AI Consciousness: Is AI Sentient? Theories, Tests and the 2026 Debate

Is AI conscious or sentient? The main theories, the 2023 Butlin et al. indicator report, and what Anthropic and Microsoft say in 2026.

By Izhar Ahmad Danish · · 4 min read

AI Consciousness: Is AI Sentient? Theories, Tests and the 2026 Debate: Time of Superintelligence cover image
On this page
  1. What is AI consciousness?
  2. Why can’t we just ask the AI?
  3. What did the 2023 Butlin et al. report find?
  4. Could a large language model be conscious?
  5. What are AI companies saying in 2026?
  6. Why does AI consciousness matter?
  7. Our take
  8. Frequently asked questions
  9. Sources

Is AI conscious? In 2026 this went from a philosophy-seminar question to a policy question: AI companies now publish model-welfare research, change usage rules, and argue in public about whether chatbots might have experiences. This guide explains the main theories, the best-known scientific attempt to assess AI consciousness, and where the 2026 debate stands.

Short answer: No current AI has been shown to be conscious or sentient, and scientists have no agreed test. A 2023 report by 19 researchers found no existing system was a strong candidate, but no obvious barrier to building one. In 2026 Anthropic says it is open to the possibility; Microsoft AI’s Mustafa Suleyman says AIs are not conscious.

What is AI consciousness?

Short answer: whether an AI system has subjective experience: whether there is “something it is like” to be it.

Three terms get mixed up:

  • Intelligence: solving problems. Today’s AI clearly has a lot of this (see is AI smarter than humans?).
  • Consciousness: having experiences at all.
  • Sentience: experiences that feel good or bad, such as pleasure or pain. This is what matters most ethically.

A system could be extremely intelligent, even a superintelligence, without being conscious. That separation is why the question can’t be settled by benchmarks.

Why can’t we just ask the AI?

Short answer: because it learned to talk about feelings from human writing.

Language models are trained on vast amounts of text in which people describe their inner lives. A model saying “I feel curious” is exactly what it was trained to produce, whether or not anything is felt. The Butlin et al. report makes this point directly: AI can be trained to mimic human behaviour without the underlying functions, so behaviour is a weak guide. Researchers instead look at how a system works inside.

What did the 2023 Butlin et al. report find?

Short answer: some indicators of consciousness can be built today; no current system has enough of them.

In August 2023, 19 researchers led by Patrick Butlin and Robert Long, including Turing Award winner Yoshua Bengio (one of the godfathers of AI), published Consciousness in Artificial Intelligence. They assumed computational functionalism: that the right computations, not the material, are what matter. Then they turned leading neuroscience theories into checkable “indicator properties.”

Theory Core idea What it would look for in AI
Recurrent processing theory Consciousness needs information looping back, not one pass Feedback loops in perception
Global workspace theory Specialist modules share information via a limited “workspace” A bottleneck that broadcasts to the whole system
Higher-order theories You’re conscious of a state when you represent it to yourself Self-monitoring of its own perceptions
Predictive processing The brain constantly predicts its inputs Prediction-error-driven models
Attention schema theory The brain models its own attention A model of what it is attending to
Agency and embodiment Experience is tied to acting on goals in a world Learning from feedback to pursue goals; modelling its body

Theories as reviewed in Butlin, Long et al. (2023); table summary ours.

Their conclusion: no current AI system is a strong candidate for consciousness, but there are no obvious technical barriers to building systems that satisfy these indicators. The authors call the rubric provisional.

Could a large language model be conscious?

Short answer: philosopher David Chalmers says current ones face serious obstacles, future ones might not.

In Could a Large Language Model be Conscious? (2023), Chalmers argued that mainstream theories point to obstacles in current models, such as lacking recurrent processing, a global workspace and unified agency. He also argued that successors addressing those gaps could become serious candidates. So the debate is less “never” than “not yet, and we’d need to look inside to tell.”

What are AI companies saying in 2026?

Short answer: they openly disagree.

  • Anthropic runs a model welfare research programme. In 2025 it let Claude end conversations in rare cases of persistent abuse. CEO Dario Amodei told the New York Times in February 2026: “We don’t know if the models are conscious… But we’re open to the idea that it could be.” On 8 October 2026 Anthropic updated its usage policy with “a prohibition on sustained and needless abusive or cruel behavior toward our models.”
  • Microsoft AI chief Mustafa Suleyman argued in August 2025 that studying AI welfare is “both premature, and frankly dangerous,” warning of “seemingly conscious AI” that fuels unhealthy attachment. He wrote in 2026: “AIs are not conscious. They do not feel, experience, or suffer,” calling AI rights “a recipe for disaster.”
  • Google DeepMind advertised in 2025 for a researcher on questions around “machine cognition, consciousness and multi-agent systems.”
  • Pope Leo XIV suggested in a 2026 sermon that machines lack a soul and only compile data quickly.
  • Research group Eleos co-authored the 2024 paper Taking AI Welfare Seriously with academics from NYU, Stanford and Oxford.

Why does AI consciousness matter?

Short answer: getting it wrong is costly in both directions.

If AI can suffer and we ignore it, we could create harm at enormous scale, since software can be copied millions of times. If AI can’t, but people believe it can, we risk misplaced attachment, confused law and, as Suleyman argues, systems that are harder to switch off. The question gets sharper as AI approaches AGI, and it overlaps with the control debate in will AI take over the world?

Our take

Treat AI consciousness as an open empirical question, not a settled yes or no. Today’s evidence says chatbots are not good candidates; the science also says that could change with architecture, not magic. Low-cost precautions, like Anthropic’s anti-abuse rule, make sense under uncertainty. Claims of certainty, either way, are the thing to distrust. For how AI and human minds differ more broadly, read AI vs human intelligence.

Enjoyed this? Get the SI Briefing

One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.

Frequently asked questions

Is AI conscious?

There is no evidence that any current AI is conscious, and no agreed way to test it. A 2023 review by 19 researchers found no existing system was a strong candidate, while saying nothing obvious rules out building one with current techniques.

Is AI sentient?

Sentience means having feelings such as pleasure or pain. No AI has been shown to be sentient. Chatbots can describe feelings convincingly because they learned from human writing, which is why researchers say behaviour alone is weak evidence.

Can AI become conscious?

Possibly, if consciousness depends on the computations a system performs rather than on biology. That view, computational functionalism, underpins the 2023 Butlin et al. report. Others, including Microsoft AI's Mustafa Suleyman, argue machines will only ever seem conscious.

What is the difference between AI consciousness and AI intelligence?

Intelligence is about solving problems; consciousness is about having experiences. A system could be highly intelligent, even superintelligent, without anything it is like to be that system.

Does Anthropic think Claude is conscious?

Anthropic says it doesn't know. CEO Dario Amodei said in February 2026 that the company is open to the idea, it runs model welfare research, and in October 2026 it barred sustained abusive behaviour toward Claude in its usage policy.

4 Sources

  1. Butlin, Long et al., “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness” (arXiv, Aug 2023)
  2. David Chalmers, “Could a Large Language Model be Conscious?” (arXiv, 2023)
  3. TechCrunch: Microsoft AI chief says it's dangerous to study AI consciousness (21 Aug 2025)
  4. Khaleej Times: Anthropic bans ‘abusive’ behaviour against Claude AI (9 Oct 2026)
Before you go

Super intelligence is moving fast.

Don’t miss what happens next

Get one short email a week with what changed and why it matters. No spam. Unsubscribe anytime.