There are more and more instances popping up where AI systems are helping science and math researchers with their research.
A Michigan State University physicist has claimed a breakthrough in both AI-assisted research and quantum mechanics, publishing what may be the first theoretical physics paper where the central insight originated from an AI system. Stephen Hsu, a professor at Michigan State University, has published a research article in Physics Letters B that tackles fundamental questions about quantum mechanics—with the core idea generated by OpenAI’s GPT-5. The work examines when modifications to standard quantum mechanics violate relativistic covariance, a question with profound implications for our understanding of reality itself.
“I think I’ve published the first research article in theoretical physics in which the main idea came from an AI,” Hsu announced on X, describing his collaboration with the frontier language model.
The Physics: Challenging Linear Quantum Evolution
The paper, titled “Relativistic Covariance and Nonlinear Quantum Mechanics: Tomonaga-Schwinger Analysis,” investigates one of the most fundamental questions in physics: whether quantum evolution is exactly linear. Using the Tomonaga-Schwinger formulation of quantum field theory, Hsu derives new mathematical conditions that any “state-dependent” modification to quantum mechanics must satisfy to remain consistent with relativity.
State-dependent modifications—changes to the standard linear Schrödinger evolution that depend on the quantum state itself—turn out to be extraordinarily difficult to implement without breaking the theory’s mathematical consistency. The research shows that nonlinear modifications affect operator relations at spacelike separation, leading to violations of the integrability conditions required for the theory to work.
The implications extend beyond abstract mathematics. Whether quantum evolution is truly linear determines whether we live in an Everettian multiverse and affects the fundamental limits of quantum computing.
The AI Method: Generator-Verifier Protocol
Perhaps equally significant is Hsu’s accompanying paper describing the “Generator-Verifier” method he developed to extract reliable research insights from large language models. The approach addresses a critical challenge in AI-assisted research: frontier models can produce brilliant insights alongside serious errors.
“Models sometimes make very simple mistakes (e.g. in calculation) and also even make incorrect conceptual leaps that are superficially plausible,” Hsu writes. “The second type of error can lead even expert researchers astray, consuming large amounts of effort to detect and correct.”
His solution employs structured orchestration of multiple AI instances. One model generates a proposed step forward, while an independent instance verifies it—significantly reducing hallucination errors compared to single-pass generation. Hsu compares working with an LLM to “collaboration with a brilliant but unreliable human genius who is capable of deep insights but also of errors both simple and profound.”
The Breakthrough Moment
In screenshots Hsu shared, GPT-5 proposes the paper’s central insight entirely on its own. When asked to compare nonlinearity treatments in non-relativistic quantum mechanics with approaches informed by quantum field theory, the AI suggested using the Tomonaga-Schwinger formulation to “show explicitly why a hypersurface-local generator that depends nonlinearly on the global state cannot remain foliation-independent without collapsing back to linear dynamics.”
This suggestion became the foundation of the published research—a genuine case of AI not merely assisting with calculations or literature review, but contributing the core conceptual innovation.
Implications for Scientific Research
The development comes as the scientific community grapples with how to integrate increasingly capable AI systems into research workflows. While AI has been used for tasks like protein folding prediction and data analysis, Hsu’s work may represent a new category: AI as a source of novel theoretical insights in fundamental physics.
The Generator-Verifier protocol could prove valuable beyond physics. Any field requiring rigorous logical reasoning—from mathematics to computer science—faces similar challenges when working with AI systems that blend genuine capability with occasional unreliability.
Hsu’s dual publication—the physics result in Physics Letters B and the methodological paper on AI-assisted research—provides both a scientific result and a potential template for how experts in technical fields can productively collaborate with frontier AI systems. AI systems are already solving Erdos problems and winning gold medals at Math Olympiads. As AI capabilities continue to advance, the boundary between human and machine contributions to scientific discovery may become increasingly blurred, and Hsu’s transparent documentation of AI’s role in generating his paper’s central idea offers an early glimpse of this emerging research paradigm.