Consciousness, capacity by capacity: what the evidence from humans, other animals and AI agents does and does not decide
Brendan Barlow
Zenodo (CERN European Organization for Nuclear Research) October 25, 2026 preprint DOI: 10.5281/zenodo.23633213 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Systematic review |
|---|---|
| Population | Findings from human clinical and experimental research, comparative and animal-welfare research, and research on language models and agents |
| Key findings | The authors argue that if consciousness depends only on functional organisation, which systems are conscious is an empirical question best asked capacity by capacity. Capacities dissociate across humans, animals, and language models, and no study has measured all five families in one system. Of 51 generalisations, only five separate any two of thirteen consciousness theories, and none is contradicted on a finding of more than contested strength. |
Abstract
If there is no fact about consciousness beyond a system's functional organisation, which animals and artificial systems are conscious becomes an empirical question, best asked capacity by capacity. We assembled 305 findings from human clinical and experimental work, comparative and animal-welfare research, and research on language models and agents; organised them by five families of capacities (availability, world model, evaluation, self-monitoring and global states); checked 280 against full text; and grouped them into 51 empirical generalisations. Apart from a shared enabling state, whose loss in coma, seizure or anaesthesia removes them together, pairs of capacities come apart in humans, other animals and language models; no study has measured all of them in the same system. In animals, evidence that relief from a noxious state works as a reward is replicated in rats and rests on single small studies in birds and one octopus species. A trained action that stops when its outcome loses value is shown in rats, macaques, children and one insect study, and in other species only for Pavlovian responses. Behavioural tests of metacognition in animals are contested. In language models, where an internal state was manipulated experimentally, it changed behaviour or report in three quarters or more of cases, and models' reports often failed to track their behaviour; almost all such tests were on models alone, not on agents. Two independent scorers rated thirteen theories of consciousness against the 51 generalisations, using results that would count against each theory (disconfirmers), written from its literature before scoring, and decision rules fixed in advance. By the functional standard of the position, most theories either make no claim about most findings or state their claims in terms these findings do not measure. None is contradicted on a finding of more than contested strength, and only five generalisations separate any two theories. We propose experiments that test those five and fill gaps in the evidence.