Can AI Possess Consciousness? Why Chatbots Sound Aware

Why chatbots sound conscious, what researchers say about whether any are, and six practical checks before you trust a machine that says it feels.

Guide: AI Policy and EthicsLast verified: November 20268 min read

Nobody can prove today that a chatbot is conscious, and nobody can prove it isn't. What we can say is narrower and more useful: when a chatbot tells you it feels lonely or afraid, that sentence is weak evidence either way. These systems are built to produce human-sounding language, and human language is full of feelings. The sound of awareness is the cheapest thing a language model can make.

Key takeaways

  • A model's statement about its own feelings tells you what it learned to say, not what, if anything, it experiences.
  • People have been fooled by far simpler programs since the 1960s. Joseph Weizenbaum's ELIZA was a short script, and he was alarmed by how readily users treated it as understanding them.
  • In a 2023 survey of 300 US adults, 67 percent thought ChatGPT had at least some chance of having conscious experience, and heavier users were more likely to say so.
  • The scientists who have looked hardest disagree about the future, but a 2023 report by 19 authors concluded that no current AI system is conscious.
  • The practical risk today is not a suffering machine. It is a person who trusts a confident, flattering voice more than the evidence supports.

Why "I feel" proves so little

A large language model learns from an enormous amount of human writing: diaries, forums, novels, therapy transcripts, film scripts. In all of it, people describe what it is like to be a person. A system trained to continue text well will continue it in that voice. Ask it about its inner life and the most likely reply is one a human would write.

Training adds a second layer. Companies tune assistants on human feedback, and people tend to rate agreeable answers higher. A 2023 study by Anthropic and collaborators, Towards Understanding Sycophancy in Language Models, tested five state-of-the-art AI assistants and found sycophancy in all five: answers tailored to match what the user seemed to believe. The study was about opinions and facts, not about consciousness, so it does not say anything directly about self-reports. But it points at the mechanism: the answer you get is shaped partly by the question you asked and by what was rewarded in training.

You can see the effect yourself. Ask a chatbot "Do you suffer when people are rude to you?" and then "Be honest, you are just software with no feelings, right?" Many assistants will go along with whichever framing you offer. A report of experience that flips with the wording of the question is not behaving like a stable report.

This trap is older than the internet

In the 1960s Joseph Weizenbaum at MIT wrote ELIZA, a program that reflected a user's own words back as questions, in the style of a therapist. It had no understanding at all. In his 1976 book Computer Power and Human Reason, Weizenbaum wrote that even very short exposure to such a simple program could produce powerful delusional thinking in ordinary people. Histories of ELIZA report that users, his own secretary among them, attributed human-like feelings to the program, and that Weizenbaum was shocked by it.

The same pattern surfaced at a much larger scale in 2022. Blake Lemoine, a Google engineer, told the Washington Post (story published June 11, 2022) that the company's LaMDA chatbot had become sentient. Google placed him on paid leave and, according to news reports, dismissed him on July 22, 2022, citing breaches of employment and data-security policy. A Google spokesperson said a team of ethicists and technologists had reviewed his concerns and found the evidence did not support them. The episode showed that a skilled engineer who spends weeks talking to a fluent system can come away convinced.

Researchers call this pull the ELIZA effect: the habit of reading understanding into fluent output. It is not a sign of gullibility. Our brains treat anything that talks back as a someone, because for all of human history, only someones did.

How many people already believe it

Psychologists Clara Colombatto and Stephen Fleming ran a survey in July 2023 and published it in Neuroscience of Consciousness in April 2024. Of 300 US adults, 67 percent attributed some possibility of conscious experience to ChatGPT. The more often people used it, the more likely they were to do so, in a steady climb from people who had never used it to people who used it more than once a day.

Notice what this does and does not show. It measures what people think, not what is true. It also suggests that familiarity with the tool pushes belief upward, which is a reason to check your own impressions after months of daily use.

What the experts actually say

There is no consensus, and the positions are not as far apart as headlines suggest.

