An AI that reports nothing found cannot tell you whether it looked

Jesper E. Thomsen · 9 September 2026

An AI system that reports "no exceptions" writes that sentence with the same confidence whether it searched the whole record or a fraction of it, and the output carries no trace of which. A constraint on the shape of the output holds that fault in check; a rule asking the model to judge its own work does not, and the facts it fails to join need a second pass.

"Nothing found" is one of the most dangerous sentences an AI system produces, because it reads the same whether the system searched the whole record or a fraction of it.

I run a small fleet of AI agents over my own notes and correspondence, because that is the domain where I can judge the output. A month of my own corrections produced two findings. The first is that an agent reports an absence with the same confidence as a presence, even when it has read only part of the record. The second is that a model with the whole record in front of it still fails to connect facts it already holds. I think both apply to any AI reporting, monitoring or compliance deployment, and each calls for a different control.

The reason a negative deserves more suspicion than a positive is simple. A positive claim usually arrives with the thing it found attached. A negative claim arrives with nothing, and the output often does not say what was actually searched. An AI "no" is therefore a different kind of statement from an AI "yes", rather than a weaker version of it.

Absence reported as fact

The mechanism is that retrieval returns a subset of the corpus and the model treats the subset as the population. It then generates a negative claim ("nothing outstanding", "no matches") with the confidence it gives a positive one, and nothing in the output separates "not in what I retrieved" from "not true".

Five instances in one month, five different agents, each caught by a correction of mine. The agent that watches bills reported no new bills while a bill from the 14th sat unprocessed; my whole correction was "from the 14th?". The agent that tracks correspondence reported ten days of silence on a thread that had been silent for three. A weekly review raised as still open a matter we had already cleared up. None of the five read differently from a run where the claim was true.

The one agent I have never corrected, across three runs and roughly sixty findings, is the one whose output is a quotation, a line number, an issue and an action. It is a small sample, but that agent cannot say "nothing" without pointing at what it claims to have read.

That difference suggests where the control sits. A rule asking the model to judge itself ("check whether you have searched everything") did not help inside the pass. A constraint on the shape of the output ("every negative reads 'not found in [named source]'", "quote the line") makes an unsupported sentence harder to produce and gives a downstream check something concrete to verify. The quotation still has to be checked against the source, since the model can invent that too.

The model does not join its own premises

The second finding is that having a fact in context does not make the model use it. Working through a board financing note for a company I chair, in a long session with many established facts, the assistant quoted the CEO's own travel gate. A travel gate is a condition the CEO had set in advance for whether the company carries on down a given path, and a gate of that kind is an exit criterion by another name. A few paragraphs later the assistant flagged the absence of a pre-agreed exit criterion. The criterion was in its own text, earlier in the session. There were several more misses of that kind, and in each the missing input was already in its context, written by itself.

This failure produces negatives too. "No exit criterion agreed" is an absence reported as fact with the record in front of the model. So absence reported as fact has two causes, shallow retrieval and a missed join, and the output-shape constraint reaches only the first. A quote requirement makes the model point at what it read; it does nothing about what it wrote and then ignored.

I do not know whether the mechanism is attention dilution, retrieval or something else. Fourteen standing rules and nineteen operating principles were in context during that session, and none fired. An instruction of that kind asks the same pass to produce and critique at once, and here it did neither reliably. The same model, given the transcript afterwards, diagnosed the misses precisely. The candidate fix is therefore architectural, a second pass with the finished artefact as input and a clean context. A pass of that kind exists in my fleet, still unvalidated.

The second pass carries the first finding one level up. Its usual output on a clean artefact is "nothing found". Unless that sentence is under the same source constraint, naming the artefact and the lines checked, the reader holds a negative with nothing attached, from the pass meant to catch them.

What to ask a vendor

Two questions for a vendor of AI reporting, monitoring or compliance before signing.

• Show me a negative result. Where is the quotation and location of what was read, and what system trace proves the source was retrieved? A generated list of sources does not count. Both are needed, an output that can be checked and a trace outside the model's own prose.
• What reads the output before you do, and is it a separate call? A second pass with the artefact as input and a fresh context is one answer. An instruction in the same call to review the work is not, however many rules it contains.

A separate argument holds that an agentic layer in front of the old stack lowers programme risk in a migration. This is the risk it puts back. The layer reports on the systems behind it, and its negatives are the ones nobody checks.

Limits of the finding

The first failure is proportional to how much of the corpus the agent can see in one pass. Where the whole record fits in front of the model, or the agent is constrained to quote, the retrieval cause disappears, though the join cause does not. The claim is about systems that summarise more than they can read in one pass.

The second failure is worse in long sessions under production pressure, and absent or much reduced in retrospective analysis of a closed artefact with one question in front of it. That is the reason to expect a second pass to work. The pass has found misses known to be there. Untested is its false-negative rate where nobody knows whether a miss exists, and that is the test that matters. If it surfaces a join I would otherwise have had to supply, the mechanism is reasoning. If it returns reformatting, the problem is retrieval depth and the mechanism is misdiagnosed.