The Well Is Poisoned. We Built ForIntel So Yours Isn't.

Three of the four largest professional-services firms on earth have now retracted or rewritten AI reports because the AI made things up. Not a junior analyst's footnote error. Not a transposed figure. Fabricated court quotes. Invented McKinsey statistics. Case studies about real companies that the real companies say never happened.

This is the business we are in, and it is worth being precise about what just occurred, because the lesson is not "AI is unreliable." The lesson is that most people producing intelligence have no idea how to make AI reliable, and they are publishing anyway.

Three firms, three retractions, eight months, one pattern

Start with the timeline, because the cadence is the story. The reports were published across 2025. What landed inside eight months, from October 2025 to June 2026, was the reckoning for all three.

In July 2025, Deloitte Australia's government compliance review for the Department of Employment and Workplace Relations went up on the department's website. The contract was worth A$440,000. The report contained a fabricated quote attributed to a Federal Court judgment, with the judge's surname misspelled, and references to academics who had not written the cited work. Dr Christopher Rudge, a health and welfare law researcher at the University of Sydney, found the fabrications and took them to the press. A corrected version appeared in October 2025, with the invented quotes and references removed and Deloitte's use of AI disclosed, and Deloitte refunded part of the fee to taxpayers.

EY Canada published Points of Attack: Uncovering Cyber Threats and Fraud in Loyalty Systems in late 2025. On 14 May 2026, the AI-detection company GPTZero published an analysis of it. In GPTZero's own words, "almost all of the URLs are broken or fake, and more than half of the titles don't correspond to real sources," across twenty-seven references in all. GPTZero put the hallucinated share at 60% of those references in the account it gave Sherwood News, which had the investigation first, while the published citation table names eight of them individually. Dead links and generic tag pages stood in for named articles at Forbes, McKinsey, Gartner, TechCrunch, WIRED and others. One market-sizing figure was attributed to a McKinsey report that McKinsey never produced, and GPTZero traced that invented citation to an earlier finance blog post carrying the identical fake reference. EY pulled the report the same day and opened an internal review.

In June 2026, KPMG pulled a flagship report titled Total Experience: Redefining Excellence in the Age of Agentic AI, published the previous October. Of its forty-five citations, only five pointed to real, intact sources. UBS, the UK's National Health Service, Swiss Federal Railways, and Transport for London each told the Financial Times that the report's claims about their AI work were untrue or misleading. One case study credited Emirates with a mobile chatbot called Sara that talks to passengers and changes their flights. GPTZero marked the chatbot claim false, the talking-to-passengers claim partly true, and the flight-changing claim false: Sara is a robot assistant Emirates introduced in 2023 rather than a chatbot, and it cannot alter a booking. Another cited a 2019 railway press release as evidence of agentic AI, half a decade before agentic AI was available to buy, and the press release does not mention AI at all. The report also contradicted KPMG's own 2025 CEO Outlook, putting the share of CEOs who rank AI as their top investment priority at 55% where KPMG's own survey said 71%.

A report selling AI to enterprise clients was pulled because the report itself was hallucinated by AI. The irony writes itself, and outlets from the FT to The Register used it. But irony is not the interesting part.

"Vibe citing" and the second-hand hallucination

GPTZero caught the EY and KPMG reports. The Deloitte Australia fabrications were found by Dr Christopher Rudge at the University of Sydney, working alone and without a detector. An engineer at GPTZero coined a phrase for the failure mode: vibe citing. It is the citation equivalent of vibe coding: a generative model stitches together fragments of real sources, invents plausible titles, paraphrases a reference until it is no longer the reference, and produces something that looks exactly like scholarship and is hollow all the way down.

The sharper warning is the one that should keep anyone in this industry awake. Fake information in a report from a trusted firm, GPTZero wrote, can "poison the well" for every researcher who comes after. GPTZero's CEO, Edward Tian, has a name for what follows: second-hand hallucinations, fabrications laundered into the next report, the next model's training data, the next analyst's "according to KPMG." The lie does not die when the PDF is pulled. It has already been cited, screenshotted, and absorbed.

That is the actual stakes. Not one embarrassing retraction. A slow contamination of the shared record that everyone downstream trusts.

Here is the part the Big Four got exactly backwards

GPTZero is a detector. It is very good at what it does, and it caught two of these firms after the fact. But detection is an autopsy. By the time GPTZero runs its check, the report is published, the client has read it, the FT has the spokesperson on the phone, and the well is already poisoned.

There is, conspicuously, no public preventer. No one is owning the narrative of catching this before the PDF is published.

That gap is the whole reason ForIntel exists, and it is why our entire architecture looks nothing like "ask a model to write a report."

What we actually do differently

Every firm above had the same working method: prompt a capable model, get fluent prose, lightly review, publish under a trusted brand. The brand was doing the verification work the process skipped. That is the trap: fluency reads as accuracy, and a model that does not know an answer will produce a beautifully formatted wrong one with total confidence.

