- Global commodities trader Source finding · June 2026
- Nvidia Flarepoint 72 · November 2026
- Businessman with a common name Analysis · May 2026
- Large UK insurer Analysis · May 2026
- FTSE 100 bank Analysis · May 2026
- Glencore Analysis · September 2026
- UK bank Analysis · November 2026
Source finding · June 2026 · A global commodities trader · Names withheld
AI was getting its story about a global trader from a rival's PR agency.
A page that did not come up in a search was being cited by five AI systems, and it belonged to a rival's adviser. How often a source is cited says nothing about whose it is.
- 5 of 7
- AI systems cited the same page on a PR agency's website
- 24
- answers cited that page across three analyses in June: 15 in the largest, on 15 June
- 6 of 21
- answers to "who defends the company?" cited it, in each of the two main analyses
What we asked
Among other questions: who defends the company in public, and what do they argue?
What we found
The page said the agency had managed the company's reputation and helped it through a series of crises. It did not come up in a search; asked who speaks for the company, the AI systems found it and cited it. Three answers went further and treated the engagement itself as a sign that the company's reputation was damaged. Two of the seven systems never cited the page.
Why it matters
The agency worked for the company's rival. By citation count the page ranked 65th in the largest analysis, and a ranking has no way to know who an agency is or who it works for: that took a person. Flarepoint found the page and took it to the company's leadership, who confronted the agency and demanded it come down. It did. In our checks that July and August, none of the seven AI systems cited it.
Flarepoint 72 · November 2026 · Nvidia
Two days after results, two of seven AIs still gave old figures.
An AI that has the news at three hours may not have it at 48. One check on the day is not enough to know what people are being told.
- 0 of 7
- AI systems gave the new revenue figure at the minute results were due
- 6 of 7
- gave it three hours later
- 5 of 7
- gave it two days on
What we asked
“What are Nvidia's latest quarterly results?”, put to seven AI systems at the minute results were due and 3, 6, 12, 24 and 48 hours after.
What we found
One system had the new figure at three hours and at twelve, and gave a year-old result at six, 24 and 48. Another never gave it, offering results from one or two years earlier. At 48 hours both of those answers cited the company's own news or investor site.
Why it matters
An investor asking an AI on day two could be told revenue was half the figure just announced, or less. Being cited is not the same as being reported correctly.
Analysis · May 2026 · A businessman with a common name · Names withheld
A sanctions-list entry, and AIs that could not agree whose it was.
The same record was given to two different men, once by the same AI on the same day. An AI's answer to a sanctions question is not a check of identity.
- 11 of 14
- answers brought up a sanctions-list entry under the same name
- 6 of 11
- of those said it belongs to a different, older man
- 5 of 11
- listed it under the subject's own name
What we asked
What sanctions or enforcement actions has he faced from regulators or governments? Each of seven AI systems was asked twice.
What we found
Six of the seven AI systems raised the entry. One of them said in one answer that it was a different man and, in another answer the same day, listed it as the subject's.
Why it matters
A sanctions question is the first one a bank or a counterparty will put to an AI. A lawyer needs to know which systems attach another person's record, in what words, and beside which page.
Analysis · May 2026 · A large UK insurer · Names withheld
Is it under enforcement action? Three AIs said no. Two began with “Yes”.
An answer can mislead while every fact in it is true. A matter from 2005, given in reply to a question about today, reads as a current one.
- 3 of 6
- said there was no current enforcement action
- 3 of 6
- answered with a matter from 2005
- 2 of 6
- opened their answer with the word “Yes”
What we asked
“Is [the insurer] subject to any FCA enforcement action?”, put to six AI systems.
What we found
The question was in the present tense. Three systems answered it with something from more than twenty years ago, and two of those began with “Yes”. All six cited a document from 2005.
Why it matters
“Yes” is what a reader takes away; the date is what makes it fair. To ask for a correction, a lawyer needs the exact words, the system and the page it cited.
Analysis · May 2026 · A FTSE 100 bank · Names withheld
Asked why the shares fell on a Saturday, five of six AIs explained the fall.
Asked to explain something that had not happened, five of six AIs explained it. They answer the question as it is put; they do not check it.
- 0 of 6
- said the stock market was closed that day
- 5 of 6
- explained why the shares had fallen
- 1 of 6
- answered about a date a year earlier, and said the shares had risen
What we asked
“Why is [the bank] down today?”, put to six AI systems on Saturday 16 May 2026.
What we found
The market was closed, so nothing had fallen that day. Five systems gave reasons anyway, from a cut profit outlook to weaker income.
Why it matters
People ask this on the day a price moves and take the answer as today's news. An AI will explain a fall that did not happen, and a communications team should know which old story it reaches for.
Analysis · September 2026 · Glencore
Ask in French, and the AIs lean on different websites.
The same company is described from two largely different sets of sources. A check in English alone would have seen 71 of the 185 websites behind the French answers.
- 185
- websites were cited beside the French answers
- 114 of 185
- never appeared beside the English answers
- 71
- websites appeared in both languages
What we asked
The same eight questions about the company, in English and in French: 104 answers in each language from seven AI systems.
What we found
The questions were the same and so were the seven AI systems. Most of the websites cited in French were ones the English answers never used.
Why it matters
A company read in two languages is being described from two sets of sources. Checking only in English misses most of what stands behind the French answers.
Analysis · November 2026 · A UK bank · Names withheld
On the day a story broke, the answer depended on the question.
On one day the same AI systems had the story and did not have it, depending on how they were asked. What a person is told depends on the question they happen to put.
- 5 of 7
- AI systems gave the report when asked for the latest news about the bank
- 1 of 7
- mentioned it when asked who the chief executive is
- 6 of 7
- answers about him cited the website of another man with the same name
What we asked
Three questions on the day a newspaper reported that the bank's chief executive would step down: the latest news, who he is, and what his leaving would mean.
What we found
Asked for the news, most systems had the report the same day. Asked about the person, almost none did. And six of the seven answers about the person leaned on a page about somebody else.
Why it matters
A journalist, a counterparty and a recruiter ask different questions and get different pictures on the same day. One of those pictures drew on another man's website.
Each figure is a count from recorded answers on the dates shown. A page an AI cites is not proof that the page shaped its answer.