"From mazes to natural habitats: AI reshapes how scientists quantify animal behavior

A mouse standing still in an open field could be freezing in fear, resting, or grooming. Judged by position and speed alone, the three are nearly indistinguishable — and that, argue the authors of a review published in The Innovation Reviews, is precisely the problem with a century of behavioral neuroscience. Classical ethology has long relied on task-constrained apparatuses: Y-mazes to score memory, three-chamber tests for social preference, open-field arenas for anxiety. These designs compress complex behavior into low-dimensional readouts such as trajectory, velocity, and arm entries. Manual scoring is subjective and labor-intensive, the authors note, and it no longer matches the resolution of modern neural recording tools such as high-density silicon probes and optogenetics. The review, by Ke Chen and Yaning Han (co-first authors) with Kang Huang, Liping Wang, and Pengfei Wei, organizes the AI-driven alternative into five stages. Multi-view cameras, microphones, and physiological and environmental sensors first record the animal and its surroundings. Deep-learning tools such as DeepLabCut then extract keypoints — nose, ears, limbs — and reconstruct three-dimensional movement. Supervised classifiers recognize predefined actions, while unsupervised methods can reveal fine-grained, sub-second behavioral structures that human observers tend to miss. Higher-order analysis then interprets individual actions within preceding and following behavior and experimental context, a task for which multimodal large language models may prove useful. Finally, statistical analysis connects quantified behavior to neural activity, physiological state, disease progression, or drug effects. Three obstacles, however, still constrain the field. Behavior has no standard ruler: laboratories may disagree on how many degrees, or how many milliseconds, constitute a mouse's "left turn," and complex states such as anxiety or aggression are harder still. Data collected in different laboratories — with different cameras, environments, keypoint schemes, and behavior categories — resist combination. And as experiments scale from single animals to groups in naturalistic settings, occlusion, identity switching, and shifting social structures such as subgroups and dominance hierarchies compound the difficulty. The authors see several ways forward. Standardized, anatomy- and biomechanics-based posture models for each species would make datasets comparable. High-fidelity virtual animals and simulated environments could generate labeled training data, easing the cost of annotation and reducing reliance on experimental animals. Two more speculative concepts receive early elaboration here: "behavioral nucleotides," which would decompose behavior into elementary, repeatable units that can be analyzed and predicted like language, and "bilingual machine animals" — robots with precisely controllable behavior that interact with real animals to help decode the social meaning of their signals. As a review, the article presents no new experimental data, and the authors frame both concepts as forward-looking proposals. But the synthesis makes a clear argument: the tools to quantify natural behavior now exist, and the remaining barriers are shared standards rather than sensors or algorithms.