Physicist Steve Hsu says he has published a peer-reviewed theoretical physics paper whose main idea came from GPT-5
Steve Hsu / @hsu_steve :
Context & Ripple Effects
Hsu’s claim places GPT-5 in the more consequential part of research work: proposing a testable idea rather than merely drafting or coding. The reported peer review supports the paper’s publication, but does not independently establish how much of the underlying insight came from the model.
It also precedes related coverage in which a user used a newer model for a 60-year-old Erdős problem, suggesting that AI-assisted discovery is being tested across formal disciplines, with human evaluation still central.
First-order effects
- Hsu can cite a peer-reviewed theoretical-physics result while attributing the paper’s main idea to GPT-5, creating a concrete—though self-reported—example of model-assisted scientific ideation.
- The immediate evidentiary distinction matters: peer review assesses the submitted physics work, not necessarily the model’s role in originating its central idea.
Second-order effects
- Researchers and journals may face greater pressure to separate AI assistance in hypothesis generation from authorship and to document how model outputs were checked before publication.
- If similar cases accumulate, AI labs will have a stronger benchmark category beyond coding: whether models help experts generate ideas that survive domain review. GPT-5’s reported practical software-engineering gains show the broader push toward task-specific utility.
Third-order effects
- The larger shift is from models as productivity tools toward models embedded in research workflows, where independent verification—not fluent output—determines their scientific value.
- That shift could make provenance, reproducibility, and credit allocation recurring governance questions for AI-assisted research, especially when a model’s contribution is hard to reconstruct after the fact.
The trend: AI models are increasingly being evaluated by whether expert users can turn their suggestions into independently validated work, not just by benchmark performance.