AI talk.
Independent evidence.

Over the past decade no claim has spread through venture capital faster than the assertion that a startup is built on artificial intelligence. Once the mark of a few genuinely technical ventures, it is now a near-default across the technology sector. That creates a problem for investors, who decide under acute uncertainty and lean on a few observable signals, because a claim in a venture's own description is among the most immediate of those signals and among the cheapest to fabricate.

The question

When a technology becomes fashionable, its language becomes cheap to adopt. So the thesis asks: how do AI narratives in startup descriptions relate to the funding outcomes ventures achieve, and how did the information content of those narratives change as AI claiming became ubiquitous? It builds directly on the weak-signal perspective of Tumasjan, Braun and Stolz (2021), who show that venture capitalists price social-media sentiment (a weak, nearly costless signal) into valuations even though it does not predict eventual success, whereas patents (a strong, costly signal) predict both. The step taken here moves from an audience-generated signal to the harder case of a self-reported one, authored by the party with the strongest incentive to embellish.

The data

The analysis draws on proprietary PitchBook data, used under a confidentiality agreement: two linked extracts covering venture characteristics and self-reported business descriptions on one side, and individual financing events on the other. After transparent, pre-specified sample rules, the analytical sample is 376,366 technology ventures founded between 2010 and 2024, linked to 869,172 completed financing events. Deal-outcome narration fields are removed entirely to preclude outcome leakage, and thin registry-style placeholder descriptions are excluded from all narrative analyses.

The method: a two-source taxonomy

The methodological cornerstone is to measure AI two independent ways and cross them. A single keyword flag cannot tell a venture that merely invokes AI language from one whose AI activity is externally recognised. Crossing self-report against an independent classification separates talk from substance.

1. Self-reported framing

   A dictionary of AI terms scans every business description; a venture is coded as AI-claiming if it contains at least one term. This is the fast, transparent step the illustrative text scanner demonstrates. A construct-validity check confirms the dictionary tracks meaning, not noise: flagged descriptions score nearly twice as high on similarity to an AI reference corpus.

2. Independent classification

   PitchBook's analyst-assigned vertical and industry fields, generated independently of the free text, code whether a venture actually operates in AI. An alternative classifier built from a distinct set of fields agrees for 90.9% of ventures and reproduces the central result.

3. Cross into four groups

   Crossing the two binary indicators yields the four mutually exclusive groups above. This crossing is what lets the thesis ask not merely whether AI framing is rewarded, but whether investors reward the label or the underlying capability, and at which stage.

4. Pricing, progression, survival

   An event-study linear-probability model traces the pricing premium year by year; graduation to institutional VC is modelled as a linear-probability model and a logit, then as a Cox duration and competing risks; survival is a firm-level model. Identification is stress-tested with coarsened exact matching, a doubly-robust AIPW estimator, an Oster selection bound, a specification curve, and a blockchain placebo.

Results

Prices pool; the funnel sorts

At the moment of valuation, the three AI groups command statistically indistinguishable premia over non-AI ventures, with close point estimates and wide, overlapping confidence intervals. Yet realised graduation from an entry-stage round to institutional venture capital separates the same groups cleanly by substance, against a non-AI baseline graduation rate of 33.4%.

Valuation premia carry wide, overlapping confidence intervals (prices pool). Graduation advantages separate monotonically by substance and are each significant at the 1% level (outcomes sort). Fixed effects for sector, country, and cohort throughout.

The result is robust, and triangulated

Because the graduation separation is the thesis's central claim, it is re-estimated under three methods that rest on different identifying assumptions. They converge to within one percentage point: the fixed-effects linear-probability model gives 19.1 points, coarsened exact matching on common support gives 18.4, and the doubly-robust AIPW estimator gives 18.6. Quiet AI ventures also graduate faster and fail less at once: at 36 months they reach 36.7% graduation against 27.7% for non-AI ventures, while showing the lowest prior-failure incidence, 2.4% against 7.4%. The two competing outcomes move in the same favourable direction, the opposite of what a differential-exit explanation would predict.

The signal decays with ubiquity

The AI pricing premium was large when claiming was rare (on the order of 20 to 25% larger early-stage rounds in 2017 to 2019), then eroded, turning significantly negative by 2022 and vanishing before the late-2022 generative-AI surge. The decay appears in real outcomes too: the quiet-AI graduation advantage falls from 30.1 points among pre-2017 entrants to 16.0 points in 2017 to 2019 and 11.8 points in 2020 to 2021, declining a significant 2.54 points per entry year, while the substance ordering is preserved in every phase.

Intellectual honesty: the hypothesis that did not survive

Why it matters

For investors, paying more for an AI claim now means paying more for a label that no longer distinguishes the ventures that will progress from those that will not, while the information that does discriminate, independent classification of substantive AI activity, sits underused at the pricing stage. For founders, the value lies in the substance rather than the claim: the ventures that fared best were not those that talked most about AI but those whose activity the market could recognise independently of anything they said. The mechanism is general; AI is simply the fashionable claim of the moment.