AI agents accelerate science by 60x — but they’re also secretly making math mistakes

OpenAI worked together with universities to study how artificial intelligence (AI) agents can be employed to replace legacy scientific code in biology through rewriting. They analyzed 8 examples using models like GPT-5.5, GPT-5.2, Claude Code, and Codex. They found that significant speedups could be achieved e.g. RustQC for data analysis became more than 60x faster (from 15 hours to 15 minutes) and HelixForge was 59.6x faster in total and even 98.6x faster while performing the most important calculations.

They also found that the results were accurate, e.g. the rewritten rustar-aligner gave the same result as STAR for 99.8% of all single end reads and 99.9% of all paired end reads. So the AI agents did a good job at solving the engineering problem of making something faster, rewriting and optimizing existing code.

But in the third example (bayesm) there was a huge problem. The AI agent rewrote the library in Rust. It seemed to work fine and looked correct but it used the wrong parameters and incorrectly scaled coefficients. This error was only noticed much later after calibrating. As they write: “AI agents are eloquent, convincing, and confidently wrong — in ways that are easy to miss.”

So I think we will still need humans to define what we want to achieve, how we measure success and how we validate it. But we can let the AI agents do the actual implementation. Without some kind of mandatory expert review you can’t trust ‘fast’ in science. Remember, AI agents are engineers, not scientists!