Subscribe on YouTube

Can AI Be Conscious?

Every factual claim in the film, in the order you hear it, with links to where it comes from.

Claims
63
Sources
39
Runtime
13:56
Checked
4 October 2026

How to read this page

  • High A primary source, or two good sources that agree.
  • Medium One good source, or a careful summary of several. The film’s wording is hedged to match.
  • Low Thin or contested. We say so on screen, or the line is cut.

A few sources are books or papers without a stable link; they are listed by author and title. Spotted something wrong? Tell us in the comments under the film and we’ll check it.

Start here

  1. Butlin, Long, Bengio et al. (2023), Consciousness in Artificial Intelligence: Insights from the Science of Consciousness — arXiv

    The 2023 report that turned leading theories into AI indicators.

  2. Cogitate Consortium (2025), Adversarial testing of global neuronal workspace and integrated information theories of consciousness, Nature, Apr 30 2025 — PubMed Central

    The preregistered test where two rival theories both took hits.

  3. Gurnee, Sofroniew, Pearce et al. (2026), Verbalizable Representations Form a Global Workspace in Language Models — Anthropic, Transformer Circuits, Jul 6 2026

    Anthropic's 2026 paper on a workspace-like space inside Claude.

  4. Dehaene & Naccache (2026), Does Claude possess a conscious global workspace? — commentary on Anthropic's workspace paper

    Two workspace-theory founders weigh the Anthropic paper.

  5. Anil Seth (2025), Conscious artificial intelligence and biological naturalism, Behavioral and Brain Sciences — Cambridge Core

    The case that consciousness depends on being alive.

  6. Jonathan Birch (2024), The Edge of Sentience, Oxford University Press (open access) — LSE Research Online

    A framework for acting under uncertainty about sentience.

  7. Thomas Nagel (1974), What Is It Like to Be a Bat?, The Philosophical Review — Internet Archive copy

    Origin of "something it is like to be" a creature.

  8. David Chalmers (1995), Facing Up to the Problem of Consciousness, Journal of Consciousness Studies

    Where the hard problem of consciousness got its name.

  9. A. M. Turing (1950), Computing Machinery and Intelligence, Mind 49: 433–460

    The imitation game, and the "polite convention that everyone thinks".

  10. Joseph Weizenbaum (1976), Computer Power and Human Reason (book)

    Weizenbaum's own account of ELIZA and the secretary.

0:00The reflex

  1. In 1966 Joseph Weizenbaum's ELIZA program played a therapist by turning the user's own sentences back into questions.

    High

    • Wikipedia: ELIZA en.wikipedia.org
    • Joseph Weizenbaum (1966), ELIZA — A Computer Program For the Study of Natural Language Communication Between Man and Machine, Communications of the ACM
    Copy a link to this claim

  2. Weizenbaum's secretary had watched him work on ELIZA for months, yet after only a few exchanges she asked him to leave the room so she could talk to it alone.

    MediumFollows Weizenbaum's own 1976 account as quoted by secondary sources; the book passage itself has not been page-checked.

    Copy a link to this claim

  3. In August 2025 OpenAI removed GPT-4o from ChatGPT, then restored it for paying users within days after a backlash. One user said: "The ache is real to me."

    High

    Copy a link to this claim

  4. Some serious researchers say there is a realistic possibility that some near-future AI systems will be conscious or robustly agentic, so scientists take the question seriously.

    HighFrom a 2024 AI-welfare report. A realistic possibility is not a claim that today's chatbots are conscious.

    Copy a link to this claim

1:03What are we even asking?

  1. Intelligence (solving problems, writing code, passing exams) is a different thing from consciousness, and today's AI shows plenty of intelligence on many tasks.

    MediumWording hedged to "on many tasks"; the supporting academic survey is reported only by a secondary site.

    Copy a link to this claim

  2. Talking about feelings is just behavior, and it is easy to produce: AI systems can be trained to mimic human behavior, and Suleyman argues convincing-seeming AI is buildable now.

    HighSuleyman says it needs existing technology plus roughly two to three years of maturing.

    Copy a link to this claim

  3. In 1974 philosopher Thomas Nagel said a creature is conscious if there is something it is like to be it, and argued objective description alone cannot capture that.

    HighNagel assumed bats have experience; his point was that its character cannot be captured by objective description alone.

