Your AI Report Needs a Receipt Drawer
KPMG pulled an AI usage report after apparent hallucinations. EFF keeps finding fake staff quoted by AI slop sites. The fix is boring: citations, owners, and a human who checks the receipts.
There is a very specific kind of embarrassment that only AI can produce: a polished report with fake scaffolding.
The sentences sound expensive. The charts look board-ready. The conclusion is tidy enough to survive a steering committee. Then someone checks the citations and the floor gives way.
That is roughly where KPMG landed this weekend, according to TechCrunch. The firm pulled a report about AI usage after apparent hallucinations were found in it. The funny part, if you are feeling uncharitable, is that the report was about AI. The less funny part is that this is exactly how weak verification becomes a reputation problem.
This is not only a consulting-firm problem. EFF published a separate piece this week about a site quoting nonexistent EFF staffers. Sarah Chen, Javier Morales, Caitlin Chin, Emma Rodriguez, and Mikko Kopponen were presented as EFF experts. EFF says they do not exist.
That should make every team using AI for research a little uncomfortable.
The model is not the source
A lot of AI-assisted writing fails at the same point. People treat the model output as if it is the work product, then use review as a light polish step.
That is backwards.
The model can draft. It can summarize. It can give you a list of places to check. Sometimes it can help you notice the obvious thing you missed because you were tired. Fine. Useful.
But it is not a source. It is not a citation database. It is not a person who interviewed anyone. It is not a lawyer, journalist, auditor, or analyst just because the paragraph sounds like one.
If the final document makes a factual claim, somebody needs to know where that claim came from.
Not "the AI said so". Not "it was in the draft". A real link. A filing. A transcript. A vendor advisory. A dataset. A named person. Something with a trail.
AI makes bad process look productive
The dangerous part is not that models hallucinate. Everyone knows that by now, at least in theory.
The dangerous part is that they make a missing process feel like speed.
Before AI, a weak research workflow was visibly weak. You had a half-empty notes doc, vague bullets, and a deadline breathing down your neck. Now the same workflow can produce a confident 18-page report in the house style. It feels done earlier than it is.
That is the trap.
A good AI workflow should leave more evidence behind, not less. If a team uses a model to draft a report, the review trail should be boring and inspectable:
- claims that need evidence are marked before publication
- every source link is opened by a human
- citations are checked against the sentence they support
- quotes are copied from the original source, not regenerated from memory
- names, dates, amounts, CVEs, laws, and product versions get a separate fact check
- one person owns the final factual integrity of the piece
None of this requires a grand AI governance program. It requires a receipt drawer.
Slop is a security issue too
For Andri's usual world of security and privacy, this matters beyond content quality.
Bad AI research can send defenders toward the wrong patch. It can invent a vulnerable version. It can cite a fake expert. It can turn a real privacy problem into a mushy claim that nobody can act on. It can also pollute search results with confident nonsense, which then gets summarized by the next tool, which then becomes input for the next report.
That loop is how slop becomes infrastructure.
The fix is not to ban AI from writing. That is not serious. People are already using it, and in many cases they should. The fix is to stop pretending that fluent text is evidence.
If you publish AI-assisted work, keep the receipts close enough that an annoyed reader can check them. If you cannot show where a claim came from, cut the claim. If the model gives you a name you did not already know, assume the person is fictional until proven otherwise.
A model can help you move faster. It should not get to lower the burden of proof.