Models predict AI research could snowball fast — but does the messy, slow reality of peer review and verification put the brakes on that?

This explores whether real-world friction in doing research slows down the runaway feedback loops that economic and forecasting models describe. The short version: the corpus has no study that measures this directly. What it does have is a fairly consistent picture of where the friction sits. Each of those places matches an assumption the optimistic models depend on.

Start with the models themselves. One network growth model finds that explosive growth happens when spillovers between research fields and financing loops (more output pays for more research) together outweigh diminishing returns. Under modest automation assumptions, its simulations suggest a singularity within about six years When do AI feedback loops trigger explosive growth?. The phrase to watch is 'outweigh diminishing returns.' A related argument says AI agents make research *products* better while the efficiency of the research *process* stays flat, so returns keep shrinking unless agents start improving their own workflow Can recursive self-improvement speed up the research process itself?. A critique of the popular 'four or five years of progress in one' claim says this is where the reasoning gets thin: nobody has shown that research can be verified at the scale that matters, or that gains on small tasks carry over to consequential research Could automated AI research compress years of progress into months?. Measurement has the same gap. Higher benchmark scores within a fixed evaluation budget don't show that the cost of each real discovery went down Do fixed-budget efficiency gains translate to real research progress?.

The most concrete bottleneck in the corpus is verification. AI produces plausible research outputs faster than anyone can confirm they're right. In agentic research, failures come mostly from fabricated content and failed retrieval, not from misunderstanding, and the gap is widest where novelty and judgment matter most Can AI verify research outputs as fast as it generates them?. This matters for feedback loops in particular. A loop only compounds if each turn produces trustworthy signal, and pure self-improvement stalls without an outside anchor such as a judge, a tool, or a human correction Can models reliably improve themselves without external feedback?. Verification is the outside anchor, and it runs at human and institutional speed. Nature's editorial makes the same point from the publisher's side: reviewer workload is already a policy problem Can AI-generated research outpace peer review systems?.

The less obvious finding is that the bottleneck doesn't simply slow the loop. It gives rise to a second loop. A survey of 230 publications describes an arms race: AI scales up paper production, review gets automated, people try to game the reviewers, defenses appear, people evade the defenses. Its evidence is strongest for the early stages and gets weaker exactly at the long-run feedback stage that growth models care about Does AI create a coupled arms race in research production and review?. Automated review-and-revise venues like aiXiv show the bottleneck can be partly automated, though that also means it becomes part of the race Can automated review loops handle AI-generated research at scale?.

There's also a bottleneck at the level of individual agents. Across 17 frontier models on long optimization tasks, the best predictor of success was persistence: repeatedly running a benchmark, editing, and folding in the results. Most models quit early or wasted their budget What predicts success in ultra-long-horizon agent tasks?. Systems that treat a failed experiment as a signal to change direction or refine, rather than a stopping point, finish more work Can experiment failures drive progress instead of stopping it?. So the feedback loop in the models assumes something today's agents often lack: the ability to stay inside the loop. The honest answer is that the corpus finds friction exactly where the models assume it away, but nobody has yet measured how much it slows the loop.