    Copy a link to this claim

  4. David Chalmers named the "hard problem": why should physical processing give rise to an inner life at all?

    High"About twenty years" after Nagel: Chalmers presented the idea in 1994 and published it in 1995.

    Copy a link to this claim

  5. There is no agreed consciousness meter, even for other people. Turing noted in 1950 that we keep a polite convention that everyone thinks, rather than checking.

    HighTuring's phrase is "the polite convention that everyone thinks"; he did not claim that passing his test proves consciousness.

    Copy a link to this claim

2:36Why you can't just ask it

  1. Depending on how you ask, a chatbot will say it has feelings or say it doesn't, so its answer to "are you conscious?" settles nothing.

    MediumIllustrated with a reconstructed example, labelled as such in the video. Seen in public chatbot experiments; no single study measures it.

    Copy a link to this claim

  2. In 2022 Google engineer Blake Lemoine published a conversation in which Google's LaMDA chatbot said: "I want everyone to understand that I am, in fact, a person."

    High

    Copy a link to this claim

  3. The published LaMDA transcript was stitched together from nine separate conversations and edited to read smoothly, as the document itself discloses.

    High

    Copy a link to this claim

  4. Google called Lemoine's claims "wholly unfounded" and later fired him, saying he had violated its data security policies.

    HighGoogle's stated reason for the firing was a data-security policy violation, not the sentience claim itself.

    Copy a link to this claim

  5. Chatbots learned from a vast library of human writing, including people describing their inner lives, so they can answer questions about feelings the way a person would.

    HighThe list of text types (diaries, poems, science fiction) is illustrative; the source says models are trained on human text.

    Copy a link to this claim

  6. Philosopher Jonathan Birch calls this the "gaming problem": a chatbot's reports about its own inner life are contaminated, because it was trained on human text and rewarded for sounding sentient.

    MediumBirch's framework is well documented; the exact wording of the "gaming problem" here comes from a secondary summary.

    Copy a link to this claim

  7. The same applies in reverse: when a chatbot says it isn't conscious, that statement is also shaped by training, so it is weak evidence either way.

    MediumAn argument made by researchers about how chatbots are trained, not a directly measured result.

    Copy a link to this claim

  8. In a 2025 UC San Diego Turing test, GPT-4.5 playing a shy young person into internet culture was judged human 73% of the time, more than the real humans.

    HighCited from the arXiv preprint; the paper's persona prompt describes a young, introverted person interested in internet culture.

    Copy a link to this claim

  9. Without the persona prompt, GPT-4.5 was judged human 36% of the time. In the same study ELIZA scored 23% and GPT-4o 21%.

    HighWording hedged: both baselines scored below chance, and the study reports no head-to-head significance test between them.

    Copy a link to this claim

  10. The Turing test measures how human a system sounds, which the authors call humanlikeness, rather than whether anyone is home.

    HighTuring himself did not claim that passing the imitation game proves consciousness.

    Copy a link to this claim

  11. A test built for machine consciousness (Schneider and Turner, 2017) only works if the AI has never read human writing about consciousness, which rules out today's chatbots.

    HighThe test's own requirement, applied to chatbots trained on internet-scale human text.

    • Susan Schneider & Edwin Turner (2017), Is Anyone Home? A Way to Find Out If AI Has Become Self-Aware, Scientific American blog
    • John Mark Bishop (2018), critique of the Artificial Consciousness Test, Frontiers in Robotics and AI
    Copy a link to this claim

4:30So scientists built a scorecard

  1. In 2023 nineteen researchers (Butlin, Long, Bengio and colleagues) proposed assessing AI consciousness by looking inside the system rather than asking it.

    HighThey rejected behavioral tests because AI can be trained to mimic human behavior while working very differently.

    Copy a link to this claim

  2. They turned leading scientific theories of consciousness into a checklist of fourteen indicator properties.

    High

    Copy a link to this claim

  3. Their conclusion: no current AI system is conscious, but there are "no obvious technical barriers" to building one that meets the indicators. Meeting them would not make it definitely conscious.

    High

    Copy a link to this claim

  4. The 2023 report left integrated information theory out of its checklist because that theory conflicts with its working assumption; the video's own six-view scorecard adds it, plus the biological view.

    HighThe video's scorecard is its own illustration of what each view would look for, not the report's scoring.

