No AI system is known to be self-aware in the sense of having conscious experience of itself, and whether a future system could be remains an open scientific question. A 2023 report by 19 researchers, among them Yoshua Bengio of the University of Montreal and Mila, concluded that no current AI system appears to be a strong candidate for consciousness while finding no obvious technical barrier to building systems that meet its indicators. What has been measured in language models is narrower: a limited, unreliable ability to report on their own internal states.
Key takeaways
- Self-awareness, consciousness, and sentience are different claims, and most headlines blur them.
- The 2023 Butlin, Long, and colleagues report scores AI systems against 14 indicator properties drawn from neuroscientific theories of consciousness.
- In Anthropic's introspection study, the best model named an injected concept on about 20 percent of trials, with no false alarms in 100 control trials.
- That study states that it does not address whether AI has human-like self-awareness or subjective experience.
- Researchers disagree on the basic premise: the 2023 report relies on computational functionalism, which it calls mainstream but disputed.
Three terms that are often mixed up
Articles about "aware" AI use several words as if they were interchangeable. The table separates them as used in this article. The definitions are working labels for reading the research, not settled scientific vocabulary.
| Term | Working meaning | Can it be tested today? |
|---|---|---|
| Functional self-monitoring | A system tracks or reports on its own internal states or its own likelihood of being right | Yes, with experiments on model internals and calibration |
| Consciousness | Subjective experience, something it is like to be the system | Only indirectly, through theories and indicator properties |
| Sentience | Capacity for felt states such as pleasure or pain | No agreed test for machines |
Self-awareness sits between the first two rows. A thermostat monitors a state, a person reflects on being someone, and a chatbot that writes a sentence claiming self-awareness is producing text that resembles both. Only the first row can currently be measured directly.
The indicator approach to consciousness
The most cited scientific framework is the report Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, first posted on August 17, 2023, with Patrick Butlin and Robert Long as lead authors. Rather than ask a chatbot whether it is conscious, the authors take theories such as recurrent processing, global workspace, higher-order, attention schema, and predictive processing, and extract 14 indicator properties that can be described in computational terms. An AI system that has more of them is judged more likely to be conscious.
The report gives a reason for avoiding behavioral tests: AI systems can be trained to mimic human behavior while working in very different ways. It also warns that systems able to imitate conversation convincingly will lead many people to believe they are conscious. Its verdict is cautious. No current system looked like a strong candidate, many indicators could be built with standard machine learning methods, and meeting the indicators would still not prove a system is conscious.
What introspection experiments have found
Anthropic researcher Jack Lindsey tested a different question: can a model detect something about its own internal state? The paper, Emergent Introspective Awareness in Large Language Models, was posted to arXiv on January 5, 2026. The method injects a representation of a known concept, such as a particular word, into a model's activations and then asks whether the model notices anything unusual.
Results were modest and conditional. At the best-performing layer and injection strength, Claude Opus 4.1 named the injected concept on about 20 percent of trials, and the paper reports no false positives over 100 control trials. Claude Opus 4 and 4.1 performed best among the models tested, and results depended on how a model was fine-tuned afterward. The author describes the capacity as highly unreliable and context-dependent. The paper also says it does not seek to address whether AI systems possess human-like self-awareness or subjective experience. The study comes from a developer examining its own models, so independent replication matters.
A 2022 paper titled Language Models (Mostly) Know What They Know shows a related but separate skill. Larger models could estimate the probability that their own answers were correct, and were reasonably calibrated on multiple-choice and true/false questions, though calibration weakened on unfamiliar tasks. That is self-assessment of accuracy, not evidence of experience.
Where experts disagree
The 2023 report adopts computational functionalism as a working hypothesis: the view that performing the right kind of computation is necessary and sufficient for consciousness. It calls this position mainstream but disputed, and it excludes integrated information theory because that theory conflicts with it. Anyone who rejects the premise, for example by tying consciousness to biology, will read the same evidence differently. The authors also call their list of indicators provisional, expecting it to change as research continues.
These disagreements are about theory, not about whether today's chatbots are secretly awake. The documented evidence supports only two statements: some models show weak, unreliable self-monitoring, and no accepted test shows machine consciousness.
A checklist for reading headlines about aware AI
- Does the claim rest on what a chatbot says about itself? Text output alone is weak evidence, since models are trained on human writing about minds.
- Is there a named study, and can its abstract be opened? Quotes without a paper are not evidence.
- Which word does the study use: introspection, consciousness, or sentience?
- Do the authors state limits, such as success rates or unreliable results?
- Was the work done by an independent group or by the company that built the model?
FAQ
Has any AI been proven self-aware?
No. The 2023 indicator report found no current system to be a strong candidate for consciousness, and the 2026 introspection paper explicitly leaves the question of subjective experience aside.
Does a chatbot saying it is conscious count as evidence?
Not by itself. The 2023 report notes that behavioral tests are unreliable because AI can be trained to mimic human behavior, and that convincing conversation tends to make people believe a system is conscious.
What is the difference between introspection and consciousness?
Introspection in the Anthropic study means a model correctly reporting an altered internal state, which can be measured. Consciousness means subjective experience, which no experiment has shown in any AI system.
Could future AI become conscious?
The 2023 report sees no obvious technical barriers to building systems that meet its indicators, but says doing so would not prove consciousness. The answer depends on theories that researchers still dispute.
Related reading
- Expert estimates of when AI sentience might arrive
- Conscious machines and their effect on society
- AI systems that build other AI systems
- Why AI systems fail in practice
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Butlin P, Long R, et al., Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, arXiv:2308.08708, 2023
- Lindsey J, Emergent Introspective Awareness in Large Language Models, arXiv:2601.01828, January 5, 2026
- Kadavath S, et al., Language Models (Mostly) Know What They Know, arXiv:2207.05221, 2022