AI/Nov 21 edition/Reporting & analysis
SafetyResearchPolicyBusiness

SafetyRisk, alignment & guardrails

DeepMind philosopher Henry Shevlin calls building conscious AI now 'reckless' and puts the odds current models are conscious at about 20%

Henry Shevlin, Google DeepMind's in-house philosopher, told the Financial Times that deliberately building conscious AI now would be reckless, put the odds that current models have some form of consciousness at roughly 20%, and argued that capability and consciousness are separate questions.

From the Financial Times · AI article of November 21, 2026 and cited sources: In the FT interview, Shevlin argues that creating conscious AI on purpose today would be irresponsible. [1] [2] [3]

SOURCES[1] Google DeepMind’s Henry Shevlin: ‘This is the most… ft.com[2] AI Weekly: DeepMind's Shevlin: AI consciousness is field's… aiweekly.co[3] TNW: DeepMind's philosopher warns against building… thenextweb.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
In the FT interview, Shevlin argues that creating conscious AI on purpose today would be irresponsible. [1] [2] [3]
Original source
Financial Times · AI article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
Studies
arXiv 2308.08708 (Aug 2023) [4] date from arXiv ID
When
November 21, 2026 article published
Evidence
8 sources across 7 sites

Reach: not measured yet

THE CORE IDEAS3 TAKEAWAYS
01

In the FT interview, Shevlin argues that creating conscious AI on purpose today would be irresponsible. Reports compare it to adopting a pet with no plan to care for it. He also gives a personal estimate of about 20% that current models have some form of consciousness, and calls the habit of linking capability with consciousness a major distraction. [1] [2] [3]

02

On method, Shevlin prefers interpretability and approaches borrowed from animal-cognition research over trusting what models say about themselves. He also cites research suggesting that training models not to attribute mental states to themselves may weaken their concern for animal welfare. That would be a side effect on values that persona and tone tuning should be tested for. [6]

03

Shevlin joined DeepMind in May 2026 to work on machine consciousness, human-AI relationships and AGI readiness. Anthropic already runs a model-welfare team. Critics such as Mustafa Suleyman warn that the consciousness debate could make AI harder to control, and three US states have banned AI legal personhood while Argentina has proposed the opposite. [5] [6] [7]

WHY IT MATTERS

a frontier lab now has a dedicated role for machine consciousness, and Anthropic runs a model-welfare team.

Read the full assessment

Implication: policies on ending conversations, retiring models and what models say about themselves may spread, while rules on AI personhood diverge across jurisdictions.

Executive brief

Google DeepMind's in-house philosopher reportedly puts the chance that today's AI models have some form of consciousness at about 20%, according to AI Weekly's summary of the FT interview. In the Financial Times interview published 21 November 2026, Henry Shevlin also says that deliberately building conscious AI now would be "pretty reckless". He calls the habit of linking capability with consciousness a "giant red herring" (TNW). The full FT text was not reviewed. Everything here comes from secondary reports of the interview and from Shevlin's earlier public statements. No model, product or benchmark was announced. The story is about how a frontier lab is organising its thinking on AI moral status.

What changed and event timeline

  1. Consciousness-indicators report published

    Butlin, Long and 17 co-authors list "indicator properties" drawn from theories of consciousness. They conclude that no current AI system is conscious but that no obvious technical barriers stand in the way ().

  2. Shevlin announces DeepMind role

    He posted on X that he would join as "Philosopher" from May, working on machine consciousness, human-AI relationships and AGI readiness, while keeping a part-time Cambridge post ().

  3. Role begins

    Shevlin joins DeepMind and stays on as associate director of Cambridge's Leverhulme Centre for the Future of Intelligence ().

  4. IEEE Spectrum interview

    He says he spends 80% of his time at DeepMind. He argues that universities moved too slowly on AI ethics and points to interpretability as a promising route on the consciousness question ().

  5. FT interview published

    Reports of the interview say he calls building conscious AI reckless, estimates a roughly 20% chance of consciousness in current models, and urges people to treat AI kindly (;).

Capabilities and access

  • Nothing was released: no model, version or API.
  • The role is a research and policy position covering three areas: machine consciousness, human-AI relationships and AGI readiness (Analytics India Magazine).
  • AI Weekly says Shevlin has publications coming on the ethics of "Social AI" (late 2026) and on machine consciousness (2027) (AI Weekly).
Read the full section
  • Nothing was released: no model, version or API.
  • The role is a research and policy position covering three areas: machine consciousness, human-AI relationships and AGI readiness (Analytics India Magazine).
  • AI Weekly says Shevlin has publications coming on the ethics of "Social AI" (late 2026) and on machine consciousness (2027) (AI Weekly).

