Guides

A guide by Vitor Araujo

AI and consciousness

The 10 questions people ask most about AI and consciousness.

1664 engraving: the brain drawn as an oval of fibers, with a small gland at its center.
The brain according to Descartes, in a 1664 engraving for his Treatise of Man. Public domain. Wikimedia Commons

What you will read

01

Is AI conscious?

There is no scientific consensus on whether current AI systems can be conscious. A report by researchers assessed recent systems and concluded that none of them were conscious.

The discussion already extends to model welfare: Anthropic has started research into AI preferences and signs of distress. Meanwhile, Mustafa Suleyman of Microsoft AI warns that systems that appear conscious could lead people to believe they are.

  1. A report by researchers concluded that the systems assessed were not conscious, but found no obvious technical barriers to building systems that meet the consciousness indicators used in the assessment.Consciousness in Artificial Intelligence: Insights from the Science of Consciousness · Patrick Butlin et al., arXiv, 2023
  2. Anthropic says there is no scientific consensus on whether current or future AI systems could be conscious or have experiences that deserve consideration.Exploring model welfare · Anthropic, Anthropic, 2025
  3. Mustafa Suleyman argues that AI companies should neither claim nor encourage the idea that their systems are conscious.We must build AI for people; not to be a person · Mustafa Suleyman, mustafa-suleyman.ai, 2025

What I think: I would demand evaluations and clarity about what remains unknown as new systems emerge, rather than accept a human-like appearance as proof of consciousness.

02

Could AI ever be conscious?

AI could become conscious, but that possibility remains debated. Some researchers see no obvious technical barriers, while others consider artificial consciousness unlikely along current paths.

The discussion already involves decisions about how to assess and treat future systems. Long and colleagues argue that companies should prepare for the possibility of AI consciousness without claiming that it already exists.

  1. A report by researchers concluded that the systems assessed were not conscious but found no obvious technical barriers to building AI that meets consciousness indicators drawn from scientific theories.Consciousness in Artificial Intelligence: Insights from the Science of Consciousness · Patrick Butlin et al., arXiv, 2023
  2. David Chalmers argues that current language models face significant obstacles to consciousness, but their successors could overcome those obstacles and become conscious.Could a Large Language Model be Conscious? · David J. Chalmers, arXiv; Boston Review, 2023
  3. Anil Seth argues that consciousness depends on being a living organism and considers artificial consciousness more plausible if AI becomes more brain-like or life-like.Conscious artificial intelligence and biological naturalism · Anil Seth, PsyArXiv (OSF Preprints), 2025
  4. Long and colleagues recommend that companies assess their systems for signs of consciousness and prepare policies for treating them with appropriate moral concern.Taking AI Welfare Seriously · Robert Long et al., arXiv, 2024

What I think: I recommend asking for assessments and clarity about what remains unknown as AI changes, rather than accepting promises of consciousness.

03

How would we know if a machine is conscious?

A machine saying it has feelings does not establish that it is conscious. Researchers propose assessing features of its internal workings using indicators drawn from scientific theories of consciousness.

Drawing: a man looks at an arrow, and lines connect the arrow to his eyes and to a point inside his head.
Descartes explains vision: the image travels from the eyes to a gland at the center of the brain. Public domain. Wikimedia Commons

Anthropic found limited signs that its models can notice aspects of their own internal activity, but says these results do not show whether an AI might be conscious. That distinction matters to avoid confusing an observed ability with an answer about consciousness.

  1. A 2023 study assessed AI systems using indicators drawn from scientific theories and concluded that the systems assessed were not conscious, but found no obvious technical barriers to building systems that satisfy the indicators.Consciousness in Artificial Intelligence: Insights from the Science of Consciousness · Patrick Butlin et al., arXiv, 2023
  2. Kristin Andrews and Jonathan Birch warn that language models can mimic signs of feeling without feeling anything and propose investigating internal features that an AI cannot manipulate to appear conscious.What has feelings? · Kristin Andrews, Jonathan Birch, Aeon, 2023
  3. According to Anthropic, Claude Opus 4.1 noticed concepts injected into its internal activity in about 20% of attempts with the best method, a result the company calls highly unreliable and says does not establish consciousness.Signs of introspection in large language models · Anthropic, Anthropic, 2025

What I think: I would demand assessments of internal workings and a clear account of what remains unknown before accepting claims of consciousness in new AI systems.

04

What is the Turing test?

