Main idea

When an AI generates notes from sources and links them to one another, the error it makes in one note does not stay in that note. The links carry it: a conclusion about a product opportunity rests on three notes, which themselves rest on others, and the false note is somewhere in that chain with nothing to flag it.

That is the difference with an error in an isolated document. A false document gets read, and the reader has a chance of stumbling on it. An error that has been linked has already become a premise: it arrives at the end of the chain as an established fact, stripped of whatever would have allowed it to be doubted.

The concrete result is defending an opportunity before a committee on the strength of false information you have no way of spotting, because nothing in the network says where it came from or what share of the analysis rests on it.

Why it matters

This shifts the question of the quality of AI-produced notes toward the question of how they are assembled. A low error rate per note stays acceptable as long as the notes are read one by one; it stops being acceptable as soon as they chain together, because the probability that a chain contains at least one error grows with its length.

It also gives the reason why traceability — which note rests on which, and on which source — is not a tidying convenience but the condition for being able to retract.

Nuances and limits

Propagation is not specific to AI: a false human note that has been linked propagates in the same way. What AI changes is the volume, and therefore the number of chains no human has read.

And a chain of notes can also correct: a later note can contradict the false one, provided the contradiction is registered rather than added as one more link.

Open questions

• What trace would a note have to carry for a retraction to reach everything that rested on it?