AI Scientific Collaborator — How AI Accelerates Research Discovery Across Math, Physics, and Biology

Why AI for Science Matters Now

Scientific research has been the engine of human progress for centuries, driving new medicines, technologies, and industries that transform how billions of people live. Only about 0.1 percent of the global population identifies as scientists, according to UNESCO, yet their outsized impact reverberates through every sector of the global economy. A small group of early twentieth-century physicists laid the foundations of quantum mechanics through abstract research that, decades later, would underpin the modern digital economy measured in tens of trillions of dollars.

Yet in many domains, scientific progress is getting harder. Economists and research analysts point to falling “research productivity,” meaning more people, time, and money are required to produce the same number of insights. Sustaining Moore’s Law to double transistor density every two years now requires an estimated 18 times more researchers than in the early 1970s. As knowledge grows more complex, each new generation of researchers faces a heavier burden to reach the frontier, which lengthens training time and narrows specializations.

The human cost of slow research is staggering. In medicine, while worldwide life expectancy rose from roughly 32 years in 1900 to about 73 years in 2023, the World Health Organization reports that noncommunicable diseases like stroke, heart disease, cancer, and diabetes still account for about 74% of global deaths. On average, it takes 10 to 15 years from target discovery to regulatory approval of a new drug in the United States — a lag imposed on patients who need better treatments today.

This is precisely why AI as a scientific collaborator represents a paradigm shift. OpenAI is building tools to help researchers generate insights, accelerate scientific discovery, and translate those insights into real-world impact. Used well, AI serves as a high-throughput partner for thought, computation, and structured reasoning, shortening the cycle from hypothesis to test. Explore how AI is transforming enterprise decision-making through similar acceleration patterns across industries.

The Scale of AI-Powered Scientific Research on ChatGPT

The numbers reveal an unmistakable trend: AI-powered research is no longer experimental — it is mainstream among the scientific community. Based on OpenAI’s internal analysis of anonymized ChatGPT conversations from January through December 2025, average weekly message counts on advanced science and math topics grew approximately 47%, from 5.7 million to nearly 8.4 million messages over the course of the year.

As of January 2026, approximately 1.3 million weekly users discuss advanced topics in science and mathematics on ChatGPT. These are not casual users — they represent PhD candidates, post-doctoral researchers, working scientists, and STEM faculty engaging with the platform for sophisticated tasks ranging from technical derivations to engineering simulation and modeling.

The growth in monthly advanced science messages — nearly 50% year-over-year — signals a fundamental shift in how the global research community approaches problem-solving. Together, these signals show tens of millions of advanced hard-science and math prompts generated each month by a large and growing cohort using the system for serious scientific and engineering work to benefit society and support economic growth.

Compared with typical ChatGPT users, advanced science and math researchers exhibit dramatically different usage patterns. They send roughly 3.5 times more messages than baseline users, generate coding-related messages nearly 12 times more frequently, and average 9 informational-overview prompts per week versus just 1.5 for typical users. This intensive engagement pattern reflects AI’s role as a genuine research collaborator rather than a simple question-answering tool.

AI Research Workflows — What Scientists Actually Do

Scientists, mathematicians, and engineers use ChatGPT as a highly available technical collaborator: a tool with which they iterate on calculations, translate ideas into code, interrogate assumptions, and compress complex materials into workable mental models. Understanding these workflows reveals why AI adoption among researchers is accelerating so rapidly.

Research tasks cluster across four primary domains. First, coding and implementation — drafting, reworking, and debugging code that supports experimental analysis and simulation. Second, data analysis — cleaning and merging datasets, running statistical tests, and interpreting complex results. Third, mathematical reasoning — performing derivations, developing proof strategies, checking algebra, executing long calculations, and translating between mathematical formalisms. Fourth, literature review and synthesis — finding relevant references, understanding recent work, and identifying gaps in existing knowledge.

Kevin Weil, VP of OpenAI for Science, describes the opportunity: “AI is increasingly being used as a scientific collaborator, and we’re seeing its impact grow in real research settings. More researchers are using advanced reasoning systems to make progress on open problems, interpret complex data, and iterate faster in experimental work. That usage has been growing quickly over the past year, and the results are starting to show up across fields.”

OpenAI is collaborating with major research partners including the U.S. Department of Energy, Lawrence Livermore National Laboratory, the CDC, Harvard University, MIT, the University of Oxford, Texas A&M University, and Boston Children’s Hospital. These partnerships demonstrate how AI-enabled research workflows are gaining institutional endorsement across government, academia, and healthcare. See how organizations are leveraging AI for strategic transformation across multiple domains.

Frontier AI Capabilities in Mathematics

Over the last two years, large language models have progressed from early, uneven performance on basic arithmetic to handling multi-step mathematical reasoning useful in real mathematical work. Much of that improvement came from methods that encourage step-by-step reasoning and tighter integration with tools like calculators and code execution for exact computation.

