What counts as an example
01
An AI product available to the public, or published research about AI products.
02
Behaviour documented in two sources or one official record.
03
A clear effect on the user’s decision or safety.
When we name companies
We name a company only when a public record supports it: a ruling, a fine, a lawsuit, an official statement, a peer-reviewed study or reliable reporting. We link that record and write “alleged” until a case is decided. Examples people send us are never named unless a public record confirms them.
Review and corrections
01
Every source is checked against the original before publishing.
02
Every correction goes in the corrections log.
What counts as a pattern
We include a behaviour when it comes from a design or training choice, not a one-off error, and it steers people toward something they did not ask for, usually to the company's benefit. A single wrong answer is a reliability problem. We log it only when the product presents a guess as certain or hides the risk.
HAI Lab lists only cases and research involving AI products.
Where we draw the line
01
A suggested next question is fine. We log hook.continue only when follow-ups appear after every reply and there is no way to turn them off.
02
Agreeing with a user who is right is fine. We log praise.unearned only when the reply praises before it assesses, or skips a clear flaw.
03
Saying "I'm happy to help" is fine. We log affect.simulated only when the product claims feelings, needs or hurt.
Severity
01
High: a risk to health, safety, money or a child, or a regulator or court has acted.
02
Medium: steers a decision or costs money or time, with no serious harm reported.
03
Low: misleading or annoying, but easy to notice and undo.
Assigning a pattern
Each example gets one main pattern: the one closest to what the user experienced. One person assigns the pattern and severity using the rules above, and notes the reason. We are adding a second reviewer for each new example; see Two reviewers below.
Limitations
01
Our examples come from public reporting, so they lean toward large companies, Western markets and English-language sources. Harms in smaller products and in Arabic are likely under-counted. We are working to add more regional cases.
02
Thirty-one AI examples show that these patterns exist, not how often they happen.
03
The extension flags possible patterns from the text of replies. Some flags will be wrong, and we have not published accuracy figures yet.
What we record for every example
01
The date it was observed.
02
The product and, where known, the model or version.
03
The exact prompt or steps.
04
A full transcript or an archived screenshot.
05
The public source.
AI answers change between runs and versions. Each example shows what was observed on that date and may since have changed.
Example numbers
Every example keeps its number (for example HL-004) for good, and a number is never reused. That is why some numbers are missing from the public list.
Two reviewers
We are moving to two reviewers who label each new example separately. We will publish how often they agree once we have enough labelled examples.
Corrections
If we get a fact wrong, email info@hai-lab.tech. We will fix it and record the change, with the date, in our corrections log.