A team of cognitive scientists has outlined a radical vision for automating the entire scientific discovery process in psychology and cognitive science. Their proposed system would use large language models to independently design experiments, simulate human behavioral data, generate new theories, and evaluate which findings are most scientifically interesting — potentially accelerating research by decades while raising profound questions about the role of human insight in scientific discovery.
The research, published by Akshay K. Jagadish and colleagues, describes what they call "a fully automated, in silico science of the mind" that could completely reshape how cognitive science research is conducted. Unlike current AI tools that assist researchers, this system would operate the entire discovery cycle autonomously.
Their proposed solution replaces each step of the research process with AI systems. Experimental paradigms would be directly sampled from large language models rather than designed by human researchers. Instead of recruiting actual participants, "high-fidelity behavioral data" would be simulated using what they term "foundation models of cognition" — AI systems trained to mimic human psychological responses.
Perhaps most ambitiously, the framework would eliminate the traditional process of researchers painstakingly crafting cognitive theories by hand. Instead, LLM-based program synthesis would conduct "a high-throughput search over a vast landscape of algorithmic hypotheses." The system would generate and test cognitive models at a scale impossible for human researchers.
The final component addresses a crucial question: how would an automated system determine which discoveries matter? The researchers propose using an "LLM-critic" to evaluate research for "interestingness" — essentially having AI judge the conceptual value and significance of its own findings.
This automated discovery engine would function as what the paper calls "a high-throughput in-silico discovery engine, surfacing informative experiments and mechanisms for subsequent validation in real human populations." The key word there is "subsequent" — human validation would still occur, but only after AI systems had already conducted the initial phases of research.
The implications extend far beyond cognitive science. If successful, this approach could be adapted to other fields that study complex systems and human behavior. The researchers envision their framework enabling "a fast and scalable approach to theory development" that could generate scientific insights at unprecedented speed.
However, the proposal raises serious questions about the nature of scientific discovery itself. Traditional research relies heavily on human intuition, serendipitous observations, and the ability to recognize unexpected patterns. Whether these qualities can be replicated or replaced by algorithmic processes remains an open question.
The framework also assumes that AI models can accurately simulate human cognition and behavior — an assumption that carries significant uncertainty. If the foundation models fail to capture important aspects of human psychology, the entire automated research pipeline could generate sophisticated but ultimately misleading results.
The researchers acknowledge that their automated discoveries would require "subsequent validation in real human populations," suggesting they view their system as a powerful hypothesis generator rather than a complete replacement for traditional research methods.
For now, the paper represents a theoretical framework rather than a working system. The researchers have outlined the architecture for automated scientific discovery but have not yet demonstrated its practical implementation or effectiveness compared to human-led research.
The work arrives as AI systems are increasingly being deployed across scientific fields, from drug discovery to climate modeling. The cognitive sciences, which have long struggled with slow data collection and limited sample sizes, may be particularly well-suited for this kind of computational acceleration.