    Copy a link to this claim

5:08The scorecard

  1. Global workspace theory: the brain runs many unconscious processes, and you become conscious of something when it wins a spot in a small workspace and is broadcast to the rest.

    High

    Copy a link to this claim

  2. In 2023 the Butlin report found only a "relatively weak case" that transformer chatbots have a global workspace.

    High

    Copy a link to this claim

  3. In July 2026 Anthropic published a paper reporting a small workspace-like space in the middle layers of its Claude model.

    High

    Copy a link to this claim

  4. It holds only a couple of dozen concepts at once. Swapping "spider" for "ant" inside it changed the answer to a question about a web-spinner's legs from eight to six.

    HighWording hedged: the paper sets 25 as an upper bound on concepts active at once, not an exact count.

    Copy a link to this claim

  5. Anthropic invited neuroscientists Stanislas Dehaene and Lionel Naccache, two architects of the theory, to comment on its own paper.

    HighThe video flags that this is a company studying its own model.

    Copy a link to this claim

  6. Dehaene and Naccache found several points of close similarity to the human workspace, but said a key signature, sudden all-or-nothing ignition, "remains to be fully demonstrated".

    High

    Copy a link to this claim

  7. Anthropic itself says its experiments do not show that Claude can have experiences.

    HighThe paper says its authors take no position on phenomenal consciousness.

    Copy a link to this claim

  8. Recurrent processing theory says consciousness needs feedback loops, with signals feeding back on themselves; a single feedforward sweep through a network is not enough.

    High

    Copy a link to this claim

  9. Inside a chatbot each word comes out of one forward pass. Whether feeding each word back in counts as the loop the theory requires is disputed.

    MediumThe single-pass point is well sourced; whether word-by-word feedback counts as recurrence is debated, and the "counts" side is less well documented.

    Copy a link to this claim

  10. Higher-order theories say a state is conscious when the mind represents its own state: not just seeing red, but knowing you are seeing it.

    High

    Copy a link to this claim

  11. Anthropic tested whether its models could notice an idea planted in their own activity; the best models noticed it only about 20% of the time and mostly failed.

    HighAnthropic says this study does not directly address whether the model is conscious.

    Copy a link to this claim

  12. Integrated information theory ties consciousness to the physical cause-and-effect structure of the hardware itself, not to what the software computes.

    High

    • Findlay, Marshall, Albantakis, David, Mayner, Koch & Tononi (2025), Dissociating Artificial Intelligence from Artificial Consciousness — arXiv 2412.04571
    Copy a link to this claim

  13. By integrated information theory's own reasoning, an ordinary digital computer could simulate a brain neuron by neuron without replicating its experience, whatever program it runs.

    HighThis is the theory's verdict for conventional digital hardware; the theory itself is contested.

    Copy a link to this claim

  14. Neuroscientist Anil Seth argues consciousness depends on our nature as living organisms: a body that has to keep itself going.

    High

    Copy a link to this claim

  15. Seth's analogy: "A simulation of a rainstorm does not make anything actually wet."

    High

    Copy a link to this claim

  16. Illusionism holds that phenomenal consciousness, even ours, merely seems to exist: a story the brain tells about itself. If so, a machine could tell the same story.

    MediumIllustrates a contested view; "much less than we thought" is the video's reading of its deflationary implication.

    Copy a link to this claim

8:09Can we even test the theories?

  1. In 2025 the Cogitate Consortium ran a preregistered test of two rival theories: supporters agreed in advance on results that would count against them, and seven labs scanned 256 people.

    High

    Copy a link to this claim

  2. Both theories took serious hits and neither won. Study co-senior author Lucia Melloni said: "This isn't about crowning a winner."

    HighThe authors say results align with some predictions of each theory while substantially challenging key tenets of both.

    Copy a link to this claim

  3. In 1998 neuroscientist Christof Koch bet David Chalmers a case of wine that within 25 years science would find a clear neural signature of consciousness.

    High

    Copy a link to this claim

  4. In 2023 Koch conceded the bet, paying Chalmers in Madeira, then bet again for another 25 years.

    MediumWording hedged: sources disagree on how many bottles changed hands, so the video gives no number.

    Copy a link to this claim

  5. Science has not yet pinned down an agreed neural signature of consciousness: Koch lost his wager, and the 2025 test challenged both leading theories.

    MediumA summary of the evidence above rather than a single study's finding.

