Because of AI

Method

Readers of this site have been burned by AI hype. So the registry earns credibility by what it leaves out, by publishing what did not work, and by dating every benefit by when it happened rather than when the paper came out.

The four tests

An entry starts "Because of AI, …" only if all four hold:

  1. It reached people. A benefit arrived for actual humans, not a benchmark score or a demo.
  2. It was measured against something. A control arm, a prior period, a human benchmark, another model. The comparison is named on every entry.
  3. The AI did something that would not otherwise have happened. The attribution grade says how cleanly we can tell.
  4. We know when the benefit occurred. Not when the result was published.

A result that passes the first three but produced no benefit goes to Not because of AI, under the same grades. A candidate that has not yet passed them goes to Not yet, with a note on what would change that.

Three grades, never one score

Every entry carries three separate grades. They are a vocabulary, not a formula. They are never added up, and no entry has a single score.

Evidence

Attribution

The cleanest counterfactual is where no human alternative existed: tuberculosis X-rays with no radiologist to read them, sperm no technician could find. Bundled and AI-versus-AI results are real results that say less about AI than they appear to. We label them rather than hide them.

Window

The window is the twelve months ending on the registry's reference date, currently 2026-10-22. Window grades are editorial and change only through a reviewed edit, so an entry can sit at the edge of the window for a short while before it is re-graded. Pre-window entries stay on the main list with a muted badge: the result is real, it is just not new.

Achieved is not published

This is the registry's most common mistake, and the one we work hardest to avoid. Every card shows the date the benefit was achieved next to the date the result was published, at the same weight.

The worked example is GOLDEN BRIDGE II, a cluster-randomized trial across 77 Chinese hospitals published in the BMJ in March 2026. The trial is superb. Its patients were enrolled between January 2021 and June 2023. So the benefit happened years before the paper, and the entry is graded pre-window, not in-window. We check methods sections and trial registries before trusting a headline date.

How entries get here, and what is and is not automated

An automated research run scans the literature every morning and a deeper one runs weekly. They read primary sources, apply the four tests, and may draft an entry; a drafted entry is published the same day once the data checks pass, and carries an "unreviewed" note in its history until the editor has read the source and confirmed or changed the grades. What is never automated: the grade vocabulary itself, the decision to keep or remove an entry after review, and the corrections readers send in.

The cautionary case is an automated index that collected thousands of entries in a couple of months by having a language model read and score papers. Unrelated studies in it carry identical sub-scores, line for line, because the scorer produces the same shape of answer whatever it is shown. An index like that looks dense and reads as evidence. It is neither.

Our rules, for the run and for the editor alike: a primary source for anything we put a number on, an absolute effect next to every relative one, company dashboards admitted only with the conflict stated and the comparison understood, press releases treated as leads rather than evidence, and no entry without a date the benefit actually occurred.

Submit or dispute an entry

Both go through the public repository. Open an issue to propose a new entry (bring the primary source and the achieved date) or to dispute an existing one (name the entry and the claim you disagree with). Every change to an entry is recorded in its history and in the changelog.