ARTIFICIAL INTELLIGENCE AGENTS AS HYPOTHESIS GENERATORS IN PSYCHOLOGY AND NEUROSCIENCE

Published: 2 June 2026| Version 1 | DOI: 10.17632/km2fk8j4w3.1
Contributor:
Milaim Delija

Description

The replication crisis in psychology and the interpretability gap in neuroscience converge on a structural weakness that methodological reforms aimed at experimental design and statistical transparency have largely left untouched: hypothesis generation remains an informal, cognitively constrained process that is rarely subjected to systematic methodological scrutiny. This paper proposes a four-phase computational framework in which artificial intelligence agents—goal-directed systems capable of extended planning, structured memory retrieval, and iterative self-correction—participate as active components of the scientific reasoning process at the stage of hypothesis formation and preliminary evaluation. The framework positions these systems upstream of the experimental process—as instruments for expanding the hypothesis space explored before experimental resources are committed. Each candidate hypothesis produced by the system is accompanied by an explicit mechanistic account, a falsification condition, and an auditable reasoning chain. The framework is developed through a case study in working memory research, in which a multi-agent system navigates conflicting empirical findings on capacity limits and attentional modulation to produce three candidate hypotheses with distinct mechanistic commitments. A computational simulation implementing the biased competition architecture (Miller & Cohen, 2001) subjects all three hypotheses to identical experimental conditions and confirms that only H3 produces the predicted differential pattern of conjunction and feature errors across parametrically varied attentional load levels, providing the computational grounding for empirical follow-up that a framework paper of this kind requires. The paper argues that systematic, documented exploration of the hypothesis space is epistemologically valuable in its own right, and that agentic architectures provide tractable means for achieving it.

Files

Steps to reproduce

https://hypothesis-engine.onrender.com/

Categories

Psychology, Neuroscience, Artificial Intelligence

Licence