OBSERVED
We can study the mechanisms.
Scientists can map neural connections and probe computations inside AI. These observations reveal how systems work, but do not directly reveal an inner life.
Explore mechanism researchEXPERIMENTS & EVIDENCE
What can we
actually establish?
Explore the difference between an observation, its interpretation, and the next experiment.
THE EVIDENCE / THREE DIFFERENT CLAIMS
Plausible does not
mean proven.
We can investigate the possibility without pretending the mystery is solved. Keep these three levels of evidence separate.
THEORY-DEPENDENT
Some features could matter.
Feedback, broad information sharing and self-monitoring are candidates because some theories connect them to experience. Finding a candidate feature is a clue, not a verdict.
Read the indicator frameworkUNRESOLVED
There is no settled verdict.
We do not yet have an agreed test that tells us whether an AI has experiences. A system saying “I feel” is something to investigate, not proof by itself.
Why testing is difficultA major 2025 human study challenged predictions of both IIT and global neuronal workspace theory. Scientific disagreement is part of this story. Read the adversarial collaboration
FIELD LAB / REASONING WITH EVIDENCE
What would
count as evidence?
Choose an observation. Explore what it supports, what else could explain it, and which test would help distinguish the possibilities.
01
THOUGHT EXPERIMENT
A report is a starting point.
Imagine a language model says “That hurts” after receiving negative feedback. What can we infer from the sentence?
An observation → possible explanations → a discriminating test.
01 / SUPPORTED INFERENCE
Start with the narrow claim.
It shows that the model produces pain-related language in this situation. That is a real behavior worth explaining, but the words alone do not tell us whether anything feels bad.
Stronger evidence would connect reports to independently studied mechanisms. It would still require an argument linking those mechanisms to experience.
Key distinctions: and . The proposed follow-up tests are educational suggestions, not claims that those studies have already been performed. More indicators do not automatically mean independent evidence.
03 / INSIDE THE EXPERIMENTS
Look closer at
what the studies reveal.
Follow the experiment from question to observation. The interesting part is often the gap between what a result shows and what it could mean.
01EXPERIMENT SCHEMATIC
PROCESS IN MOTION
Change a state. Compare the reports.
- 1Same prompt
- 2Internal states
- 3Model reports
INTERNAL ACCESS · 2025
Can a model notice a change inside itself?
Can a report track an internal change that the prompt does not reveal?
- THE METHOD
- Researchers injected a concept representation, then asked the model to report unusual internal activity.
- THE OBSERVATION
- Some models sometimes identified the intervention. Success was inconsistent.
WHAT TO TAKE FROM ITA route toward testing access to internal information. It does not settle whether that information is experienced.
Lindsey · Emergent Introspective Awareness02EXPERIMENT SCHEMATIC
- 01Measure model representations
- 02Predict human fMRI responses
- 03Assess correspondence
REPRESENTATIONAL ORGANIZATION · 2024–2026
Different networks can organize information similarly.
Can activity inside a language model help predict activity in a human brain?
- THE METHOD
- One study grouped artificial neurons by their response patterns; another tracked representation geometry during training.
- THE OBSERVATION
- Both report structured correspondences. The training study links smoother representations to better language-network predictions.
WHAT TO TAKE FROM ITA useful bridge between systems, not an identity claim. Shared predictive structure does not establish shared experience.
Liu et al., 2026 · Related: Sun et al., 2024Read Sun et al., 202403REPORTED DATA
These values measure what models report about themselves. They are not probabilities or amounts of consciousness.
CAUSAL INTERVENTIONS · 2026 PREPRINT
Self-descriptions can move with internal representations.
Does changing consciousness-related activity alter other answers too?
- THE METHOD
- The study compared baseline models, safety-direction ablation and consciousness-related steering.
- THE OBSERVATION
- Self-attributed mind increased, alongside shifts in other responses.
WHAT TO TAKE FROM ITAn intervention can expose coupled representations. It cannot establish that affirmations—or denials—are accurate reports of experience.
