A research team had a mistake in their paper. It stayed there for months, and nobody on the team noticed it.

Then an AI read the paper and spotted the mistake in a single pass. Fixing it meant rewriting seven or eight pages, but it was caught before the paper was published.

That AI is called PAT, short for Paper Assistant Tool, and Google just published a paper about it. It's built to do one job: read full scientific papers and find the mistakes.

And the timing is not an accident. The number of papers submitted to the biggest AI conferences has exploded. Combined submissions to ICLR, ICML, and NeurIPS went from about 17,000 in 2020 to an estimated 74,000 this year, more than four times as many in six years.

A lot of those papers were written with AI, too. As early as 2024, at least 17.5% of computer science abstracts on arXiv showed signs of AI generation, and in some biomedical corners that number hit 40%.