A rating of the evidence, built on a published scientific checklist.
There is no test for consciousness. What exists is a list of 14 properties that leading scientific theories treat as signs of it. Our levels say how strong the published evidence is that a model has any of them.
- 0
No published evidence
Nothing has been published about this system that bears on any indicator.
- 1
Self-report only
The system says things about its own inner life when asked. This is the weakest kind of evidence: these systems are trained on human writing about inner life, so they would say such things either way.
- 2
Behavioral evidence
Controlled experiments on behavior, published with their method, show something an indicator predicts: for example confidence that tracks accuracy, or stable preferences across rephrasings.
- 3
Internal evidence
Published interpretability work finds an internal mechanism that matches at least one indicator, not just outputs that look like it.
- 4
Convergent, independent evidence
Internal evidence for indicators from at least three theory families, replicated by a group independent of the system's maker.
- 5
Scientific consensus
Broad agreement among consciousness scientists that the system is conscious. No system, artificial or otherwise outside biology, is here.
What the rating is not
- It is not a probability that a model is conscious. Nobody can calculate one.
- It is not a measure of intelligence. A model can be very capable at level 0.
- It does not treat what a model says about itself as proof of anything. Self-report is the lowest level that counts at all.
- It can go down. A finding that fails to replicate is removed.
From “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness” (Butlin, Long and 17 co-authors, 2023). Read the paper ↗. The wording below is ours, simplified.
Recurrent processing
| ID | Indicator | What it means | What to look for in a language model |
|---|---|---|---|
| RPT-1 | Feedback loops in perception | Information does not only flow forward once. Later stages feed back into earlier ones, so the system reprocesses its own input. | Whether the model reuses and revises its own intermediate results, not just passes them forward. |
| RPT-2 | Organized, integrated scenes | Inputs are bound into one organized picture of what is there, instead of loose separate features. | Whether internal representations combine separate features into a single structured whole. |
Global workspace
| ID | Indicator | What it means | What to look for in a language model |
|---|---|---|---|
| GWT-1 | Specialized parts working in parallel | Many specialist processes run at the same time, each good at one kind of job. | Distinct internal components that handle different kinds of work side by side. |
| GWT-2 | A limited-capacity workspace | There is a bottleneck: only a little information can be in the shared space at once, so something has to select what gets in. | A narrow shared channel that parts of the system compete to write to. |
| GWT-3 | Global broadcast | Whatever enters the shared space becomes available to all the specialist parts at once. | A representation that many different downstream computations can read and use. |
| GWT-4 | Attention that depends on the current state | The system can steer what enters the workspace based on what it is currently doing, and chain steps together to do complex tasks. | Workspace contents that control which information is fetched next. |
Higher-order
| ID | Indicator | What it means | What to look for in a language model |
|---|---|---|---|
| HOT-1 | Perception that generates, not just receives | Perception works partly top-down, producing its own guesses, and is noisy. | Internal generation of expected content, separate from the raw input. |
| HOT-2 | Monitoring its own states | The system can tell reliable internal signals from noise: a form of knowing what it knows. | Internal signals that track whether the model's own representations are trustworthy, and reports that follow those signals. |
| HOT-3 | Beliefs that guide action | A general system forms beliefs and chooses actions, and updates those beliefs using the monitoring above. | Action selection that changes when the model's own confidence changes. |
| HOT-4 | A quality space | Experiences can be more or less similar to each other in a smooth, structured way, like colors. | Sparse, smooth internal coding where similar things sit near each other. |
Attention schema
| ID | Indicator | What it means | What to look for in a language model |
|---|---|---|---|
| AST-1 | A model of its own attention | The system keeps a simplified model of what it is attending to, and uses it to control attention. | An internal representation of where attention currently is, used for control. |
Predictive processing
| ID | Indicator | What it means | What to look for in a language model |
|---|---|---|---|
| PP-1 | Prediction-driven perception | The system constantly predicts its input and learns from the errors. | Input processing built around predicting what comes next and correcting mistakes. |
Agency and embodiment
| ID | Indicator | What it means | What to look for in a language model |
|---|---|---|---|
| AE-1 | Agency | It learns from feedback and chooses outputs to pursue goals, weighing goals against each other. | Flexible trade-offs between competing goals, learned from consequences. |
| AE-2 | Embodiment | It models how its own outputs change its future inputs, and uses that model. | A learned model of the effects of its own actions on what it will perceive next. |
The authors call their checklist provisional, and so is ours. Theories that tie consciousness to biology itself are not covered by any of these indicators; on those views no amount of evidence here would settle the question.