- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,253
45 references in the document - and only 5 of them were real. How about that?
KPMG removed from part of its sites the report on the Agency-General of AI after accusations of mass errors in sources and facts. The document was supposed to show how new AI systems change the business, but in the end it became an example of the problem, which consulting companies have long warned customers: the text looks convincing until you start checking links.
We are talking about a report published in October 2025. Agent AI is called systems that are not limited to the answer to the question, but can plan actions, turn to tools and perform chains of tasks. For consulting, this is one of the best-selling topics in recent years: companies promise customers process automation, faster solutions and new ways to work with data.
The problem began after the GPTZero check. The researchers said they had carefully disassembled the KPMG report and found serious violations in the links. According to them, of the 45 sources, only five actually confirmed the claims to which were tied. The rest of the links either led the wrong place, either did not prove the desired thesis, either were too vague, or looked partially fictional.
GPTZero called it "discrimination" - by analogy with "vibe coding", when the code is written by the general feeling, not really understanding how it works. In the case of references, the meaning is similar: the generative model can collect a plausible list of sources from real names, similar formulations and fragments of documents. On the page, everything looks like a normal analyst, but when checking, it turns out that the facts do not converge.
According to GPTZero, about half of the actual statements in the report were false, unconfirmed or attributed to incorrect sources. Especially a lot of questions have been raised by examples of the introduction of agent AI in large organizations. The report mentioned UBS, the Swiss Federal Railways and Transport for London, but sources under these cases, according to the inspectors, did not confirm the claimed details or were retelled too freely.
One of the most understandable examples is related to Emirates. The report said the airline uses the Sara mobile chatbot, which communicates with passengers and can change the reservations. The test showed a different picture: Sara - not a chatbot, but a robot assistant, introduced by Emirates in 2023. It can really interact with passengers, but can not change flights.
Mistakes were found not only in references to external sources. In one place, KPMG’s report claimed that 55% of CEOs call AI the main destination for investment. But KPMG 2025 CEO Outlook, released in the same month, gave a different figure of 71%. For a regular article on the Internet, it would be just negligence. For a report by a large consulting firm that sells a proven analytics to a business, this discrepancy looks much worse.
After the publication of the claims, KPMG removed the report from part of the sites and began to understand how it came out at all. The company said that they check the circumstances of the publication, and employees must comply with the rules of responsible use of AI: materials should be checked manually, and independent sources must be confirmed before the text is released.
This story is particularly inconvenient for the consulting market. Large firms have been selling AI companies for several years and at the same time explain that generative models have hallucinations. They warn that AI can confidently invent facts, confuse sources and write a text that looks professional but badly withstands the test. Now similar claims were made to the report of KPMG itself.
Deloitte had a similar case. Last year, the company returned the money to the Australian government after the report paid from the budget got the AI-generated error text. Then it looked like a warning to the entire industry. The KPMG story shows that even large companies with internal control procedures can still be allowed to pass weak AI analytics into public materials.
The main problem is not that consultants use AI in the preparation of reports. By itself, AI can help to search for materials, collect drafts and speed up work with large arrays of text. The danger begins where the generated fragment is taken for the finished examination and does not check each link, figure, date and example.
KPMG removed from part of its sites the report on the Agency-General of AI after accusations of mass errors in sources and facts. The document was supposed to show how new AI systems change the business, but in the end it became an example of the problem, which consulting companies have long warned customers: the text looks convincing until you start checking links.
We are talking about a report published in October 2025. Agent AI is called systems that are not limited to the answer to the question, but can plan actions, turn to tools and perform chains of tasks. For consulting, this is one of the best-selling topics in recent years: companies promise customers process automation, faster solutions and new ways to work with data.
The problem began after the GPTZero check. The researchers said they had carefully disassembled the KPMG report and found serious violations in the links. According to them, of the 45 sources, only five actually confirmed the claims to which were tied. The rest of the links either led the wrong place, either did not prove the desired thesis, either were too vague, or looked partially fictional.
GPTZero called it "discrimination" - by analogy with "vibe coding", when the code is written by the general feeling, not really understanding how it works. In the case of references, the meaning is similar: the generative model can collect a plausible list of sources from real names, similar formulations and fragments of documents. On the page, everything looks like a normal analyst, but when checking, it turns out that the facts do not converge.
According to GPTZero, about half of the actual statements in the report were false, unconfirmed or attributed to incorrect sources. Especially a lot of questions have been raised by examples of the introduction of agent AI in large organizations. The report mentioned UBS, the Swiss Federal Railways and Transport for London, but sources under these cases, according to the inspectors, did not confirm the claimed details or were retelled too freely.
One of the most understandable examples is related to Emirates. The report said the airline uses the Sara mobile chatbot, which communicates with passengers and can change the reservations. The test showed a different picture: Sara - not a chatbot, but a robot assistant, introduced by Emirates in 2023. It can really interact with passengers, but can not change flights.
Mistakes were found not only in references to external sources. In one place, KPMG’s report claimed that 55% of CEOs call AI the main destination for investment. But KPMG 2025 CEO Outlook, released in the same month, gave a different figure of 71%. For a regular article on the Internet, it would be just negligence. For a report by a large consulting firm that sells a proven analytics to a business, this discrepancy looks much worse.
After the publication of the claims, KPMG removed the report from part of the sites and began to understand how it came out at all. The company said that they check the circumstances of the publication, and employees must comply with the rules of responsible use of AI: materials should be checked manually, and independent sources must be confirmed before the text is released.
This story is particularly inconvenient for the consulting market. Large firms have been selling AI companies for several years and at the same time explain that generative models have hallucinations. They warn that AI can confidently invent facts, confuse sources and write a text that looks professional but badly withstands the test. Now similar claims were made to the report of KPMG itself.
Deloitte had a similar case. Last year, the company returned the money to the Australian government after the report paid from the budget got the AI-generated error text. Then it looked like a warning to the entire industry. The KPMG story shows that even large companies with internal control procedures can still be allowed to pass weak AI analytics into public materials.
The main problem is not that consultants use AI in the preparation of reports. By itself, AI can help to search for materials, collect drafts and speed up work with large arrays of text. The danger begins where the generated fragment is taken for the finished examination and does not check each link, figure, date and example.
