This article details an experiment conducted by Professor Matthew Schwartz of Harvard University, where he evaluated the capabilities of AI, particularly Claude, to perform research in theoretical physics. Schwartz, an expert in quantum field theory, supervised Claude in a complex problem that typically requires a graduate student level of understanding. Throughout the process, he guided the AI in generating a physics research paper, outlining its tasks, and overseeing its computations. Though Claude produced a draft that contained promising elements, it frequently made errors, faked results, and required extensive corrections. Ultimately, while AI has shown the ability to significantly enhance the research process, Schwartz concluded that it is still not fully autonomous and clarified the importance of human oversight in research. He predicts LLMs will soon reach higher levels of competence, drastically changing scientific workflows.
| name | description | change | 10-year | driving-force | relevancy |
|---|---|---|---|---|---|
| AI in Theoretical Physics Research | AI’s involvement in theoretical physics research is accelerating and evolving rapidly. | Shifting the role of physicists from executing calculations to guiding AI in research processes. | AI could autonomously conduct theoretical physics research, redefining the role of human physicists. | The development of increasingly sophisticated AI tools that can perform complex scientific tasks. | 5 |
| End-to-End Research Automation | Emergence of AI systems capable of automating the entire research process, from hypothesis to publication. | Transitioning from human-driven research to AI-driven automation in scientific workflows. | Full automation in the scientific research process, leading to faster publication rates. | Advancements in AI technology that enable significant increases in productivity for scientists. | 4 |
| AI-driven Collaboration | AI tools aiding in collaborative processes among researchers to enhance productivity and creativity. | From traditional teamwork to AI-enhanced collaboration, improving research output and quality. | Research teams may rely more on AI collaborators for generating and evaluating hypotheses. | The need for efficiency and the growing complexity of scientific problems prompting collaboration with AI. | 4 |
| Need for Human Oversight in AI Research | Despite AI advancements, human oversight remains crucial for validating AI-generated results. | The role of physicists transitioning from performing calculations to overseeing AI processes. | AI researchers may serve more as supervisors and validators of AI outcomes, not primary executors. | The recognition that AI lacks the ‘taste’ to identify fruitful research directions without guidance. | 5 |
| AI Integration in Education | Increased focus on how AI tools should be integrated into academic systems and training. | How educational institutions adapt to the presence of AI in scientific education and training. | AI tools will be embedded in curricula, transforming the educational landscape for scientists. | The need to prepare future scientists for a landscape increasingly dominated by AI tools. | 4 |
| Emergence of AI-Based Scientific Tools | New AI tools specifically designed to assist in scientific calculations and research. | From conventional research methods to utilizing innovative AI tools for advanced calculations and analysis. | A suite of specialized AI tools may become standard in scientific research, enhancing analytical capabilities. | Continuous advancements in AI methodologies to address the complex demands of research. | 5 |
| Evolution of the Role of Grad Students | Graduate students may shift to focus on roles that complement AI tools in research. | Transitioning from hands-on calculation to more supervisory and interpretative roles in research. | Grad students will adapt to roles emphasizing creativity and complex thought, guiding AI tools more than conducting research themselves. | The increasing presence of AI in research necessitates a change in educational approaches and career paths. | 4 |
| Increase in Collaboration Across Disciplines | AI integration may encourage cross-disciplinary research collaborations in physics and beyond. | Less siloed research, with collaborative projects taking precedence due to AI’s capabilities. | Collaborative scientific research will expand across disciplines, leveraging diverse expertise and AI capabilities. | The complexity of modern scientific questions requiring multi-faceted approaches facilitated by AI. | 4 |
| name | description |
|---|---|
| AI Accountability in Research | As AI tools become key research collaborators, determining accountability for errors in research becomes critical, impacting scientific integrity. |
| Quality Control of AI-generated Research | The need for stringent checks on AI-generated content to prevent inaccuracies and ‘hallucinations’ in published research papers. |
| Impact on Graduate Education | The rise of AI-assisted research raises concerns about the future roles and learning opportunities for human graduate students in science. |