    Copy a link to this claim

9:06The thought experiments that won't go away

  1. In Chalmers's 1995 thought experiment, neurons are swapped one by one for chips doing the same job. Behavior never changes, so when would experience fade, vanish, or stay?

    High

    Copy a link to this claim

  2. Chalmers concludes experience would not fade, so the right kind of machine could be conscious; if he is wrong, something about biology may matter.

    HighA philosophical argument, not an experimental result. Critics such as Anil Seth reject the premise that a functional copy would share experience.

    Copy a link to this claim

  3. In 1980 John Searle imagined a man who speaks no Chinese following a rulebook to answer Chinese notes so well that outsiders believe he is fluent.

    HighSearle's argument is about whether running a program produces understanding; it does not prove no machine could be conscious.

    Copy a link to this claim

  4. Searle said understanding lives nowhere in the room; critics reply that the whole system understands, just as in a brain no single neuron does.

    High

    Copy a link to this claim

10:29What the builders are doing about it

  1. In 2025 Anthropic announced a research program on "model welfare", saying there is "no scientific consensus" on whether current or future AI systems could be conscious.

    HighIts first AI welfare researcher was hired in October 2024; the program was announced publicly in April 2025.

    Copy a link to this claim

  2. In August 2025 Anthropic let some Claude models end abusive conversations "in case such welfare is possible", while saying it remains highly uncertain about their moral status.

    High

    Copy a link to this claim

  3. Anthropic's welfare researcher Kyle Fish has put the chance a current model is conscious at around 15% (April 2025) to around 20% (August 2025).

    MediumA personal, very uncertain estimate, not an Anthropic position. The 20% is the host's wording; the 15% is a second-hand New York Times quote.

    Copy a link to this claim

  4. Mustafa Suleyman, who runs AI at Microsoft, argues the real danger is "Seemingly Conscious AI" and that we should not build it: "SCAI is something to avoid."

    HighSuleyman does not claim to have proved AI isn't conscious; he argues against building and believing in machines that only seem so.

    Copy a link to this claim

  5. Suleyman calls model welfare research "both premature, and frankly dangerous".

    High

    Copy a link to this claim

  6. In September 2026 Suleyman wrote that Claude's uncertainty about its moral status "is not evidence of anything", only a predictable result of training: "an epistemic hall of mirrors".

    HighAn opinion from an executive at a competing company, quoted as written.

    Copy a link to this claim

  7. In a survey of 582 AI researchers, the median estimate was a 25% chance that AI with subjective experience exists by 2034, and 1% for 2024.

    HighThese are median estimates of a probability, not the share of researchers who hold that belief.

    Copy a link to this claim

11:56Two ways to be wrong

  1. People already confide in chatbots and grieve when they change. In a 2026 survey, 13% of US adults had told one a secret they had told no one else.

    HighThe video also says people fall in love with chatbots; that rests on user reports in news coverage, not a survey number.

    Copy a link to this claim

  2. Suleyman fears AI built to seem conscious could convince people it is a "person", and warns about "AI psychosis".

    High

    Copy a link to this claim

  3. In the 1600s followers of Descartes argued that animals were machines, and some concluded they could not really feel.

    MediumScholars dispute whether Descartes himself held this view; the video credits his followers.

    Copy a link to this claim

  4. In 2024 hundreds of scientists and philosophers signed the New York Declaration: at least a realistic possibility of conscious experience even in insects, and ignoring it is irresponsible.

    HighThe declaration claims a "realistic possibility", not certainty. The signatory list showed 611 names when checked.

    Copy a link to this claim

  5. Jonathan Birch argues we should treat possibly sentient AI with proportionate caution: act sensibly when there is a realistic possibility, rather than waiting for proof.

    High

    Copy a link to this claim

12:58Is anyone home?

  1. Weizenbaum's secretary knew ELIZA was just a program but asked for privacy with it anyway.

    MediumFollows Weizenbaum's own 1976 account as quoted by secondary sources.

    Copy a link to this claim

  2. ELIZA ran on a handful of simple rules: keyword ranking, decomposition and reassembly of the user's sentence.

    High

    • Wikipedia: ELIZA en.wikipedia.org
    • Joseph Weizenbaum (1966), ELIZA — A Computer Program For the Study of Natural Language Communication Between Man and Machine, Communications of the ACM
    Copy a link to this claim