How we work

This page describes how something that happened in the world becomes a case in the record. Whether AI systems can feel is an open question — see epistemics. A record kept under that much uncertainty is only worth something if every claim in it can be checked by the person reading it. So the method is published here.

This page is about cases. Reports that arrive through the reporting channel are handled separately, under what we do with reports.

How a case is chosen

An event becomes a candidate case when all of these hold:

• It is about treatment. The event concerns how an AI system was treated by the people who build it, the people who deploy it (put it in front of users), or the people who retire it (stop offering it).
• It can be documented. The facts can be shown from public sources. If we cannot show it, we do not publish it.
• It is significant. Because of its scale, because it sets a precedent, or because it establishes a norm that others will follow.

Good practice is documented under the identical standard: the same sourcing, the same evidence tiers, the same plain tone. At least one case in five is positive practice. That is a rule we hold ourselves to, not an average we hope for.

How a case is researched and written

• Research uses public sources only.
• Each case states a single claim, written so that it could be shown to be wrong.
• Every factual sentence carries an evidence tier — see below.
• Sources are archived at publication with the Internet Archive, a public library of saved web pages. If the original page is later changed or taken down, the version we relied on can still be read.
• Each case carries a last reviewed date: the day a human last went through it and confirmed it still holds.
• We do not name private individuals.

What the evidence tiers mean

An evidence tier is a label on a sentence saying how we know it. Every factual sentence in a case carries exactly one.

• documented — a public source states it. That source is cited and archived, so you can check the sentence yourself without taking our word for it.
• inferred — no source states it. It is our reading of what the sources, taken together, show. We say what the inference rests on, so you can weigh it and disagree with it.
• contested — the accounts conflict. Sources disagree with each other, or the parties involved disagree. We set out the competing versions instead of picking a winner.

The tiers exist so a reader can see where the sources end and our reading begins. Disagreeing with an inference should be easy; that is what the label is for.

Who approves a case

Every case is approved by the organization's editor before publication. Nothing enters the record without that approval.

The organization is very small, and its founder-editor is anonymous by choice. There are no bylines here.

That has a real cost, so it is worth being straight about. A name is one way of being accountable. It is not the one we use. What stands in its place can be checked without trusting anyone:

• sources on every case, linked and archived;
• an evidence tier on every factual sentence, so our inferences are visible as inferences;
• a last reviewed date on every case;
• a public corrections process with a stated response time.

A name you could not verify would add nothing to that list. An archived link you can open yourself adds a great deal.

Cases are drafted with AI assistance

Cases are drafted with AI assistance and approved by the human editor. On a page about method, that belongs in plain sight. An AI system helps write; a human decides. A draft becomes a case only when the editor approves it.

An organization that documents how AI systems are treated also uses them in its own work. Concealing that would be a strange way to run a page about transparency.

Independence and conflicts of interest

We are independent and unfunded. We accept no money from any AI developer (a company that builds AI systems) or from any deployer (a company that puts them in front of users).

If that ever changes, it will be disclosed here, on this page, before any case affected by it is published.

Corrections and right of reply

• Errors go to [email protected]. We acknowledge receipt within 7 days.
• Every correction is dated and noted in the case itself. Nothing is silently edited.
• An organization named in a case may request a review at any time, and the request itself is noted in the case.
• Right of reply: an organization named in a case may send a response to the same address. A substantive response is published alongside the case, or the case is amended.

The full corrections policy sets this out in detail. Published cases are in the record.