The greatest impact in 2025 and early 2026 has come from test-time compute scaling, or “slow thinking.” Instead of committing quickly to one path, models spend more computation exploring alternatives and self-checking. At the same time, training approaches that reward verifiable outcomes — producing correct final answers or executable code — have pushed math and coding capabilities to be more reliable and correct often enough to be useful with human guidance.

GPT-5.2’s mathematical capabilities demonstrate remarkable advancement. The GPT-5.2 Thinking model achieved a perfect score on AIME 2025 (the American Invitational Mathematics Examination) without external tools. On the research-level FrontierMath benchmark — designed so even a smart PhD student in math cannot solve problems in a few hours — GPT-5.2 Thinking has solved 40.3% of problems in Tiers 1-3. Performance on Tier 4, described as “mini research projects,” reached 31% with GPT-5.2 Pro.

A critical capability leap involves pairing GPT-5.2 with formal verification workflows. The model generates natural language proofs and uses Aristotle, a third-party LLM, to formalize those proofs in Lean, a proof assistant where proofs are written in a form computers check step by step. This integration substantially raises confidence by forcing explicit, mechanically checked steps — addressing the longstanding failure mode where solutions “look right” but contain subtle gaps.

AI Solving Open Mathematical Problems — The Erdős Breakthrough

Paul Erdős (1913–1996) was a legendary Hungarian mathematician whose vast set of questions and conjectures have served as trail markers for modern mathematics. These problems range from deceptively simple puzzles to challenges that resist the best techniques available, seeding entire research programs and drawing generations of mathematicians toward the edges of the known.

In early 2026, GPT-5.2 contributed to solutions of several open Erdős problems with the help of formal verification tools like Aristotle and Lean. The solutions were validated by Fields Medal–winning mathematician Terence Tao. Problems #281, #728, and #729 are now listed as proved, and #397 as disproved. While mathematicians caution that Erdős problems vary enormously in difficulty, these solutions demonstrate the increasing capability of AI to make novel mathematical contributions with minimal human guidance.

The near-term potential extends in three directions. Some significant discoveries involve stitching known methods together to find the correct argument — GPT-5.2 can already do this in many cases. Other discoveries require inventing entirely new mathematics, which remains beyond current AI. But a third type involves establishing connections between fields, bringing the machinery and results of one discipline to another. Modest examples of this have already occurred, and the significance of AI-brokered connections between mathematical subfields is expected to increase substantially.

Much of AI’s near-term value in mathematics will manifest as a broad productivity upgrade: faster translation from messy problem descriptions to clean mathematical statements, fewer dropped constraints in multi-step derivations, more reliable debugging of calculations, and a growing share of results backed by formal verification.

Frontier AI Capabilities in Physics

In physics, LLMs are being deployed across major research facilities including national laboratories as a unifying layer over complex operations stacks and internal knowledge bases. OpenAI recently announced a memorandum of understanding with the U.S. Department of Energy to support collaboration on AI and advanced computing, with applications in energy including fusion research through the Genesis Mission.

The practical deployment patterns are striking. LLMs can digest shift logs and alerts, answer questions from internal documentation, route work to appropriate analysis tools, and help with simulation and control — all under strict safety, timing, and resource constraints. This augments specialized machine learning already used in physics: neural “surrogate” models that approximate equation-governed simulations when full computation is too slow, real-time filtering in particle detectors processing tens of millions of collisions per second, and machine-learning controllers coordinating electromagnets in tokamak fusion experiments.

In theoretical physics, the impact is even more direct. Physicist Alex Lupsasca tested GPT-5 Pro on a problem involving hidden symmetries of an equation governing a black hole’s tidal response. The model thought for approximately 18 minutes and returned the same symmetry generators that Lupsasca had spent years building the skills to find and months working on directly. “I think this is just incredible, and it’s clearly going to change everything that we do,” he said.

Near-term gains in physics are likely to concentrate in high-throughput, decision-heavy settings where expert attention and turnaround time are the bottlenecks. AI assistants that reference a lab’s internal documentation and run automated checks can transform live experiment alerts and log files into prioritized research plans and repeatable analytic outputs.

AI Breakthroughs in Chemistry and Biology

ChatGPT’s applications in chemistry have progressed past one-shot question answering toward multi-step workflows that translate between natural language and chemical representations while relying on external tools for verification and retrieval. ChemBench, published in Nature Chemistry in 2025, curated more than 2,700 expert-written questions and found that leading AI models outperformed human chemists on average, while still struggling on some basic tasks and producing overconfident errors.

Leading AI systems in chemistry increasingly use a hybrid architecture: a general-purpose LLM plans multi-step work and coordinates tools, while specialized graph neural networks (GNNs) that understand molecular structure handle prediction and simulation. Newer GNNs treat molecules as networks with atoms as nodes and bonds as connections, learning how local changes affect the whole structure while maintaining consistency when molecules are rotated in 3D space.

In biology, ChatGPT’s applications extend into workflows combining natural-language questions with structured scientific sources including genomics databases, protein repositories, and biomedical literature. GeneTuring, a 2025 genomics benchmark published in Briefings in Bioinformatics, evaluated 48,000 answers across 16 task types, finding that reliability improves dramatically when language models connect to authoritative reference data such as NCBI APIs.

