries. Anything it points to must appear on the source page word for word before we use it. AI summaries restate the source without judging it, and they never set a stage, a score, a count or a quote. That we use one AI company's model is a fact you should weigh: it gets no special treatment in our coverage, and its maker's claims are labeled as claims like everyone else's.

Advertising and independence

Nobody can pay to change a stage, a score, a quote, our coverage or the order of a list. Sponsored content, if we carry it, is always labeled and kept apart from everything else. Books by the site's founder are marked as such and get no special placement. Advertising and sponsorship.

Corrections

If something is wrong, tell us through the contact form or email [email protected] with the page and what should change. We correct errors and note the change on the page.