WHAT, IF
ANYTHING, FEELS?
Sentience is the capacity for Valence is the pleasant or unpleasant quality of a feeling. A burn can feel bad; relief when it stops can feel good. experience: the ability to feel something as pleasant or unpleasant.
This page follows research into how feelings might arise, how researchers test for them, and what comparisons with living systems can teach us.
No current AI system has been shown to experience pleasure, distress or another subjective feeling.
Researchers can measure internal emotion and preference representations, test how they affect decisions, and build agents whose internal needs regulate behaviour. The open question is whether any such state feels good or bad to the system.
SENTIENCE RESEARCH
IN MOTION.
Four rows separate theory, architecture, observed mechanisms and proposed or performed tests. Pick a moment for its finding, limit and primary source.
The marker date is the cited work’s publication or public release date. Theory, architecture, mechanism and test describe the kind of work; they are not stages of sentience.
BROWSE ALL NINE SOURCE RECORDS ↘
| DATE | RESEARCH | TYPE | STATUS | SUBSTRATE | EMBODIMENT | PRIMARY SOURCE |
|---|
THE EVIDENCE
MATRIX.
Ten systems or findings, seven evidential questions. Select a row to read its strongest claim, an alternative explanation, and the next test that would matter.
Choose a system above for its interpretation.
ACROSS
SUBSTRATES.
This dated grid compares evidence used to infer sentience across biological, living-hybrid and artificial systems. It does not rank how sentient they are.
The relief column brings together different tests: reported relief, learned place preference, analgesic self-selection and drug response. Labels such as “No test reported” describe the linked work, not all research on that system. “Mixed results” flags disagreement; select the cell for the studies and methods. SOURCE CHECK marks a citation or criterion under review. SOURCE GAP identifies an account whose original report has not been located. SOURCE CONFLICT marks a study that challenges the displayed assessment. METHOD LIMIT identifies a test-design caveat. Supplied source ratings remain unchanged.
Research last reviewed September 2026
Tap a coloured cell for the evidence behind its label. Tap a system name for sources and context.
Biological reference rows change when major evidence or consensus shifts. Artificial and wetware rows are reviewed with the monthly Sentience research cycle.
WHAT WE
TRACK.
These are research source classes, not eight traits or a score. The timeline is selective; a missing marker means the work is not yet documented here.
WHO IS ACTUALLY
STUDYING THIS?
Research groups, labs and field-building organisations working on machine consciousness, digital minds and AI welfare. Listing a group does not endorse a sentience claim.
HOW TO READ
THIS PAGE.
The timeline shows dated research and gives the finding, its limit and a source. The evidence matrix asks what each study establishes. The comparison table shows how researchers test similar questions across living and artificial systems.
What do “valence” and “functional valence” mean?
Valence is the pleasant or unpleasant quality of a feeling. A burn feels bad; relief when it stops can feel good. Those are examples of negative and positive valence.
Functional valence is an internal signal that makes a system treat one state as better or worse and change its behaviour. A robot might seek a charger when its battery runs low. That action shows how its battery signal guides decisions. Researchers would need further evidence to establish that the robot feels discomfort or relief.
Why include animals and living tissue?
Biological cases show how researchers combine behaviour, physiology, intervention and, where available, report. The same questions can expose missing tests in artificial systems without assuming that all substrates work alike.
How will this change?
A cell should move only when new evidence bears directly on that question. Record its previous state, new state, date and source. Research programmes and commentary remain in the supporting ledger until a specific result changes the evidential picture.