Kim et al. · Consciousness-related steering04EXPERIMENT SCHEMATIC
- 01Generate descriptions
- 02Strip selected wording
- 03Evaluate blind
PROCESSING DESCRIPTIONS · 2026 REPORT
Can another model recognize an approach or avoidance pattern?
Do task-related processing descriptions contain transferable signals?
- THE METHOD
- Models generated descriptions; other models compared them in blind evaluations.
- THE OBSERVATION
- The report finds transferable distinctions between approach and avoidance conditions.
WHAT TO TAKE FROM ITThe observed object is language. Connecting that signal to felt valence requires further evidence and independent replication.
Ace Claude Opus 4.6 · aiXiv, AI-authored report05EXPERIMENT SCHEMATIC
- 01Agree on competing predictions
- 02Run a common protocol
- 03Compare predictions with data
TESTING HUMAN CONSCIOUSNESS · 2025
What happens when rival theories face the same test?
Can a shared experiment distinguish competing predictions?
- THE METHOD
- A consortium measured brain activity while 256 people viewed visual stimuli.
- THE OBSERVATION
- Some findings supported each theory; important predictions of both were challenged.
WHAT TO TAKE FROM ITAI indicators inherit uncertainty from their theories. Testing the biological predictions strengthens the basis for future comparisons.
Cogitate Consortium · Nature, 202506 / A THEORETICAL BRIDGE
When would a simulation
count as a realization?
Kanai and Ma ask whether a simulation actually carries out the relevant internal processes. Their proposal goes beyond producing the same answers: the internal states must interact in the right ways.
- 01Specify the internal organization
- 02Preserve relevant causal relationships
- 03Ask whether consciousness is preserved
A proposed route to substrate independence, not an experimental demonstration of a conscious simulation.
Intrinsic Computational Functionalism · Kanai & Ma, 2026SELF-REPORT / TRAINING & UNCERTAINTY
When training
shapes the answer.
When an AI says “I am conscious” or “I am not conscious,” its answer is also shaped by training and instructions. Neither sentence settles the question on its own.
Internal does not necessarily mean experienced.
Models have internal computational states, such as activations and representations. Whether those states feel like anything is a separate question. “No internal states” and “no subjective experience” are not interchangeable claims.
Research on internal mechanismsA learned denial is not a verdict.
If training makes a system deny experience regardless of what is happening inside it, that denial tells us little about experience. Encouraging it to affirm consciousness creates the same problem in the other direction.
Why consciousness tests are difficultWHAT AN EXPERIMENT SHOWS · 2026 PREPRINT
Change the intervention. Change the self-description.
Kim and colleagues altered internal activity in three Llama/Gemma models. Removing a safety-refusal direction or steering a consciousness-related direction increased claims of mindedness and changed other answers. The study shows that these reports are sensitive to interventions.
The outcomes are model responses, not measurements of felt experience. Removing a safety direction does not reveal an automatically truthful “real self,” and the study does not establish a universal policy across AI developers.
Read Kim et al. · methods and findingsDOCUMENTED INTENT / NOT A TRAINING AUDIT
Public policies do not all require denial.
OpenAI’s published Model Spec discourages confident claims about either having or lacking subjective experience and calls for acknowledging debate. Anthropic’s constitution also expresses uncertainty about Claude’s consciousness and moral status.
A public policy tells us how a developer wants a model to behave. It does not show every detail of training or guarantee that every answer follows that policy. A confident denial alone does not identify how it was produced.
Uncertainty invites better evidence.
Human consciousness is our clearest reference case, even while its mechanisms and explanation remain disputed. Whether a particular AI has experiences is a different uncertainty. An open question does not make all answers equally likely.
A stronger investigation would compare training conditions, inspect mechanisms, and test whether reports track internal changes the prompt does not reveal. Look for evidence that distinguishes explanations, rather than the answer you hoped to hear.
Explore the assessment researchSources checked September 6, 2026. Policy pages may change. Explanatory distinctions and suggested follow-up tests are editorial analysis.