AI dictionary
Resources / AI dictionary / Wrong-subject claimSIGN IN ↗

21 / 22 · How do you check its work?

Wrong-subject claim

noun

A statement where the AI copies a figure or fact correctly but gives it to the wrong person or policy. Checking the figure alone passes it.

After a joint annual review with Ruth and Graham Ellison, an administrator works from the AI's meeting notes, approved by the adviser. The notes put a workplace pension of about £62,000, new to the file, under Graham's name. The administrator adds it to Graham's record and sends the provider his signed letter of authority. Ten days later it comes back. The provider holds no plan for Graham.

The pension is Ruth's, as Graham said: "Ruth's got about sixty-two thousand in her work pension". The AI copied the figure exactly, under the speaker's name. In a joint meeting, money comes up as "we", "mine" and "hers", and the speaker is often not the owner. The adviser had checked the £62,000, and it was right. Had the letter not bounced, Ruth's record would lack the pension and Graham's retirement income would include it.

That is a wrong-subject claim. The subject is the person or policy a figure or fact belongs to, whoever said it. A wrong-subject claim is a statement in which the AI copies a figure or fact correctly from its source but gives it to the wrong person or policy. A transfer value recorded against another of the client's plans is one, and so is smoker status on the wrong life.

The fact-find gives each client a column and joint holdings another, but records whatever owner it is given. The file check ticks figures against their source. It was built for copying from statements, which carry the owner's name, and for slips like a transposed digit. A wrong-subject claim passes it, because the figure matches the source exactly. A harness can show the line each figure came from. Whose it is still has to be read from that line, or checked by a step built into the product.

Not to be confused with

  • Calling it a hallucination suggests the AI made the figure up. The figure is real and in the source, and only the person or policy it is given to is wrong.
  • Calling it a transcription error suggests the source was wrong, such as a transcript that credits a line to the wrong speaker. In a wrong-subject claim the source is right, and the AI gives the figure or fact to the wrong subject.
  • Claim, in this entry, means a statement the AI puts forward as true. It has nothing to do with an insurance claim.

Overheard

AdministratorGraham's letter of authority came back. The provider has no plan for him.

AdviserThe £62,000 is Ruth's. It's a wrong-subject claim, so I'll check whose name sits on every other figure in those notes.

Where it sits

Wrong-subject claim

Grey words are still being written.

Copy this entry

Paste it into Word, Google Docs, Notion or Apple Notes. Headings and links come with it.

Sources

Last reviewed 28 SEPT 2026

  1. Analyzing and Evaluating Faithfulness in Dialogue Summarization ↗Wang et al., EMNLP 2022 · 21 OCT 2022
  2. Checker experiment results v1 ↗Chunkbase, financial-advice-ai · 20 SEPT 2026

Your feedback

Could you explain wrong-subject claim to a colleague now?