Beyond behavioural markers: causal-representational corroboration in AI sentience assessment

Abstract Current generative AI systems increasingly elicit attributions of understanding, emotion, and even consciousness. This creates an evidential problem for AI sentience assessment, because the behavioural and linguistic markers that motivate these attributions may be generated by multiple underlying processes. One source of uncertainty is the gaming problem: AI systems may satisfy sentience indicators because they have learned the relevant behavioural regularities. A further observer-side concern is that anthropomorphic or socially responsive behaviour may influence how strongly these markers are weighted. This paper treats these as distinct sources of evidential uncertainty and focuses on the problem of evidential certification. Behavioural similarity and verbal self-report are not evidentially significant merely because they resemble human or animal expressions of sentience. Their evidential value depends on whether they can be connected to internal processes that causally explain why the relevant outputs are produced. Cross-validating multiple behavioural markers remains unreliable in the LLM case, since models may learn not only individual markers but also their patterns of co-occurrence. Moreover, even under controlled training conditions, verbal self-report remains vulnerable to alternative computational explanations. Since assessments of AI sentience may bear on questions of welfare and moral status, errors in evidential attribution can have ethically significant consequences. In response, I propose the Causal-Representational Corroboration Standard (CCS). The CCS directly evaluates whether behavioural or verbal markers are supported by a stable and causally relevant internal organization whose evidential role is not exhausted by training pressure that directly selects for the assessed marker. The corroborated mechanism becomes relevant to sentience if it also participates in an affectively structured organization and satisfies functional or architectural conditions independently motivated by a plausible theory of consciousness. Therefore, the CCS provides a more disciplined basis for managing the ethical risks of both over-attributing and under-attributing sentience to AI systems. Keywords: AI sentience; Causal-representational corroboration; Behavioural markers; Large language models; Gaming problem