Sentient AI in robots and agents: prolegomena for an evidence-based research program

Abstract

The possibility of sentient artificial intelligence has moved from speculative philosophy to a practical interdisciplinary problem for AI, robotics, and human-robot interaction. Large language models, multimodal agents, and embodied robots can now produce first-person reports, maintain dialogue, use tools, act through sensors and effectors, and participate in socially meaningful contexts. These capacities invite two symmetrical errors: anthropomorphic over-attribution and premature dismissal. This Perspective proposes a prolegomenal framework for future research on sentient AI. Its distinctive contribution lies in operationally integrating four elements that have largely been developed in separate literatures: conceptual disambiguation, multi-theory indicator profiles, causal-mechanistic testing, and robotics-specific evidence and governance. The paper distinguishes sentience, consciousness, self-modeling, metacognition, agency, moral patienthood, and AI welfare; separates evidence about an AI system from evidence about human attribution; and proposes domain-specific ordinal evidence levels rather than binary verdicts or an aggregate sentience score. It further specifies welfare- and valence-relevant tests, a preregistered rating procedure, and a concrete protocol for an embodied care robot using sensorimotor lesions, self-location manipulations, memory ablations, and anti-anthropomorphism controls. The aim is not to offer a definitive test for machine sentience, but to show how research could become more scientifically tractable, psychologically informed, robotics-relevant, and ethically responsible.