AI output is overwhelming the people who check it
arXiv's new rate limit policy is the clearest data point: 9,869 submissions in September 2016, 20,569 in 2024, 40,363 this September. In the same news cycle, Harvard particle physicist Matthew Schwartz dropped 36 papers written with Claude, and a Substack post described Opus 5.5 finding a new eyewitness record of the dodo in historical texts. Commenters on the arXiv thread split on whether the limit helps or whether it just pushes high-quality posters to other platforms. Over on the Schwartz thread, one person asked the right question: is this actually good, or just a tool that lets an LLM do a lot of numerical analysis?
The key bit: generation got cheap, verification did not. The dodo post notes that models are bad at judging the historical significance of what they find. Even HN is feeling it. An Ask HN asks for a flag on AI-generated articles, and people argue that detectors are unreliable and voters often can't tell. A thread voting on which of HN's old AI challenges have now been met shows how fast the goalposts moved.
The real results are not fake. People are finding things with these tools. The bottleneck is now a human with time and judgment.
So what?
If your product produces content, code, or research, the scarce thing you can sell is trust and review, not volume. Build provenance, verification, and rate-aware workflows now, because every platform that takes submissions is heading toward the same wall arXiv just hit.
Read these
ArXiv's Updated Rate Limit Policy
Harvard particle physicist Matthew Schwartz drops 36 papers authored with Claude
Using Opus 5.5 to discover a new eyewitness record of the dodo
Vote on which of Hacker News' challenges for AI have been met
Ask HN: Add flag for AI-generated articles