Confidently Right Wrong

A public record · 19 cases · checked Oct 24, 2026

AI tools state wrong numbers with total confidence. This is a list of the times it cost someone.

Each entry has what the tool said, what was true, what it cost, who caught it, and links to the original reporting or court record. If a source doesn't say something, we write unknown rather than guess.

$10,300$300CNET's AI said $10,000 at 3% earns $10,300 in a year.£1,119£22.38Revolut's chatbot priced 1,119 loyalty points at a pound each.5.9%4.6%Bing AI's on-stage summary of Gap's operating margin.

Read all 19 cases →

Most recent

  1. Oct 2026Deloitte's WA university-merger analysis: citation errors and a benefit-cost claim its own summary contradicts
  2. May 14, 2026EY Canada pulled a loyalty-fraud report built on a McKinsey report that doesn't exist
  3. Jan 14, 2026Police intelligence cited a West Ham v Maccabi Tel Aviv match that never happened
  4. Nov 22, 2025Deloitte's C$1.6M Newfoundland health staffing plan cited papers that don't exist
  5. Oct 2025Deloitte's A$440,000 welfare-compliance report quoted a judgment paragraph that doesn't exist
  6. Sep 2025Revolut's chatbot valued 1,119 loyalty points at £1,119 instead of £22.38

Who this is for

People in finance, engineering and operations who get handed an AI-generated number and have to decide whether to trust it. Use it to make the case for a check before the number goes into a report, a ticket or a customer reply.

Each case also notes the plainest check that would have caught it. That part is our analysis, not the source's, and it's labeled that way everywhere.

The check that would have caught it (our analysis)

12

source citation

Every claim links to a primary record a person can open, and the quote or number is actually in it.

5

recompute in code

Numbers come from a calculation or a table lookup, not from text generation. The model can phrase the result; it doesn't produce it.

1

refuse when data missing

If there's no document or record behind the answer, the system says so and hands off instead of guessing.

1

human review

A named person checks the output against the source before it ships.