Signal From Sameness – 94% of AI-Assisted Ideas Overlap, and Tailoring Lost Half Its Payoff

In a study published in Nature Human Behaviour in 2025, nine people working separately with ChatGPT gave the same toy the same name. None of them saw the others' work. They were using the same tool on the same task.

Meincke, Nave and Terwiesch at the Wharton School found that 94 percent of the ideas produced with ChatGPT's help shared overlapping concepts. The tool narrowed the range of ideas across the whole group.

The Same Brief Produces the Same Argument

The finding reaches well past toy names. A founder raising a round, a fund manager in front of LPs and a vendor answering an RFP now start from the same models, the same prompts and the same public data. What comes back reads well, and it reads like every competitor's. Lucas Mittelmeier of Heal Capital, a Berlin health-tech venture fund, wrote in October that "since most decks are AI generated, they all look the same."

Tailoring Stopped Paying

Hiring data shows what that does to the reader. In a 2025 paper, Cui, Dias and Ye studied applications before and after an AI cover-letter tool was introduced. Afterward, the link between how closely an application was tailored and whether it earned a callback fell 51 percent.

A second paper, by Galdin and Silbert using Freelancer.com data, found employers paid more for workers who sent customized applications before large language models arrived, and stopped paying that premium after. In their simulation of a market where the writing no longer signals anything, the strongest fifth of workers were hired 19 percent less often and the weakest fifth 14 percent more often.

Tailoring used to be expensive, so a tailored argument told the reader something about the person behind it. Once a prompt can produce it, it tells the reader nothing, and the reader looks for what a prompt cannot produce – the operator's record, a number nobody else has, and an argument built for this one reader and checked before anyone else checks it. That is where the signal went.

More Submissions, Made the Same Way

The pile grew as the arguments converged. Tony Corrigan, a tender evaluator who also sells AI bid-scoring software, wrote in May that in one recent evaluation about 30 percent of submissions were entirely AI-generated and another 40 percent largely so. Competitions that drew three or four bids in 2023 now draw 12 or 15.

Loopio's 2026 benchmark of more than 1,500 response teams found 79 percent now use generative AI, up from 68 percent. The same benchmark put the average win rate at 39 percent, against 45 percent across 2019 to 2026, and the shortlist rate fell from 54 percent to 46 percent. Loopio does not attribute the decline to AI, and only 13 percent of teams blamed their losses on proposal quality.

Capital shows the same crowding. The PitchBook-NVCA Venture Monitor counted 8,625 first financings in the US through the third quarter of 2026, already past 2021's full-year record of 8,507. Cheap to Build, Hard to Fund – Capital in the Age of AI traced the cause in M004. AI made building cheap and made funding harder.

Investors on the Record

DC Palter, an angel investor, wrote in October that at one pitch day at least ten of fifteen decks were made with AI, and he rejected every one of them. "If you see ten, they all look the same." Leslie Feinzaig, founder and managing director of the Graham & Walker venture fund, wrote in August that "when you build your deck with AI, you sound exactly like everyone else."

Palter does not call the AI versions bad. He rates them better than about half of what founders build by hand. His group invests in the top 2 percent, and in his view a prompt does not get a company there.

Every Reader Now Lines Them Up

Allocators and buyers work with the same tools. A survey by the Emerging Allocators Association, reported by CIO in July, found about 97 percent of the LPs surveyed use at least one AI tool at work, most often to draft memos and investment committee notes and to summarize documents. G2's March survey of 1,076 software buyers found their most common use of AI chatbots was comparing vendors' strengths and weaknesses, and 69 percent ended up choosing a different vendor than they first planned.

Five managers or five vendors making the same argument in the same order give the reader nothing to choose between. The comparison falls to whatever differs.

What Readers Fall Back On

When the arguments match, the decision moves to what cannot be generated. In the Angel Investment Network's 2026 global survey, the two fastest routes to a rejection were a founder who could not explain the problem and a founder with no evidence of demand, each named by 23 percent of angels. Neither failure is a writing problem, and no prompt fixes either one.

Corrigan describes the same retreat in procurement. When submissions read alike, he argues, the rational response is to weight known suppliers even more heavily. Every tender he evaluated this year, he notes, was still read and scored by people.

The Tool Raised the Floor

None of this argues against the tools. In the Cui, Dias and Ye data, access to the cover-letter tool lifted callbacks slightly, and the weakest writers improved their letters twice as much as the best ones. Palter rates AI-built material above half of what he sees. The cost does not fall on any single user. It falls on the market, where everyone's work improved at once and became harder to tell apart.

There Is Nothing Artificial About Intelligence described the tool in U002 as statistical, a system that predicts what comes next. Given the same brief, it predicts the same thing for everyone who asks.

Signal Is What a Prompt Cannot Supply

The difference is in what goes in. A prompt cannot supply an operator's record, a proprietary number or an argument built for one reader. It cannot check whether a brand name is already owned, whether a market figure is the one every competitor also found, or whether a competitor is already making the same argument. That checking is diligence, and it now reaches well beyond the financials.

The tools lifted everyone to the same level. Signal is whatever stands above it.