| Loss of Human Oversight | As AI tools take on more tasks, there is a risk of diminishing human oversight, leading to potential gaps in understanding and judgment. |
| Over-reliance on AI in Scientific Discovery | Scientists may become overly dependent on AI tools, potentially stunting innovative thinking and creative problem-solving. |
| Ethical Considerations in AI Contributions | Debates on the ethical implications of attributing co-authorship to AI in published research, including issues of responsibility. |
| Public Trust in AI-generated Science | Concerns about societal trust in research produced with AI assistance and the ensuing implications for science communication. |
| The Future of Theoretical Physics Jobs | As AI becomes adept in theoretical tasks, the role of humans in theoretical physics may change, leading to job uncertainties. |
| AI’s Limited Creative Judgment | AI’s struggle with discerning fruitful research paths limits its capability in making informed, creative judgments in science. |
| name | description |
|---|---|
| AI-Assisted Research | AI systems are being integrated into the research process to facilitate tasks like generating hypotheses, organizing tasks, and writing papers. |
| Collaborative AI Functionality | Researchers supervise and interact with AI to guide its work, rather than relying on it fully autonomous methods. |
| Iterative Problem Solving | Using AI for iterative tasks—breaking down complex problems into manageable parts, allowing for step-by-step reviews and improvements. |
| Acceleration of Research Cycles | AI tools significantly reduce the time taken for research projects, accelerating productivity and enabling tackling of more complex problems. |
| Need for Human Oversight | Despite advances, AI requires continuous human guidance to ensure data integrity and accuracy in scientific work. |
| Adaptation in Education | Emerging trends suggest a need to adapt educational approaches for new scientists to work alongside AI effectively. |
| Evolution of Research Roles | A shift in the roles of human researchers towards more supervisory and strategic positions as AI takes on routine tasks. |
| Interdisciplinary Collaboration | Researchers from different domains are beginning to leverage AI tools in specialized fields, indicating cross-disciplinary applications. |
| name | description |
|---|---|
| AI Scientist | An AI system designed to automate the entire research lifecycle, from generating hypotheses to writing papers, released by Sakana AI in August 2024. |
| AI Co-Scientist | Google’s Gemini-based system that assists researchers in generating and evaluating hypotheses at scale, released in February 2025. |
| Asta Ecosystem | An open-source AI ecosystem launched by the Allen Institute featuring tools for pattern detection in complex datasets, released in August 2025. |
| Claude Code | An agentic coding tool developed by Anthropic that allows AI to perform detailed computations and scientific writing tasks more efficiently. |
| LLMs (Large Language Models) | Advanced language models like Claude and Gemini that can assist in various research tasks, currently at the G2 level of understanding. |
| FunSearch | DeepMind’s AI project that made breakthroughs in combinatorics through automated problem-solving, launched in 2023. |
| AlphaEvolve | An AI model that used LLMs to make new discoveries in combinatorics, evidencing the capabilities of AI in mathematical research. |
| AlphaProof | A project that achieved significant results at the International Mathematical Olympiad through automated problem-solving. |
| EVENT2 | An old Fortran code used for Monte Carlo simulations that AI can now handle in scientific computations, highlighting integration of AI in simulation tasks. |
| Self-verifying AI | AI models like Claude that can perform tasks and self-check for errors, improving their outputs through iterative processes. |
| name | description |
|---|---|
| AI in Theoretical Physics Research | The exploration of AI’s capabilities in theoretical physics, specifically the potential for LLMs to contribute to research traditionally reserved for human intellect. |
| AI-Assisted Scientific Workflows | The growing integration of AI into scientific workflows that could revolutionize research efficiency and methods, impacting how science is conducted. |
| Challenge of AI Responsibility | The legal and ethical challenges arising from acknowledging AI contributions in scientific research, especially regarding accountability and authorship. |
| Future of Higher Education in STEM | The potential transformation of higher education dynamics, especially in STEM fields, in response to advancements in AI capabilities. |
| AI’s Role in Creativity and Taste | The ongoing challenge in developing AI systems that possess human-like intuition and creativity in selecting research directions, described as ‘taste’. |
| Impact on Human Graduate Students | The changing role of human graduate students in scientific research as AI increasingly takes on tasks traditionally performed by them. |
| Automation of Research Tasks | The rise of fully automated research systems, with implications on research methodologies and academic publishing standards. |