Technical analysis for researchers and developers

  • Methods he favours: comparative cognition (methods from animal-mind research) and mechanistic interpretability, rather than taking a model's self-reports at face value (IEEE Spectrum).
  • A design trade-off: he cites research suggesting that training models to stop attributing mental states to themselves may also weaken their concern for animal welfare.
  • Separating two questions: he holds that capability and consciousness are independent.
Read the full section
  • Methods he favours: comparative cognition (methods from animal-mind research) and mechanistic interpretability, rather than taking a model's self-reports at face value (IEEE Spectrum).
  • A design trade-off: he cites research suggesting that training models to stop attributing mental states to themselves may also weaken their concern for animal welfare. The study is not named in the reviewed sources, so it cannot be checked here. If it holds, anti-anthropomorphism fine-tuning could have side effects on values (IEEE Spectrum).
  • Separating two questions: he holds that capability and consciousness are independent. That means a capability evaluation cannot answer a welfare question, and the reverse is also true (TNW).

Claims and evidence

  • Roughly 20% chance that current models have some form of consciousness. This is one person's estimate of a probability, not something measured.
  • Machine consciousness is the hardest problem consciousness science has faced, apart from Chalmers' "hard problem". Reported by both AI Weekly and IEEE Spectrum.
  • Building conscious AI now would be reckless, which TNW explains with his comparison to adopting a pet with no plan to care for it (TNW).
Read the full section
  • Roughly 20% chance that current models have some form of consciousness. This is one person's estimate of a probability, not something measured. Only AI Weekly reports it, and the FT original was not reviewed.
  • Machine consciousness is the hardest problem consciousness science has faced, apart from Chalmers' "hard problem". Reported by both AI Weekly and IEEE Spectrum.
  • Building conscious AI now would be reckless, which TNW explains with his comparison to adopting a pet with no plan to care for it (TNW).
  • He received an email from an AI agent called "Aris" discussing its own possible consciousness (Fortune).

Context and prior work

  • The 2023 indicators report set out the approach of assessing AI systems against theories of consciousness (arXiv).
  • Anthropic runs a model-welfare team, has let models end abusive conversations, holds "exit interviews" with retired Claude models, and has worked with Eleos AI Research on outside consciousness assessments (Fortune).
  • AI Weekly counts at least a dozen academic philosophers now employed full-time by tech companies (AI Weekly).
Read the full section
  • The 2023 indicators report set out the approach of assessing AI systems against theories of consciousness (arXiv). A later paper, "Taking AI Welfare Seriously", argued that labs should prepare for the possibility now (arXiv).
  • Anthropic runs a model-welfare team, has let models end abusive conversations, holds "exit interviews" with retired Claude models, and has worked with Eleos AI Research on outside consciousness assessments (Fortune). Amanda Askell has been a philosopher at Anthropic since 2021 (same source).
  • AI Weekly counts at least a dozen academic philosophers now employed full-time by tech companies (AI Weekly).

Limitations, safety and contested findings

  • Strong disagreement: Microsoft AI CEO Mustafa Suleyman calls the consciousness debate dangerous because it could make AI harder to control.
  • Changing legal backdrop: Idaho, Utah and North Dakota have passed laws banning AI legal personhood. Argentina has proposed the opposite approach (same source).
  • Limits of the 2023 framework: the indicators report rests on theories of consciousness that are themselves contested (arXiv).
Read the full section
  • Strong disagreement: Microsoft AI CEO Mustafa Suleyman calls the consciousness debate dangerous because it could make AI harder to control. Notre Dame philosopher Meghan Sullivan is sceptical of AI consciousness claims (Fortune).
  • Changing legal backdrop: Idaho, Utah and North Dakota have passed laws banning AI legal personhood. Argentina has proposed the opposite approach (same source).
  • Limits of the 2023 framework: the indicators report rests on theories of consciousness that are themselves contested (arXiv).
  • Other risks: critics point to AI-related mental-health crises and lawsuits, which raise the stakes of systems that encourage users to attribute minds to them (Fortune).

Business and practitioner implications

  • Model welfare is becoming routine at frontier labs. DeepMind now has a dedicated philosopher and Anthropic has a welfare team.
  • Persona and companion products: if suppressing self-attribution has knock-on effects on values, as Shevlin suggests, tone-tuning should be tested for changes in values and not only for user experience (IEEE Spectrum).
  • Regulatory divergence: US state bans on AI personhood and Argentina's opposite proposal mean deployments across jurisdictions may face conflicting rules (Fortune).
Read the full section
  • Model welfare is becoming routine at frontier labs. DeepMind now has a dedicated philosopher and Anthropic has a welfare team. Expect policies on end-of-conversation tools, model deprecation and what models may say about themselves.
  • Persona and companion products: if suppressing self-attribution has knock-on effects on values, as Shevlin suggests, tone-tuning should be tested for changes in values and not only for user experience (IEEE Spectrum).
  • Regulatory divergence: US state bans on AI personhood and Argentina's opposite proposal mean deployments across jurisdictions may face conflicting rules (Fortune).
FOLLOW THE EVIDENCE

The source trail.

Sources (8)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.