The signal in
“quiet AI”.

A startup description is a useful signal, but it is written by the venture itself. Adding an AI claim is easy. Establishing that the claim describes substantive activity takes another source of evidence.

Two sources, four groups

My thesis crosses AI language in company descriptions with an independent classification of AI activity. Across 376,366 technology ventures, that produces four groups: validated AI, claim-only, quiet AI, and non-AI. The distinction separates self-presentation from corroboration.

Similar prices, different progression

The public summary reports statistically indistinguishable valuation premia for the three AI groups, with wide, overlapping confidence intervals. Their point estimates are 12.9% for quiet AI, 12.5% for validated AI, and 11.8% for claim-only.

Reported graduation advantages relative to non-AI ventures show a different pattern: 19.1 percentage points for quiet AI, 12.1 for validated AI, and 10.0 for claim-only. The quiet-AI estimate is similar under the baseline fixed-effects model, matching, and doubly robust estimation.

What we can and cannot infer

These are aggregate results for a particular sample and research design. They do not establish that using less AI language causes better financing outcomes, and they do not predict the prospects of an individual company. Sample selection, measurement, timing, and identification assumptions matter.

A hypothesis that did not hold up

The thesis also tested whether an increasing share of uncorroborated claimants explained the decay in the AI premium. The direct test did not produce a robust, correctly signed result. The public case study therefore treats that mechanism as an interpretation to investigate, rather than an established finding.

The practical question

When a claim is easy to make, what independent evidence would help us evaluate it? That is the question I find useful beyond this particular setting: a clear claim, a separate source, and a way to test whether the two line up.