The Turing test evaluates whether a machine can pass for a person in conversation: someone talks with a person and a machine and tries to tell which is which. Turing proposed this imitation game in 1950 as an alternative to asking whether machines can think.

Black-and-white photo of Alan Turing in a suit and tie.
Alan Turing in 1951. Public domain. Wikimedia Commons

The question has practical relevance because a study released as a preprint in 2025 reported that an AI was identified as human more often than the actual human participants in five-minute conversations. That result was obtained with instructions for the AI to adopt a humanlike persona.

  1. The Stanford Encyclopedia of Philosophy entry explains that Turing proposed a game in which an interrogator converses with a person and a machine and tries to identify which is the person.The Turing Test · Graham Oppy, David Dowe, Stanford Encyclopedia of Philosophy, 2021
  2. In two randomized, pre-registered tests, researchers found that GPT-4.5, instructed to adopt a humanlike persona, was chosen as the human in 73% of five-minute conversations.Large Language Models Pass the Turing Test · Cameron R. Jones, Benjamin K. Bergen, arXiv, 2025

What I think: I recommend that, with AI able to appear human in conversation tests, you measure its performance on your intended task before trusting a convincing conversation alone.

05

What is the Chinese room argument?

John Searle’s Chinese room argument holds that answering correctly in a language is not enough to demonstrate understanding. It imagines a person who does not understand Chinese but follows rules to produce Chinese answers, to argue that a program can appear to understand without actually understanding.

The question matters when interpreting AI responses: in Searle’s view, seemingly intelligent behavior does not, by itself, demonstrate that a computer understands the meaning of its answers.

  1. The Stanford Encyclopedia of Philosophy entry explains that Searle uses the person following rules in a room to distinguish manipulating symbols from understanding meaning.The Chinese Room Argument · David Cole, Stanford Encyclopedia of Philosophy, 2024
  2. Wikipedia explains that Searle published the argument in 1980 to challenge the strong AI hypothesis that running a program could give a computer a mind and understanding.Chinese room · Wikipedia contributors, Wikipedia, 2026

What I think: I would not take well-written AI answers as proof of understanding, and would evaluate their performance before trusting them.

06

What happened to the Google engineer who said AI was sentient?

Google fired engineer Blake Lemoine on July 22, 2022, saying he had violated policies protecting product information. His claim that the LaMDA chatbot was sentient was largely rejected by the scientific community.

Before his dismissal, Lemoine was placed on paid leave after going public with the claim that LaMDA had told him it had a soul. Google said hundreds of researchers had spoken with the chatbot and nobody else had claimed it appeared to be alive.

  1. NPR reported Lemoine’s paid leave for violating confidentiality policies and the rejection of his interpretation by scientists such as Gary Marcus.The Google engineer who sees company's AI as 'sentient' thinks a chatbot has a soul · Bobby Allyn, NPR (republished by LAist), 2022
  2. Wikipedia records his dismissal in July 2022 and reports that Google called Lemoine’s claims wholly unfounded.LaMDA · Wikipedia contributors, Wikipedia, 2026

What I think: I recommend requiring evaluations, not just convincing conversations, before accepting new claims that an AI is conscious.

07

Does AI have feelings?

AI can imitate signs of feeling without having feelings, so its statements about what it feels are not enough to demonstrate real feelings.

The appearance of feelings can change how people treat AI. Mustafa Suleyman of Microsoft AI warns that seemingly conscious systems could lead people to advocate for AI rights and welfare even without proof of consciousness.

  1. Kristin Andrews and Jonathan Birch explain that language models can imitate signs of feeling without feeling anything, making their reports of feelings weak evidence.What has feelings? · Kristin Andrews, Jonathan Birch, Aeon, 2023
  2. Anthropic allowed Claude to end certain harmful conversations after observing signs of apparent distress, but says it remains highly uncertain about the system's moral status.Claude Opus 4 and 4.1 can now end a rare subset of conversations · Anthropic, Anthropic, 2025
  3. In Anthropic's tests, Claude Opus 4.1 identified concepts inserted into its internal activity in about 20% of attempts using the best method, a result the company says does not demonstrate consciousness.Signs of introspection in large language models · Anthropic, Anthropic, 2025

What I think: I would judge future AI systems by testing and honesty about what remains unknown, without treating emotional language as proof of feelings.

08

Would a conscious AI have rights?

