Physicists Use LLMs, Skip the Panic

Physicists have quietly absorbed the LLM revolution that is tearing through mathematics, and according to a new YouTube breakdown by researcher Sabine Hossenfelder, they are doing it without the existential panic.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with funding and a one-liner from our database.

OpenAI

An artificial intelligence research organization developing and promoting friendly AI for the benefit of humanity.

OpenAI

Private / $100B+ est

OpenAI is an AI research and deployment company dedicated to ensuring that artificial general intelligence benefits all of humanity.

OpenAI

$190.6B

An AI research and deployment company building safe and beneficial artificial general intelligence.

OpenAI

$122.0B

Artificial intelligence research and deployment company focused on developing advanced AI models like GPT-5.6 and GPT-Live, offering products such as ChatGPT, DALL-E, and Whisper.

arXiv's high-energy theory (hep-th) category ran roughly 30% above prior-year baselines in May and June. Peter Woit flagged the spike, and Hossenfelder ties it to researchers offloading derivations to frontier models rather than flooding the server with slop.

How AI physics research is actually landing

Gerardo Adesso at Nottingham used ChatGPT 5.6 to prove there is no shortcut to a class of entangled states, then verified every step himself. A separate team disproved a 150-year-old Maxwell rule about equilibrium points in collections of electric charges, crediting ChatGPT for suggesting the construction. A silicon-membrane optics group modeled laser behavior with Claude and matched measurements cleanly.

Other recent uses: Claude for numerical quantum calculations, Claude writing ion-routing software for a trapped-ion computer, and a GPT agent steering stellarator design choices in fusion research. The pattern is the same. The model drafts, the physicist checks.

Why physicists aren't panicking like mathematicians

Hossenfelder draws the contrast sharply. Young mathematicians are calling for moratoria, accusing LLMs of academic fraud, or describing spiritual crises. Physicists, she says, ask whether the model can finish the calculation by Friday.

That gap reflects a cultural split. Mathematics prizes elegant proofs that a single mind can carry from hypothesis to conclusion. Physics prizes agreement with experiment, which gives LLMs a built-in quality filter. A wrong number fails to match a detector reading. A wrong proof can still look beautiful.

What this means for builders and labs

For research groups, the playbook is already clear. Treat the LLM as a derivation engine with a human reviewer. Keep the human accountable for the final result. Document the model's role, the way Adesso's paper does, so peer review can audit the chain.

For vendors, the wedge is narrow and defensible. ChatGPT 5.6, Claude, and the unreleased OpenAI Astra are not selling novelty here. They are selling a faster path to a publishable figure.

What still needs scrutiny

Most cited examples rely on authors verifying their own AI-assisted work. No preprint from this batch has been independently retracted, but no independent audit of LLM-derived physics proofs exists at scale either. The arXiv volume bump is suggestive, not conclusive. The open question is whether the 30% surge reflects genuine productivity or a flood of lightly-checked submissions that the review system hasn't caught yet.

Frequently Asked Questions

Is AI actually doing physics research now?

Frontier models are contributing derivations, numerical work, and proof ideas that end up in published or preprint papers, with physicists verifying the results. The work is AI-assisted, not autonomous.

Which AI models are physicists using?

ChatGPT 5.6, Claude, and unreleased models like OpenAI Astra appear in recent preprints. Choice depends on the task: symbolic algebra, numerical work, or coding for lab software.

Why are mathematicians more worried than physicists?

Mathematics lacks physics's experimental referee. A wrong theorem can still pass peer review if the logic holds, so AI-generated proofs carry higher reputational risk without a natural check.