The one-sentence version
"Whether AI has consciousness" cannot be measured directly, because even human consciousness can only be experienced from the inside and inferred from the outside. A more solid approach is: take several major scientific theories of consciousness, identify the functional or structural features each theory claims are necessary for consciousness, and then check whether existing AI possesses them item by item. In 2023, a report co-authored by a dozen philosophers and neuroscientists, including Butlin and Long, did exactly this. Their conclusion was: Existing AI systems likely lack consciousness, but there are no obvious technical barriers to building systems that satisfy these indicators.
Positioning
Compilation and commentary. All theories of consciousness are controversial; "satisfying theoretical indicators" does not equal "being conscious."
Layer 1: The Intuition
1.1 Two Questions
| Question | Description |
|---|---|
| The Easy Problem (relatively speaking) | How the brain processes information, reacts, and reports "I see red" |
| The Hard Problem (Chalmers) | Why these processing operations are accompanied by subjective experience, and why "seeing red" feels a certain way |
Scientific theories of consciousness primarily address the former: what kind of information processing structures are related to consciousness. They cannot answer the hard problem, but they can provide us with a "checklist."
1.2 Why we cannot ask the AI itself
Large Language Models (LLMs) have learned to talk about consciousness from human text. When it says "I have feelings" or "I do not have feelings," it may simply be reproducing patterns found in its training data. Therefore, self-reporting cannot serve as reliable evidence; we must examine the internal structure of the system.
Layer 2: Key Theories and Indicators
2.1 The Method of Butlin, Long, et al. (2023)
This report, titled "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness," adopts "computational functionalism" as its working hypothesis (consciousness depends on what computations are performed, rather than the material used for implementation; see Is AI Materialist or Idealist?). It extracts over a dozen "indicator properties" from several major theories and uses them to evaluate AI systems at the time.
2.2 Theories and Indicators
| Theory | Core Claim | Representative Indicators | Status of Existing LLMs |
|---|---|---|---|
| Recurrent Processing Theory (Lamme) | Consciousness requires information to feedback back and forth between brain regions | Presence of recurrent processing within input modules | Transformers are essentially feedforward in a single forward pass; reading back their own output during generation is a form of external recursion, but whether this counts is controversial |
| Global Workspace Theory (Baars, Dehaene) | Multiple specialized modules compete to enter a capacity-limited "workspace," and selected information is broadcast to all modules | Specialized modules + limited capacity workspace + global broadcasting + state-dependent attention | Partially present (attention mechanisms and residual streams can be compared to broadcasting); whether a capacity-limited "bottleneck" exists is controversial |
| Higher-Order Theories (Rosenthal et al.) | A mental state is conscious if and only if there is a higher-order representation about it ("I am aware that I am seeing red") | Metacognitive monitoring: distinguishing reliable from unreliable perceptual representations | There is some evidence that models can "know what they know" (calibration), but whether this constitutes true higher-order representation remains unclear |
| Attention Schema Theory (Graziano) | The brain constructs a simplified model of its own attention | A predictive model of attention itself | Very little evidence |
| Predictive Processing (Clark, Seth et al.) | Consciousness is related to hierarchical prediction and prediction error | Predictive models + hierarchical structure | Pre-training is prediction, but whether it is "hierarchical prediction of perception" is a matter of debate (see Predictive Processing and Next-Token) |
| Agency and Embodiment | Consciousness is related to pursuing goals, interacting with the environment, and modeling one's own body | Learning from feedback, pursuing goals; modeling "how output affects input" | Agent systems partially possess this; embodied robots are closer |
2.3 Why Integrated Information Theory (IIT) is discussed separately
IIT (Tononi) posits that consciousness is the degree of integration of information within a system (measured by ) and depends on physical causal structure rather than computational function. According to IIT, any program running on a standard digital computer, no matter how intelligent, has a that is very low, and therefore lacks significant consciousness.
This directly conflicts with computational functionalism, so the report by Butlin et al. did not include IIT in the checklist. In 2023, over a hundred scholars signed a joint critique calling IIT "pseudoscience," which sparked further rebuttals, illustrating the deep divisions in this field.
Layer 3: Review
3.1 Conclusions of the Report
- Existing systems (as of 2023) do not satisfy most indicators and likely lack consciousness;
- However, most indicators are not difficult to implement: there are no fundamental technical barriers preventing the construction of systems that satisfy more indicators;
- Therefore, research is needed in advance, rather than waiting for it to appear suddenly one day.
3.2 Key points to note
- The conclusion depends on computational functionalism. If you believe in IIT or biological naturalism (consciousness must depend on a biological substrate), the conclusion will be entirely different;
- Indicators are candidates for necessary conditions, not sufficient conditions. Satisfying all indicators does not prove consciousness;
- The dilemma of self-reporting: The stronger the model, the more it will "claim" to have experiences, but this is precisely what is most unreliable;
- Interpretability is a key tool: Instead of asking a model "Do you have feelings?", it is better to use interpretability methods to check whether structures like global broadcasting or metacognitive monitoring actually exist within it.
3.3 Why this question matters
- Ethics: If certain AI systems have the potential to possess consciousness or sentience, it raises questions of moral status. Some AI labs have begun researching "model welfare," such as allowing models to end persistently abusive conversations;
- Two-way risks: There are costs to both mistakenly believing an AI is conscious (misled by anthropomorphism) and mistakenly denying it (ignoring real moral issues);
- Science: AI provides an unprecedented "open-box" intelligent system that may, in turn, help test theories of consciousness.
Summary
| Layer | Key points |
|---|---|
| Method | Extract indicators from major consciousness theories and check AI systems item by item |
| Theory | Recurrent processing, Global Workspace, Higher-Order, Attention Schema, Predictive Processing, Agency and Embodiment; IIT discussed separately |
| Conclusion | Existing systems likely lack consciousness; but there are no obvious obstacles to building systems that satisfy the indicators |
| Note | Depends on computational functionalism; self-reporting is unreliable; interpretability is key |