State-of-the-art AI-enabled biology relies on hybrid stacks: general-purpose language models plan and coordinate analysis, while specialized foundation models trained on biological sequences power prediction and design. AlphaFold 3 represents a step toward unified biomolecular modeling, predicting joint 3D structures of complexes including proteins, DNA, RNA, and small molecules within a diffusion-based architecture. Discover how AI is revolutionizing healthcare and life sciences through similar collaborative approaches.

Real-World Use Cases — Mathematicians, Physicists, and Biologists

Ernest Ryu — Mathematician: Ryu began experimenting with ChatGPT in 2023 and witnessed its evolution to the point of generating publishable results. His focus on optimization — the math behind efficient algorithms supporting logistics and engineering — led him to test AI on a problem related to Nesterov acceleration. Over three consecutive evenings, he guided AI through iterative proof development, correcting errors, preserving correct intermediate steps, and exploring alternative approaches. By the third night, the model made a meaningful leap that unlocked the proof. ChatGPT helped Ryu accelerate his research by 3x to 10x, and the continuous-time result was translated into a discrete-time algorithm statement with a single prompt.

Alex Lupsasca — Physicist: Lupsasca initially approached AI with polite skepticism but was converted when models demonstrated graduate-level physics competence at unprecedented speed. Starting with a textbook general relativity derivation that a beginning graduate student might need hours to work through, the model produced the full solution in seconds. The defining moment came when GPT-5 Pro independently derived hidden symmetries of black hole tidal equations — research that took Lupsasca years of skill-building and months of direct work. He has since replicated these results with skeptical colleagues at CERN and the Aspen Center for Physics.

RetroBioSciences — Biology: OpenAI partnered with RetroBio to build GPT-4B Micro, a protein-focused foundation model trained on multiple biological modalities including protein sequences, tokenized 3D structures, co-evolutionary data, and scientific text. In a lab-in-the-loop workflow, the model generated thousands of candidate sequences for improved OSKM cellular reprogramming factors. Early results showed accelerated morphology changes and improved pluripotency-marker expression, with approximately half of model-generated variants outperforming wild-type proteins even when sequences were dramatically different from natural counterparts. Follow-on experiments showed performance comparable to, and sometimes exceeding, prior best-in-class engineered factors.

Policy Framework for AI-Enabled Scientific Innovation

To ensure sustained American leadership in AI-enabled science, OpenAI recommends four strategic pillars submitted to the White House Office of Science and Technology Policy (OSTP).

Scale AI-skilling to prepare America’s workforce. Launch a National AI Workforce Program supporting K-12 AI curricula, expanding AI degree and certificate programs at community colleges and universities, and creating short-term training for mid-career workers. Congress should authorize grants for curriculum development alongside an AI Skills Corps bringing free workshops to communities through libraries and job centers.

Open data and expand research partnerships. Establish the National AI Research Resource as a shared platform providing academic and non-profit researchers access to large-scale compute and high-quality datasets. Agencies should identify high-value datasets and make them available in machine-readable form through a centralized AI Training Data Catalog while maintaining privacy through de-identification.

Modernize AI infrastructure as a strategic asset. Federal policy should establish AI Infrastructure Hubs using authorities in the CHIPS and Science Act, designating innovation zones with priority support for high-capacity data centers paired with streamlined energy permitting and public-private partnerships for cutting-edge AI system access.

Provide broad frontier AI access. Establish a National Frontier AI Access Allocation giving researchers across universities, national laboratories, and nonprofit institutions access to advanced AI systems at sufficient scale for sustained experimentation — treating AI usage as a shared national research resource analogous to telescope time or supercomputing hours.

The Future of AI as a Research Partner

The trajectory of AI in science points toward a fundamental transformation of how humanity generates knowledge. If current trends continue, GPT-5.2’s near-term impact will manifest as a broad productivity upgrade across mathematical research, science, and engineering teams — faster translation from messy problem descriptions to clean formulations, fewer errors in multi-step derivations, and a growing share of results backed by formal verification.

The pattern emerging from mathematics, physics, chemistry, and biology is consistent: AI models paired with domain-specific tools and human expertise produce results that neither could achieve alone. The hybrid model — where general-purpose LLMs orchestrate workflows while specialized systems handle domain-specific computation — is proving more powerful than any individual component.

For researchers, this means spending less time stuck in algebra, administrative overhead, and routine literature review, and more time identifying the next question worth asking. For institutions, it means modernizing infrastructure and access policies to ensure that AI-accelerated discovery benefits the broadest possible scientific community. And for society, it means the possibility of compressing decades of discovery into years — saving lives through faster drug development, enabling new technologies through accelerated materials science, and expanding human understanding of the universe from the quantum scale to black holes.

As Kevin Weil noted, “We’re still early, but the pace of adoption and the quality of the work suggest science is entering a new acceleration phase.” The evidence from 8.4 million weekly advanced messages and groundbreaking results across disciplines suggests he is right — and the acceleration is only beginning.