ForIntel is built on the opposite assumption: the model is the least trustworthy part of the work, so everything around it is built on not trusting it. Concretely:

Every quantitative claim traces to raw data, not to model output. When a ForIntel brief says an import is concentrated 71.3% on one origin, that number is recomputed from the underlying line items, not paraphrased from a summary the model wrote. We have caught our own intermediate steps producing inflated figures, and the discipline is that the recomputed-from-source number wins, every time, even when it is less dramatic.

We read a finding against a second, independent measure wherever one exists. A value-based concentration is cross-checked against a physical-weight concentration. A search-demand read is cross-checked against labor-market or filing data. When two unrelated instruments point the same way, the brief says the finding is grounded. When they disagree, the brief reports the disagreement and either carries the finding at a lower confidence or drops it, rather than quietly keeping the tidier number. And where no independent second measure exists at all, the brief says so in its own limits section rather than implying a corroboration it does not have.

Statistical significance is claimed only where it is tested. If a trend isn't significant, we say "flat." If a sample is too thin to weight, we say "directional, not deep-corpus." We do not dress an underpowered signal as a confident one, which is precisely the move that turned a KPMG survey number into a contradiction with KPMG's own data.

Boundaries are named, never filled. When a data source returns empty, the ForIntel brief says so, "this layer returned no data on every attempt," and fabricates nothing to cover the hole. The Big Four's failure was the inverse: a gap in the data became an invented case study about a real company. We would rather show you the edge of what we know than manufacture confidence past it.

The challenge to a finding is a separate step, by design. One part of the system builds the read. A different part works from the same raw data, never sees that read, and exists only to attack it. Whatever it comes back with is printed in the brief next to the finding, before a client ever sees either. Detection after publication is what GPTZero sells. Publishing the challenge alongside the finding is what we deliver.

The pitch, stated plainly

Three of the four biggest intelligence brands in the world just demonstrated, in public and inside eight months, that prestige is not a substitute for engineering. The verifier catches it after the damage. We are the engineering layer that catches it before, because we assumed from day one that the model would lie, and we built everything downstream to keep that lie from reaching you.

The well is poisoned. We would like to show you what clean looks like under the hood.

FAQ

Why did Deloitte, EY, and KPMG pull their AI reports?

Each report contained material fabricated by a generative AI model and published without adequate verification. Deloitte Australia's government compliance review, published in July 2025, included a fabricated quote attributed to a Federal Court judgment and references to academics who had not written the cited work. A corrected version and a partial refund followed in October 2025. EY Canada's loyalty-fraud report carried twenty-seven references, and GPTZero found almost all of its URLs broken or fake and more than half of its titles matching no real source, putting the hallucinated share at 60% and naming eight of the references individually, among them a market-sizing figure attributed to McKinsey that McKinsey never produced. EY pulled the report in May 2026. KPMG's report, pulled in June 2026, had only 5 of 45 citations pointing to real, intact sources, with named organizations stating their described AI work was wrong or invented.

What is 'vibe citing'?

Vibe citing is a term coined by an engineer at GPTZero for the citation equivalent of vibe coding: a generative model stitches together fragments of real sources, invents plausible titles, and paraphrases references until they no longer match the original, producing something that looks exactly like scholarship but is hollow underneath. The output reads as authoritative because fluency is mistaken for accuracy.

What is a second-hand hallucination?

A second-hand hallucination is a fabrication from one report that gets laundered into the next one, cited by another analyst, absorbed into a model's training data, or repeated as 'according to KPMG.' The error does not disappear when the original PDF is retracted, because it has already been quoted, screenshotted, and propagated. This is how a single fabricated claim slowly contaminates the shared record everyone downstream relies on.

How does ForIntel prevent AI hallucinations in its reports?

ForIntel treats the model as the least trustworthy part of the work and is built around not trusting it. Every quantitative claim is recomputed from raw source data rather than paraphrased from model output. ForIntel reads each finding against a second, independent measure wherever one exists, and where the measures disagree the brief reports the disagreement instead of keeping the tidier number. Statistical significance is claimed only where it is tested. Data boundaries are named rather than filled when a source returns empty. And a separate step works from the same raw data without seeing the read, tries to break it, and whatever it finds goes into the brief next to the finding before a client sees either.

Isn't detecting AI hallucinations enough?

Detection happens after publication: it is an autopsy. By the time a detector flags a fabricated report, it is already published, read by the client, and quoted elsewhere. The well is already poisoned. Prevention means verifying before the report is published. ForIntel's architecture is built to catch fabrication before publication, not to diagnose it afterward.

ForIntel is a research-first business-intelligence product from Foragentis. Every quantitative claim in a ForIntel brief traces to source data and carries the confidence the evidence actually supports. Before publication, a second, independent read of the same data tries to break the finding without seeing how it was reached, and whatever it comes back with is printed in the brief next to the finding. Where two measures of the same thing disagree, the brief reports the disagreement instead of keeping the tidier number. To see a sample brief or commission a read, reach the ForIntel desk at forintel@foragentis.com.