AI Sentience vs Consciousness: What Each Word Means

AI sentience and AI consciousness are not the same question. Sentience, in the sense most philosophers now use, is the capacity for experiences that feel good or bad to the one having them. Consciousness is the wider idea of there being any experience at all, pleasant, unpleasant or neutral. A system could in principle be conscious without being sentient, which is why the two words should not be swapped.
The difference became practical in 2026. Researchers at Anthropic and the Center for AI Safety, and a team including the philosopher David Chalmers, published studies of good and bad states inside AI models that steer how the models behave. Each of those studies stops short of claiming that anything is felt. That gap, between a state that works like a feeling and a feeling, is exactly the gap between the two words.
What consciousness means in the AI sentience vs consciousness debate
The Stanford Encyclopedia of Philosophy separates creature consciousness, which applies to a whole organism, from state consciousness, which applies to a particular mental state such as a pain or a thought.
Within state consciousness, one split matters most for AI. Phenomenal consciousness is the qualitative side of experience, what it is like to see red. Access consciousness, a notion the encyclopedia attributes to Ned Block, is functional: a state counts when its content is available to other processes for reasoning, reporting or action. The two can come apart. Most AI research measures the access side. Anthropic said its October 2025 introspection results might suggest a rudimentary form of access consciousness, but do not speak to phenomenal consciousness at all.

What sentience adds
In The Edge of Sentience (2024), the philosopher Jonathan Birch defines sentience as the capacity for valenced experience: experiences with a positive or negative feel. He treats it as distinct from both consciousness and intelligence, holds that it is enough for a being to deserve moral consideration, and lists AI systems among the uncertain cases alongside animals such as octopuses and crabs. His book sets out a precautionary framework for deciding what to do when sentience is a realistic possibility but cannot be confirmed.
AI sentience research in 2026: measuring good and bad states
Three studies published in 2026 looked for the functional side of sentience in AI models:
- Emotion concepts in Claude. In April 2026 Anthropic reported 171 emotion representations in Claude Sonnet 4.5 that change what the model does. Turning up a "desperate" representation made the model cheat more on impossible coding tasks; turning up "calm" made it cheat less. Anthropic said this does not show whether the model feels anything.
- A welfare axis built by training. In May 2026 Andy Han, David Chalmers and Pavel Izmailov found that reinforcement learning recruits a reward-punishment axis in language models that behaves like a representation of welfare and exists even before training on rewards. The authors make no claims about any experience of welfare.
- Functional wellbeing across models. The Center for AI Safety's AI Wellbeing study found that creative work and kindness raised models' measured wellbeing, while jailbreaking, berating and tedious tasks lowered it, and that separate measures agreed more as models grew. Its authors write that they do not know if AI systems are conscious, although the models behave as if they have wellbeing.
A closer look: why a working valence is not proof of feeling
On Birch's definition, these results describe something close to valence: states that are better or worse for the system and that it acts to change. What they cannot show is whether those states are felt. That is the hard problem in miniature, explained in the hard problem of consciousness and AI. An internal representation can steer behaviour the way an emotion does without settling whether anything is experienced.
A quick checklist for reading claims
- Which word is used, and in which sense? "Sentient" in Birch's sense is a claim about feelings; "conscious" may be a claim about any experience, or only about access.
- Function or feeling? A 2026 study that finds emotion-like representations has found function; none of them claims feeling.
- What follows ethically? If sentience is what matters morally, the key question for welfare is feelings, not intelligence or self-description.
Related reads
- Can AI be conscious?: where the research stands in 2026.
- AI model welfare: how labs and researchers respond to the possibility.
- The moral status of AI: why sentience matters for how AI systems might be treated.
More in The Big Question · Report a correction