Overview

The SAGE study’s abstract describes preregistered experiments in which selected LLMs evaluated 171 archived new-venture pitches using varied scientific, contextual, and social cues. Their outputs were compared with fundraising outcomes, post-campaign survival, and human-investor judgments. The models reportedly mirrored fundraising outcomes strongly, but survival accuracy largely reflected prior digital-footprint exposure rather than reasoning under uncertainty. The abstract supports this boundary, not unreported model names, prompts, effect sizes, or protocols.

Learning outcomes

Matching fundraising outcomes is useful, but it is not a pure measure of venture quality or long-run execution.
The abstract links much of the apparent survival-prediction accuracy to prior digital exposure, not proven reasoning under uncertainty.
Scientific, contextual, and social cues can improve evaluations while also activating anchoring or herding.

Keywords

Generative AILLMsVentureFundraisingSurvivalExposure

When the Evaluator Is the Model

The ifo experiment changed the sender's pitch and watched human receivers interpret it. This chapter turns to the other side of the relationship: an LLM evaluating early-stage ventures. The question is not whether a model can produce a plausible investment opinion. It is what kind of evidence its apparent accuracy represents—fundraising-outcome matching, genuine reasoning about survival, prior exposure to a venture's digital traces, or a judgment shaped by the cues placed in front of it.

What the venture-evaluation study compares

The SAGE study's abstract describes preregistered experiments in which selected large language models generated investment evaluations and survival predictions for an archival dataset of 171 new-venture pitches. The researchers systematically varied the information cues supplied to the models. They compared the generated evaluations with three reference points: realized fundraising outcomes, post-campaign survival rates, and evaluations from a benchmark sample of human investors.

That design puts the LLM in a different role from the senders in the ifo experiment. In the earlier experiment, a candidate or entrepreneur used ChatGPT while a human receiver tried to infer expertise. Here, the model is the receiver. It reads venture material and produces an evaluation that can be checked against outcomes or compared with human judgments.

The difference changes what “good performance” means. A sender-side tool can help someone communicate a capability. An evaluator-side model must decide what to notice, how to weight cues, and how to translate incomplete information into a prediction. The output may be a ranking, an investment assessment, or a survival forecast, but the visible confidence of the text does not reveal which information produced it.

sequenceDiagram participant V as Archived venture pitch participant C as Information cues participant L as LLM evaluator participant O as Observed outcomes participant H as Human-investor benchmark V->>L: Provide venture material C->>L: Add scientific, contextual, or social cues V-->>L: Prior digital footprint may be available L->>L: Generate investment and survival evaluations O-->>L: Compare with fundraising and later survival H-->>L: Compare with human-investor evaluations

The sequence shows three different comparisons, not one master score. Fundraising is an observed event. Survival is a later outcome under uncertainty. Human investors are a benchmark of judgment, not an unquestionable ground truth. A model can resemble one reference point more than another, and the reasons for that resemblance matter.

The evidence boundary is especially important in this study. The material available supports the abstract's stated design and conclusions, but not unreported model names, prompts, effect sizes, subgroup analyses, or full methodological details. The article can therefore explain what the study claims to compare and why the distinctions matter, but it should not invent a more detailed protocol around the abstract.

Two forms of apparent accuracy

The abstract reports that the LLMs showed strong capabilities in mirroring real fundraising outcomes. This means their evaluations aligned well with what happened in the fundraising record. That is potentially useful.

Published by DeepMox

FAQ

What did the SAGE study have LLMs evaluate?

It had selected LLMs generate investment evaluations and survival predictions for an archival dataset of 171 new-venture pitches, with scientific, contextual, and social cues varied. Researchers compared the outputs with fundraising, survival, and human-investor benchmarks.

Why doesn’t accurate survival prediction prove reasoning under uncertainty?

It does not prove it because apparent survival accuracy largely reflected prior exposure to some ventures’ digital footprints rather than reasoning from limited information. A prediction can match a later event without showing transferable explanation.

Can extra information improve an LLM’s venture evaluation and still bias it?

Yes. Scientific, contextual, and social cues can improve evaluations while also activating anchoring or herding, so more context is not automatically more reliable judgment.

Previous9 minGenerative AI, Language, and Unequal Gains

Next10 minGenerative AI Screening: How to Validate Signals

Generative AI as a Venture Evaluator

When the Evaluator Is the Model

What the venture-evaluation study compares

Two forms of apparent accuracy

FAQ