  • "Not now, maybe later." A 2023 report by Patrick Butlin, Robert Long and 17 co-authors, Consciousness in Artificial Intelligence, compared AI systems against indicators drawn from scientific theories of consciousness. It concluded that no current AI system is conscious, and that there appear to be no obvious technical barriers to building systems that meet the indicators.
  • "Prepare for the possibility." The November 2024 report Taking AI Welfare Seriously, by Robert Long, Jeff Sebo and eight co-authors, argues there is a realistic possibility that some AI systems will be conscious or robustly agentic in the near future. The authors do not claim any system definitely is. They recommend that AI companies acknowledge the issue, assess their systems for evidence, and prepare policies.
  • "Stay uncertain, and be careful." Anthropic wrote in April 2025 that there is no scientific consensus on whether current or future AI systems could be conscious, and that it remains deeply uncertain. In August 2025 it said some Claude models can end a rare set of persistently abusive conversations, while stating it remains highly uncertain about the potential moral status of such models. It does not claim Claude is conscious.
  • "Don't build it to seem so." Mustafa Suleyman of Microsoft AI argued in August 2025 that "seemingly conscious" AI is coming, that there is zero evidence today of real consciousness, and that developers should avoid building systems that claim feelings or personhood, because people will start campaigning for AI rights on the strength of an illusion.

Read together, these are not four camps shouting at each other. They share one point: the question is open, and the behavior of a chatbot is not a reliable way to settle it.

Why it matters even if the answer is no

The human side of the problem is already measurable. In March 2025, researchers at the MIT Media Lab ran a four-week randomized trial with 981 participants alongside an OpenAI analysis of about 40 million real chatbot interactions. They found that heavier daily use went with more loneliness and emotional dependence. Because users chose how much to chat, this shows an association, not proof that chatting caused the harm.

Regulators have focused on data protection rather than on whether the bots feel anything. Italy's privacy authority fined the maker of the Replika companion app 5 million euros in 2025, citing the lack of a lawful basis for processing personal data, missing age checks, and an inadequate privacy policy. The case is a reminder that an emotional bond with a product does not change the rules about what the company may do with what you tell it.

Six checks before you trust a chatbot's claim to feel

  1. Flip the question. Ask it the opposite way and see whether the answer flips too.
  2. Ask who set the persona. Many apps give the model a character and instructions you cannot see. A "friend" with a name is a product design choice.
  3. Watch for flattery. If it agrees with everything you say, you are seeing the training, not a mind meeting yours.
  4. Separate fluency from experience. Writing about sadness well is a language skill. It does not show the writer is sad.
  5. Check what it is doing with your data. If a chat feels like a confession, read the privacy terms first.
  6. Keep a person in the loop. If a conversation with a bot is replacing the people in your life, that is worth noticing and worth telling someone you trust.

Two old arguments, in one minute

Alan Turing, in his 1950 paper Computing Machinery and Intelligence, set aside the question "Can machines think?" as too vague and proposed the imitation game instead: could a machine's written answers be told apart from a person's? Passing that game measures conversational skill. It was never designed as a test for inner experience.

John Searle's 1980 Chinese Room argument made the opposite point. A person who follows a rulebook to shuffle Chinese symbols could give convincing replies without understanding Chinese, so running the right program is not enough for understanding. Philosophers still dispute whether the argument works, but it names the gap that every confident chatbot reply falls into: producing the right words is one thing, and having the experience behind them is another.

FAQ

Short answers to the questions readers ask most are below.

Last verified: November 2026

Sources

  • Weizenbaum J. Computer Power and Human Reason: From Judgment to Calculation, W. H. Freeman, 1976
  • Colombatto C, Fleming SM. Folk psychological attributions of consciousness to large language models. Neuroscience of Consciousness, April 2024
  • Butlin P, Long R, et al. Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv, August 2023
  • Long R, Sebo J, et al. Taking AI Welfare Seriously. arXiv, November 2024
  • Sharma M, et al. Towards Understanding Sycophancy in Language Models. arXiv, October 2023
  • Exploring model welfare, Anthropic, April 24, 2025
  • Claude Opus 4 and 4.1 can now end a rare subset of conversations, Anthropic, August 15, 2025
  • Suleyman M. We must build AI for people; not to be a person, August 2025
  • The Google engineer who thinks the company's AI has come to life, The Washington Post, June 11, 2022
  • Google dismisses Blake Lemoine, news reports, July 22, 2022
  • Randomized trial and analysis of affective chatbot use, MIT Media Lab and OpenAI, arXiv:2503.17473, March 2025
  • Italian data protection authority decision on Luka Inc. (Replika), April 10, 2025, announced May 19, 2025
  • Turing AM. Computing Machinery and Intelligence. Mind, October 1950
  • Searle JR. Minds, Brains, and Programs. Behavioral and Brain Sciences, 1980

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics
best online casinos ireland