Rewriting 26,000 crowdfunding pitches with AI cut the variation in length and quality by about 80%

The design is cleaner than the argument it serves. Each of the 26,000-plus crowdfunding descriptions supplies its own baseline, because every AI-augmented text is a rewrite of one specific original ^[3], so the narrowing cannot be an artefact of comparing unlike campaigns with each other ^[18]. What contracted was dispersion. Abstractness varied 43% less across the rewrites, and variation in length and in writing quality fell by roughly 80% ^[4][5]. The second figure is about 1.9 times the first ^[15], which reads as the tools compressing surface form faster than rhetorical stance: a pitch can still be concrete or lofty, but it lands at a similar length in a similar register.

This is best read as a ceiling on convergence rather than a measure of what actually happens in practice. The twins were produced by the researchers from narratives humans had already written ^[3], so nobody in the sample chose to use a model, and nobody decided how much of its draft to keep ^[17]. Real adoption is partial and edited, which puts observed homogeneity somewhere below the corpus figures. Karl Taeuscher, the lead author at Alliance Manchester Business School, makes a version of this point when he says the issue is using the tools well and knowing when to depart from what they suggest ^[13]. The paper's own caution is firmer still: a plausible scenario, not a prediction, with the crowdfunding analysis serving to illustrate the central argument rather than to test it ^[14].

What the finding leaves open is whether converged pitches raise less money. The reported measures are all properties of the text; no measure of backer or evaluator response to either version appears in the account ^[16]. And the mechanism the paper proposes runs entirely through audiences. Repeated exposure to similar AI-assisted communications makes those patterns the expected norm ^[8], the quality floor rises as polish gets cheap ^[9], and weight shifts from plausible-sounding claims toward evidence for them ^[10]. Each link is reasonable, but none of them is measured by the variance statistic drawn from the 26,000 rewrites.

Where that leaves an operator depends on which half of the paradox you sit in ^[6]. If you write badly, the tools are still a gain, which is why the leveling story took hold ^[7]. If everyone writing badly gets the same gain from the same model, the gain shows up as a higher floor rather than a wider gap, and the differentiating input becomes the cultural read of a specific audience that Taeuscher says the tools cannot supply ^[12]. I would hold that view conditionally. The compression inside the corpus is documented ^[3][5]; the market consequence is a stated hypothesis with a described mechanism and, so far, no outcome data attached ^[14][16]. The paper is in Entrepreneurship Theory and Practice, which is the right venue for an argument of that shape ^[2].