What’s Happening in Mathematics Will Likely Happen in Psychology

OpenAI recently released hundreds of AI-generated mathematical results. What I keep coming back to is that what’s happening in mathematics will likely happen in psychology.

By that, I mean AI performing these tasks at the level of top humans, at superhuman speed and a scale people didn’t expect, including tasks they didn’t expect AI to take on at all. You could say that, in some sense, this will happen in all fields, and while it’s not certain, I believe it will most likely happen in psychology.

I was listening to the Lex Fridman podcast with three mathematicians discussing the release and their reactions. Much of the discussion concerns where AI systems are now, how much better they might get, and what will still be left for humans. One possibility is a “centaur” model, humans and AI working together. Alex Kontorovich brings up centaur chess, saying “that lasted like 10 seconds” before the computers got so good that the human’s contribution became accepting the computer’s move. He also considers what it would mean for mathematics if AI could solve any problem that a human working with AI could solve, though he doesn’t necessarily expect us to get there.

This connects to Zvi Mowshowitz’s three AI pills. Pill 2 is the AGI pill, artificial general intelligence. AI will continue to improve and become capable of doing much of the work humans currently do. Pill 3 is the ASI pill, artificial superintelligence. Intelligence doesn’t stop at the human level. Taking this pill means expecting AI to become better than humans at approximately all tasks within our lifetimes, including the things we imagine will remain our contribution. Zvi leaves room for exceptions, but I think much of the discussion about what will be left for humans reflects Pill 2 thinking. Pill 3 says AI will likely eat it all.

The Generalization Question

We’ve seen progress in software engineering and coding, math competition problems, mathematical proofs, and formalization in Lean. We’re also seeing work in physics, especially physics problems that are heavy in math. In mathematics, models are extending known ideas and applying them in novel ways that have surprised specialists.

How much will these abilities generalize to related fields? Generalization is part of what makes LLMs general purpose. They’re trained on enormous amounts of internet data and other sources, and labs keep changing the proportions and types of data and increasing its quality. One reason ChatGPT was so successful beginning in 2022 was that it could engage in tasks that weren’t specifically presented in its training data.

Reinforcement learning, or RL, can take existing abilities and hone them, building on a base layer of knowledge in specific domains. But there is also generalization in the ability of agentic AI to search, reason through problems, and discover solutions. How far that extends remains a question.

Progress in Research and Psychology

In a recent post, Benjamin Todd discusses how forecasters, including himself, have often underestimated AI performance. Capabilities can advance faster than expected while economic impacts and diffusion move at a different pace. AI researchers themselves are experiencing uplift from using AI in their work.

General research benchmarks also show progress. On AstaBench, which examines literature understanding, coding, data analysis, and discovery, GPT-5.5 improved on GPT-5.4 using the same agent framework. The gains varied across tasks. Research abilities are improving, even while individual areas remain jagged.

How does this progress extend to psychology? The domain includes areas such as social, industrial-organizational, personality, and cultural psychology, where research involves observing people and systems, conducting experiments, and developing theories. General improvements in reasoning and knowledge, alongside agentic abilities to search the literature, write code, and analyze data, apply to tasks within these fields. My field is clinical psychology, including the assessment and treatment of mental health concerns through psychotherapy. Here, and in the related medical field of psychiatry, the question extends from research assistance to clinical reasoning and interactions with patients.

Models are jagged in their psychology abilities. The same model can perform well in some areas and much less well in others, but those published benchmarks are historical snapshots. Models have continued to improve in clinical knowledge and reasoning, mental health conversations, and research tasks such as literature search and data analysis. On MentalHealthBench, for example, OpenAI reported higher scores for newer models on responses to mental health conversations.

I used models as research assistants to help me carry out a project in J-space and clinical AI safety. I developed the questions and scenarios, while Codex helped set up the environment, manage the datasets, and assist with analysis. Agentic AI made it possible for me to carry out that project.

As general model abilities, psychology-specific data, and agentic research abilities improve, I expect continued encroachment on psychology-related research and clinical tasks.

Reinforcement Learning and Psychology

In reinforcement learning, a model learns through repeated interactions with an environment, guided by a defined reward signal. Training adjusts the model to favor actions and strategies that earn higher rewards. The environment defines the tasks, available actions, and feedback, giving the model opportunities to practice and improve.

Could these environments be applied to psychology, as they have been applied to coding and math? Could they mimic psychotherapy and clinical psychology, research tasks, or consulting and supervision?

ResidencyRL is a medical example. The system trains through simulated patient encounters, with feedback for diagnostic accuracy, management, communication, documentation, and safety. The researchers report improvements on simulated clinical evaluations compared with the base model. TherapyGym is an example closer to psychotherapy. It uses simulated patient conversations and rewards based on CBT skills and safety, with clinician ratings used in evaluation.

What could a reward look like across therapy sessions? We might look at measurement-based care, using the PHQ-9 or other patient-reported outcomes, and use an approximation of change simulated over multiple sessions. We could also look at aspects of the interaction, including empathy. Part of the success of this approach comes down to the ability to simulate realistic therapeutic encounters. This is where clinical psychologists and other domain experts can help create data and evaluate the simulations.

I think gathering this data is more a matter of motivation for labs and startups. They have the funding to gather data in the domains they’re interested in. There are millions of therapy sessions happening every year, and useful signals could be gathered from even a small portion of them. There are a lot of potential RL benefits waiting in that data, as there are in other domains the labs consider valuable.

From Simulation to the Real World

This brings us back to generalization. How far can we get with simulated environments, and how much will what is learned generalize to the real world?

In math, proofs can be formally checked in Lean, and AI-generated work has extended to unsolved problems. In hypothesis generation, the result is less directly verifiable and more exploratory. Generating and testing ideas is a skill that can likely continue to improve, but a promising idea still has to be tested.

In clinical psychology, the question is how well interactions in a simulated environment generalize to individual people. For research, it’s how much skills developed in RL environments generalize to research that has not yet been conducted. The same question extends outward to social and systems-level research, where models may simulate groups of people and their interactions.

Some abilities may come more quickly to LLMs, such as empathy in the form of writing, while others come more slowly. But the human limit isn’t necessarily a limit in the AI training pipeline. In specific domains, such as chess, AI has learned to perform better than humans. The question returns to the quality and availability of data, the training environments, and whether what models learn transfers to the tasks we care about.

In my opinion, when it comes to AI’s continued progress into psychology, it’s a matter of when, not if. What we’re seeing in software development and mathematics is a preview of what will happen in other fields, including human-centered domains such as psychology. As we watch people struggle with the societal implications, questions of identity, and what these changes mean in the long term, we can and should prepare now for what that will look like in our own fields.

Process note: AI was used as an editorial and publishing collaborator for research, organizing, drafting, editing, fact-checking, and publication support. The ideas and argument are my own.