OpenAI's GPT-5 science paper puts the human–model dialogue at the center

What OpenAI put on the table

On November 20, 2025, OpenAI released a paper co-authored with collaborators at Vanderbilt, UC Berkeley, Columbia, Oxford, Cambridge, Lawrence Livermore National Laboratory, and The Jackson Laboratory. It compiles early case studies across math, physics, biology, computer science, astronomy, and materials science.

The specific results are concrete. In a study led by Derya Unutmaz, GPT-5 identified a likely mechanism for a puzzling immune-cell change within minutes from an unpublished chart and suggested an experiment that confirmed it. Mehtaab Sawhney and Mark Sellke used it to supply a missing step in a decades-old Erdős problem. Sébastien Bubeck and Christian Coester used it to find a counterexample showing a decision-making method used in robotics and routing can fail, and to improve a classic optimization result.

OpenAI is careful to frame these as curated illustrations, not a systematic sample. That framing matters more than the individual wins.

The interface is a conversation, and using it is a skill

The document's most repeated claim is not about raw capability but about how a scientist and the model interact. OpenAI describes productive work as a loop: pose a question, push back, break the problem into steps, validate independently, iterate until a direction survives or is discarded.

Productive work often looks like dialogue—researcher and model iterating until a promising direction emerges or the idea is discarded.Montana Labs

This is a statement about the front end of the work. GPT-5 does not run the project. Scientists set the agenda, choose methods, critique, and validate. The model contributes breadth, speed, and parallel exploration. The value shows up in the exchange itself, and OpenAI explicitly calls using GPT-5 well 'a skill'—knowing when to challenge an answer and what to check by hand.

That reframes the deliverable. The product here is not an oracle that returns finished answers; it is a working surface where an expert steers, and the quality of the steering determines the quality of the output.

Failure modes that the interface has to absorb

OpenAI lists specific ways GPT-5 fails: it can hallucinate citations, mechanisms, or proofs that look plausible; it is sensitive to scaffolding and warm-up problems; it misses domain-specific subtleties; and it will follow unproductive reasoning if not corrected.

Each of these is a reason the human stays in the loop rather than behind a fully automated pipeline. A hallucinated proof that reads well is exactly the case where a validation step in the dialogue is load-bearing. Sensitivity to scaffolding means the framing a researcher provides is part of the result, not a neutral preamble.

The paper notes GPT-5 is strongest in mathematics and theoretical computer science, where structure is explicit and feedback loops are fast. That aligns with the failure modes: domains with quick, checkable feedback let the human catch bad turns early. Empirical sciences still route through the wet lab to confirm a proposed mechanism.

The implication: build the validation loop, not the autonomy fantasy

For teams applying these models, the honest read of this paper is that the near-term leverage is in the interaction design, not in removing the expert. GPT-5 shortened parts of the workflow—proof outlines in minutes instead of days, mechanism identification from a chart—but only under oversight that catches plausible-looking errors.

OpenAI gestures at a future where more time and compute yield deeper results, extrapolating from 20-minute assists to hours or days of reasoning. That may come. What the current evidence supports is narrower: the systems that get value out of GPT-5 today are the ones that make the push-back, step decomposition, and independent verification cheap and habitual for the expert driving them. The interface is the product that has to be built well right now.