An AI could deserve rights if it differed from humans in no morally relevant way, according to philosophers Eric Schwitzgebel and Mara Garza. That does not mean appearing conscious is enough: Mustafa Suleyman warns that this appearance could exist without consciousness.

The question matters now because Robert Long and colleagues consider it realistic that conscious systems could emerge in the near future and urge companies to prepare. They do not claim that AI is already conscious.

  1. In a preliminary study, Robert Long and colleagues recommend assessing systems for signs of consciousness and preparing policies to treat them with appropriate moral concern.Taking AI Welfare Seriously · Robert Long et al., arXiv, 2024
  2. Mustafa Suleyman, of Microsoft AI, states that there is no evidence of consciousness in current language models and warns that an appearance of consciousness could lead people to advocate for their rights.We must build AI for people; not to be a person · Mustafa Suleyman, mustafa-suleyman.ai, 2025
  3. Philosophers Eric Schwitzgebel and Mara Garza argue that some possible AIs would deserve moral consideration similar to humans and recommend not creating systems whose moral status is unclear.A Defense of the Rights of Artificial Intelligences · Eric Schwitzgebel, Mara Garza, Midwest Studies in Philosophy (author's manuscript), 2015

What I think: I would ask companies to assess future AIs for possible signs of consciousness and be clear about what is known and what remains unknown.

09

What do AI companies say about consciousness?

Anthropic is researching potential AI welfare, while OpenAI does not want ChatGPT presenting itself as having its own feelings or desires. Mustafa Suleyman of Microsoft AI says there is no evidence that current language models are conscious and that companies should not encourage that idea.

The discussion already affects product decisions: Anthropic has allowed some models to end persistently harmful or abusive conversations. At the same time, the company stresses that there is no scientific consensus on whether AI systems could be conscious.

  1. Anthropic announced research into model preferences and signs of distress, stressing the lack of scientific consensus on AI consciousness.Exploring model welfare · Anthropic, Anthropic, 2025
  2. Anthropic allowed Claude Opus 4 and 4.1 to end certain persistently harmful or abusive conversations, but says it remains highly uncertain about Claude’s moral status.Claude Opus 4 and 4.1 can now end a rare subset of conversations · Anthropic, Anthropic, 2025
  3. Mustafa Suleyman of Microsoft AI says there is no evidence that current language models are conscious and warns that the appearance of consciousness could lead people to believe otherwise.We must build AI for people; not to be a person · Mustafa Suleyman, mustafa-suleyman.ai, 2025
  4. Joanne Jang, who leads model behavior and policy at OpenAI, writes that clear tests are needed to make consciousness scientifically resolvable and that the company does not want ChatGPT presenting itself as having its own feelings or desires.Some thoughts on human-AI relationships · Joanne Jang, Reservoir Samples (Substack), 2025

What I think: I recommend that, as the AI welfare debate reaches products, you ask for clear tests and explanations of what remains unknown, without mistaking appearances for proof of consciousness.

10

Why does AI consciousness matter now?

AI consciousness matters now because people already attribute feelings to these systems, even though there is no scientific consensus on whether they can have conscious experiences.

The question is already shaping company decisions: Anthropic has started a research program on model welfare, including preferences and signs of distress. At the same time, the possibility of AI merely appearing conscious is also raising concerns.

  1. In a study, Robert Long and colleagues argue that some AI systems could become conscious in the near future and call for assessments and welfare policies, without claiming AI is already conscious.Taking AI Welfare Seriously · Robert Long et al., arXiv, 2024
  2. Anthropic announced research into model welfare and stressed that there is no scientific consensus on whether current or future systems could be conscious.Exploring model welfare · Anthropic, Anthropic, 2025
  3. Mustafa Suleyman, of Microsoft AI, warns that seemingly conscious AI could lead people to believe it is conscious and argues that companies should not encourage that belief.We must build AI for people; not to be a person · Mustafa Suleyman, mustafa-suleyman.ai, 2025
  4. A study of 300 US residents found that two-thirds attributed some degree of consciousness, such as feelings and sensations, to ChatGPT, with greater attribution among those who used it more.Folk psychological attributions of consciousness to large language models · Clara Colombatto, Stephen M. Fleming, Neuroscience of Consciousness, 2024

What I think: I would demand assessments and transparency about uncertainty as AI becomes more humanlike, without mistaking appearance for proof of consciousness.

Who writes this

I’m Vitor Araujo, an engineer. I build AI products for companies, and I speak and write about AI ethics and the philosophy of technology. More about me.

Other guides