Defamation by AI output

What it is

A generative model asked about a named individual or company may invent a crime, a scandal, a bankruptcy or a regulatory finding, or attach a real one to the wrong person. If the statement is communicated to someone else and causes or is likely to cause serious harm to reputation, it may be defamatory. The organisation whose system made the statement, or which republished it in a report, an article or a customer reply, is the natural defendant. That the words came from a model is unlikely to be an answer. The data protection side is covered in inaccurate personal data in outputs.

How it happens

• A public assistant answers questions about named people from the model's memory.
• Generated research or due diligence reports are sent to others unchecked.
• An automated news or summary feature misattributes an allegation.
• People with the same name are merged.

Why it matters

Defamation claims are costly, slow and public. The harm to the person is real and may be hard to undo once a statement has spread. A system that has made one such statement about a person is liable to make it again.

Warning signs

• The system will discuss named private individuals with no source.
• Generated content about people or companies is published without review.
• No procedure exists for a complaint that an output is false and damaging.
• Statements complained of cannot be traced or suppressed.

Who owns it

The owner of the publication or channel is accountable. Legal sets the review standard and handles complaints. Engineering restricts what the system will say about named persons.

In practice

A local news site uses a model to summarise court lists. It reports that a named shopkeeper was convicted of fraud. The conviction was of a different person with the same name in another town. The same pattern generalises to any automated statement about real people that reaches an audience unchecked.