Pronouns as Safety Controls for Human-AI Teams

PREPRINT UNDER REVIEW: Nature Human Behaviour (Manuscript ID: NATHUMBEHAV-26010307, submitted Jan 17, 2026). Humans automatically attribute intentional agency to entities referred to with human pronouns ("he"/"she"), even when those entities are non-human AI systems. This linguistic cue creates three risks in safety-critical domains: (1) over-trust calibrated to human reliability rather than system performance, (2) deepfake vulnerability via undetected pronoun-identity mismatches, and (3) accountability confusion obscuring who decided and who acted in hybrid human-AI decisions. This Perspective proposes Agent Category Markers (ACMs)—compact pronouns (AE for disembodied AI, XE for embodied AI, with optional PA/PE/PH for posthuman hybrids)—predicted to suppress automatic mentalizing within the 150–420 ms neurocognitive window, trigger pragmatic incoherence detection in deepfakes (~400 ms N400 response), and stabilize responsibility attribution in incident documentation and clinical communication. ACMs integrate dual-process mentalizing models, pragmatic presuppositions about agency, and biomedical requirements for sex-as-a-biological-variable in humans (sex-differentiated; pharmacologically relevant) versus AI (strictly sex-neutral; scientifically honest). Design requirements ensure compatibility with NIST AI Risk Management Framework, ISO/IEC 42001, and FDA medical device governance. A pre-registered empirical agenda (to be registered at OSF.io) tests trust calibration, clinician override behaviour, fMRI mentalizing suppression, deepfake detection latency, accountability clarity, and conversational hazards ("I"/"you") across nine studies. Study A recruitment begins February 2026. ACMs offer minimal, deployable controls preserving human oversight, while remaining agnostic to debates about AI personhood or consciousness.