Answers are grounded in this paper and cite their sources.
Generative artificial intelligence (AI) can now produce, within minutes, a venture risk assessment that reads like the work of an experienced analyst. Whether founders, and the investors who read their documents, can still tell how much such an assessment rests on is an open question with consequences for entrepreneurship, finance and the governance of AI. This article develops a theory of cue–evidence decoupling. People judge the strength of evidence from cues such as fluency, structure and agreement, and generative AI supplies these cues whether or not the evidence is there. Because founders act when the uncertainty they perceive falls to a level they can accept, decoupling lowers perceived uncertainty without lowering actual uncertainty. It also weakens written venture materials as costly signals to investors. Because feedback in entrepreneurship is slow and noisy, founders should learn to correct for decoupling only slowly. To make the theory testable before outcomes are known, the article draws on risk science to define the knowledge-strength gap: the number of strength-of-knowledge criteria on which a founder rates an AI-assisted risk assessment as better supported than blind raters judge a normalised version of it to be. Eight propositions specify when the gap widens (novel ventures; estimating rather than listing risks), when it narrows (an AI critic; feedback on past estimates), how it raises resource commitment and how disclosure of AI use shifts investors' errors. Each proposition is paired with a randomised test and a result that would refute it. The theory implies that AI may mislead founders most where they need it most. It recasts the debate on AI and Knightian uncertainty as a question about founders' beliefs, and it suggests that safeguards for AI-assisted decisions, including transparency rules, should target the grounds of confidence and not only the accuracy of AI output.
No takes yet. Share an insight, caveat, or question.
Sulaiman et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: