How hallucinated metrics get into AI reports and who notices last

Slide seven of the post-show board deck said onsite rebooking was 71 per cent. It was a good number, up on the prior year, and it did the work a good number does in a board meeting.

There was no query behind it. The real figure was 389 rebookings from 623 eligible exhibitors, which is 62.4 per cent. Hallucinated metrics in AI reports get discussed as though a model dreamed them up, and in this case no model was involved at all. The 71 had been typed into the report template as a sample value eleven months earlier and never replaced.

Four routes in, and only one involves a model

Sorting the ways a fabricated figure reaches a report is worth doing properly, because the fixes are different and only one of them is about AI.

A model invents a figure. Asked to summarise and given incomplete material, a language model produces a plausible number to fill the shape of the sentence. This is the route everyone talks about and, in a well-built reporting pipeline, the rarest.

A template ships a placeholder. Somebody builds the report layout before the data exists, types a realistic value so the design can be reviewed, and the value survives. The 71 above.

A synthetic seed survives into production. A deterministic test dataset is built so the interface can be demonstrated, and a fallback path renders it whenever a real query returns nothing. The report then looks complete on exactly the days when the data failed.

A rounding step runs twice. A rate of 0.6249 gets rounded to 0.625 in a transform layer, then rounded again to 0.63 for display, and appears as 63 per cent when a single rounding of the original gives 62. One point, in a figure nobody will question, produced entirely by arithmetic that was individually correct at each step.

The last three predate generative AI by decades. What changed is that a language layer over the reporting stack converts all of them into confident prose, and prose travels further than a dashboard tile and carries less of the context that would have let someone catch it.

What NIST calls this, and why the word is useful

The NIST Generative Artificial Intelligence Profile, AI 600-1, published in July 2024, lists twelve risks that are unique to or exacerbated by generative AI. Confabulation is the second of them, defined as "The production of confidently stated but erroneous or false content (known colloquially as 'hallucinations' or 'fabrications') by which users may be misled or deceived."

Two things in that definition earn their place. Confidently stated, which identifies the mechanism of harm as presentation rather than error rate. And by which users may be misled, which locates the risk in the reader instead of in the model.

The document also notes that confabulations include outputs diverging from the input, which is the distinction that separates route one from routes two, three and four. A model faithfully repeating a placeholder it was given has not confabulated anything. The system has still published a fabricated metric. Keeping those two failures under separate names stops a team from buying a better model to fix a template problem.

A rebooking rate with no query behind it

Follow the 71 per cent through.

The template was built in the previous September with sample values so the layout could be signed off. Rebooking rate, 71 per cent, alongside sample counts that were later wired to real queries. The rebooking tile was wired last, and the wiring did not land before the first edition using the new template.

At the show, 623 exhibitors were eligible to rebook onsite and 389 did, which is 62.4 per cent. The deck's 71 per cent implies 442 rebookings against 623, so the slide overstated by 53 exhibitors.

Then the summarisation layer read the deck. It wrote that rebooking strengthened to 71 per cent, an improvement on the prior edition, and that sentence went into the covering email. The sentence is grounded, in the technical sense: every figure in it traces to something the model was shown. It is also false about the world, and the grounding made it read as verified.

The commercial consequence is not abstract. Sales planning for the following edition assumed 442 renewals and built a target on the remainder. The gap of 53 accounts is the difference between a comfortable year and a difficult one, discovered in February.

Why does nobody catch it?

Because the people who could are not the people who are looking, and NIST names the mechanism.

Under Human-AI Configuration, AI 600-1 describes automation bias as excessive deference to automated systems, and states plainly that it can exacerbate other risks including confabulation. A figure that arrives inside a generated report inherits the authority of the system that produced it, and the fluency of the surrounding prose is itself a signal of care that has nothing to do with whether the number is right.

The evidence from other domains on how bad this gets is worth carrying around. Dahl, Magesh, Suzgun and Ho published a study in the Journal of Legal Analysis in 2024 testing large language models on specific verifiable questions about randomly selected federal court cases. Hallucination rates ran from 58 per cent for ChatGPT 4 up to 88 per cent for Llama 2. Their second finding is the one that transfers directly: the models could not reliably predict when they were producing a hallucination, and they often failed to correct a user's incorrect premise when the question contained one.

Both failures show up in event reporting. Ask an assistant why rebooking improved to 71 per cent and it will explain why rebooking improved to 71 per cent, because the premise arrived with the question.

Now the ordering of who notices. The exhibitor sales representative knows her own book and would say within a minute that 71 per cent is not what her territory looked like, and she has no standing to stop a deck that has already been circulated. The analyst who could check it was not sent the deck. The show director reads it as one of forty numbers that week. The board sees it last, has no independent means of checking anything in it, and is the audience whose decision it changes. Detection capability and authority to act run in opposite directions, and that ordering is a design property of the reporting process rather than a failure of any individual.

Which route is most common in event reporting?

The placeholder and the synthetic seed, and it is not close.

The reason is sequencing. Event reporting platforms get built while the warehouse is still being wired, because the show happens on a fixed date whether or not the data model is finished. So the interface is developed against sample values, demonstrated against sample values, and approved against sample values, and every one of those sample values is a live fabrication waiting for a fallback path to render it.

The tell is a report that never has a gap. Real event data has holes: a sensor that failed on the Wednesday, a channel that stopped populating a field, an exhibitor category with four members. A reporting layer that always renders a full set of tiles, for every show, in every period, is either extraordinarily well instrumented or filling gaps with something.

Making a fabricated figure detectable

The control that works is structural rather than vigilant, because vigilance decays.

Give every real figure a provenance record, then refuse to render any figure that lacks one. A tile with no query behind it displays as unavailable, in the interface and in every export, and the summarisation layer is given nothing to describe. This inverts the current default, where a number with no lineage looks exactly like a number with lineage. Building that stamp into every report line is Q5's subject and it is the single highest-value change available here.

Two cheap tests find most of what already exists. Point the reporting layer at an empty database and screenshot what renders; anything still showing a value is a fabrication route. And search the codebase for the numbers on your most recent deck, because a figure that appears as a literal in source is not a measurement.

Then keep the generated prose honest at the point of writing, which means supplying a closed list of aggregates and matching every output figure back against it, as Q6 sets out.

Where this stops

None of this addresses the fifth route, which is a figure that is genuinely computed from genuinely wrong data. A rebooking rate of 62.4 per cent calculated over an eligibility list that omits 40 exhibitors is fully stamped, fully traceable and wrong, and every control described above will pass it.

The second limit is that refusing to render unstamped figures produces gaps in reports, and gaps generate pressure. Somebody will ask why the rebooking tile is empty this month, and the honest answer, that nobody has wired it yet, is less comfortable than a number. Holding that line is a management decision more than a technical one, and it is the decision that determines whether any of the machinery matters.

The third is that counting fabrications found tells you nothing about fabrications present. Turning this into a rate you could report needs a denominator and a sampling method, which is Q8's subject.

Open your most recent post-show deck and pick the five numbers that most influenced a decision. For each one, name the table and the filter that produced it. Any figure you cannot resolve in two minutes is the place to start work on your reporting layer, and in most organisations at least one of the five will not resolve at all.