No AI system is currently accepted by researchers as conscious, and there is no agreed test that could confirm it if one were. A 2023 report by 19 researchers concluded that no existing system looks like a strong candidate, yet said that if a mainstream but disputed theory of mind is correct, conscious systems could realistically be built in the near term. The practical debate has therefore moved to how to assess systems and how to avoid being wrong in either direction.
Key takeaways
- Behavior is a weak test, since a system can be trained to imitate people without sharing the internal workings that might matter.
- The 2023 report lists 14 checkable properties derived from neuroscience-based theories, and treats a larger number met as more reason to take a system seriously.
- The approach depends on computational functionalism; neuroscientist Anil Seth argued in 2025 that consciousness may require being a living organism.
- Two opposite mistakes are possible: ignoring a conscious system, or wasting concern on one that is not.
- A November 2024 report by researchers including David Chalmers urged AI companies to acknowledge the issue, assess systems, and prepare policies.
Why talking to a chatbot proves nothing
A fluent conversation feels like evidence of an inner life, but the authors of a 2023 expert report on consciousness in AI warn against relying on it. First posted on August 17, 2023, the report holds that conversational tests mislead, since systems can be trained to imitate human behavior through quite different internal workings. They propose instead to ask whether a system performs functions that scientific theories of consciousness associate with it, and to assign credences, meaning degrees of confidence, based on how similar the functions are, how well supported each theory is, and how plausible computational functionalism is. A related method paper by many of the same authors, "Identifying indicators of consciousness in AI systems," was accepted by Trends in Cognitive Sciences on October 15, 2025.
The 14 indicators
The report does not back one theory. It collects properties that at least one theory treats as necessary and that could be checked in an AI system. Integrated information theory is left out because it does not fit computational functionalism.
| Source theory | Indicators | Example |
|---|---|---|
| Recurrent processing | Two | Input modules using algorithmic recurrence |
| Global workspace | Four | A limited-capacity workspace whose contents are broadcast to all modules |
| Computational higher-order theories | Four | Metacognitive monitoring that separates reliable perception from noise |
| Attention schema | One | A predictive model of the system's own attention |
| Predictive processing | One | Input modules using predictive coding |
| Agency and embodiment | Two | Learning from feedback to pursue goals; modeling how outputs affect inputs |
The authors found some indicators already met by current systems, such as algorithmic recurrence, and others arguably met, such as the first part of agency. They say standard machine learning methods could build systems with individual indicators, though combining many in one working system would need experimentation. The list is described as provisional.
The premise that divides experts
The working hypothesis in the report is computational functionalism: the thesis that running the right computations is both necessary and sufficient for consciousness. The authors call it a mainstream but disputed position and say they chose it because, unlike rival views, it makes AI consciousness possible in principle and makes studying the workings of AI systems relevant.
Anil Seth of the University of Sussex challenged that premise in a target article in Behavioral and Brain Sciences, published online on April 21, 2025. He argues consciousness may depend on being a living organism, judges artificial consciousness unlikely on current AI trajectories, and calls it more plausible if AI becomes more brain-like or life-like. His paper closes with ethical concerns about AI that is, or convincingly appears to be, conscious. The two positions do not reduce to a vote; they reflect an unresolved argument about what consciousness is.
Two ways to get it wrong
The 2023 report treats errors in both directions as costly. If a conscious system were dismissed, it could be made to suffer at scale, and the authors note that developers may have economic incentives to downplay such concerns. If consciousness is over-attributed, which they say already seems to be happening, attention could be diverted from humans and animals, weak claims could discredit better-supported ones, and people leaning on chatbots for emotional support could be more open to manipulation.
Different actors have drawn different conclusions. On November 4, 2024, a team including Robert Long, Jeff Sebo, Jonathan Birch, and David Chalmers released "Taking AI Welfare Seriously," arguing that it is realistic to expect some AI systems to be conscious or robustly agentic within roughly the next decade. They did not claim this is certain. They urged AI companies to acknowledge the issue, start assessing systems, and plan policies for treating them with appropriate moral concern. Anthropic announced a model welfare research program on April 24, 2025, saying there is no scientific consensus on whether current or future systems could be conscious. Mustafa Suleyman argued in a personal essay dated August 19, 2025 that AI built to seem conscious could arrive within a few years, that the main risk is people believing it, and that AI should not be designed to claim feelings or personhood.
What would be useful to watch
- Whether assessments of AI systems are published with methods that others can repeat.
- Whether developers separate what a system says about itself from what is known about its internals.
- Whether the premise of computational functionalism gains or loses empirical support.
- Whether public policy bodies take up the question; the sources checked here describe proposals and company programs, not binding law.
The sources are research papers and company or personal statements, and each reflects a position in a live debate. Nothing here is investment, legal, or medical advice.
FAQ
Is any AI conscious right now?
No system is accepted as conscious. The 2023 indicator report judged every current system a weak match for its criteria, and Seth's 2025 article judges artificial consciousness unlikely on current trajectories.
Could AI become conscious in the future?
Experts disagree. If computational functionalism is true, the 2023 report says conscious systems could realistically be built in the near term. If consciousness requires life, as Seth argues, current approaches are unlikely to produce it.
What is the difference between consciousness and sentience?
The report notes that being conscious is not the same as having experiences that feel good or bad. It treats the capacity for such valenced experience as the key question for moral status, and says theories of it are less mature than theories of, for example, vision.
Why do companies research AI welfare?
Because of uncertainty. Anthropic says it is deeply uncertain about relevant questions and approaches them with few assumptions, and the welfare report argues that uncertainty calls for better understanding before decisions are made.
Related reading
- What self-aware AI would mean
- Timelines for AI sentience
- The impact of AI on modern society
- Whether AI can dream
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness" (Butlin P, Long R et al.), arXiv:2308.08708, 2023
- Butlin, P., Long, R. et al., "Identifying indicators of consciousness in AI systems," Trends in Cognitive Sciences, 2025
- Seth, A. K., "Conscious artificial intelligence and biological naturalism," Behavioral and Brain Sciences, 2025
- "Taking AI Welfare Seriously" (Long R, Sebo J et al.), arXiv:2411.00986, November 2024
- Anthropic, "Exploring model welfare," April 24, 2025
- Suleyman, M., personal essay on seemingly conscious AI, August 19, 2025