Skip to content
Back to the question

Evidence record

Could a machine be conscious?

Every sentence on the exploration belongs to at least one claim below. Each says what kind of statement it is, what it rests on, what the evidence does not establish, and how far it was checked. Nothing here is a score, and none of it is a verdict on a belief or a theory.

Last checked 24 September 2026.

  • A judgement, not a fact: 11
  • Source checked: 31
  • Presented as open: 1

If something here is wrong, misquoted, out of date or unfair, write to support@unknownoceans.com and give the claim number, which does not change.

What the question takes for granted

A question that is part science and part philosophy. Whether a given system has the properties theories of consciousness point to can be investigated; whether those theories, and the assumption that computation is what matters, are right cannot yet be. The page reports the method and where it stops.

  • Consciousness is a real property a system either has or lacks

    Contested

    The report the page leans on writes as though it is all or nothing, for convenience, and says it may instead come in degrees or be indeterminate. Chalmers assumes consciousness is real and not an illusion, and calls that a substantive assumption.

  • A machine could be conscious at all

    Contested

    Possible in principle if computational functionalism is true, which the 2023 report adopts only as a working hypothesis. Biological views and integrated information theory deny it for conventional digital hardware.

How the question is labelled

Claim: The question we cannot ask it

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s title. It follows from claims 13 and 14: asking a system trained on human text whether it is conscious gives an answer that says little either way.

Claim 1 · checked 24 September 2026 · AI-assisted audit

Claim: Machines

Kind
Framing. Signposting. It asserts nothing a source could check.

The question

Claim: A system that describes its inner life fluently is weak evidence of one, because that is what it was trained on.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

The subtitle states the point the first section makes.

  • Supports

    Could a Large Language Model be Conscious?section 2, “X = Self-Report”

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: LaMDA has actually been trained on a giant corpus of people talking about consciousness. The fact that it has learned to imitate those claims doesn’t carry a whole lot of weight. ... If it has been trained on this material, the evidence is much weaker.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of Consciousnesssection 1.2.3

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: we are sceptical about whether behavioural approaches to consciousness in AI can avoid the problem that AI systems may be trained to mimic human behaviour while working in very different ways

Corrected

24 September 2026. Previously: “A system that describes its inner life fluently is evidence of nothing, because that is what it was trained on.” Neither source says such reports are no evidence. Chalmers says they do not carry a whole lot of weight and that the evidence is much weaker when a system was trained on this material; the 2025 restatement of the indicator method says carefully designed behavioural tests could provide some evidence.

Claim 3 · checked 24 September 2026 · AI-assisted audit

Claim: There is no validated test for consciousness in a system unlike us

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Identifying Indicators of Consciousness in AI SystemsThe problem of AI consciousness; Looking ahead

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Review of the field

    The source reads: it remains unclear how they should (or even could) be validated, and tests for AI consciousness are an especially challenging case ... Validating our method in the AI case would be challenging because it would require the development of alternative assessment methods suited to AI

  • Supports

    Could a Large Language Model be Conscious?section 1; section 4, list of challenges

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: there’s no standard operational definition of consciousness ... Evidence: Develop benchmarks for consciousness.

Corrected

24 September 2026. Previously: “There is no test for consciousness that works on a system unlike us” Whether a test “works” is what cannot be known; what the sources say is that no method has been validated for AI, and that validating one would need an independent method that does not exist.

Claim 4 · checked 24 September 2026 · AI-assisted audit

Claim: and the usual informal test, asking it, is unreliable on anything trained to produce human text.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Identifying Indicators of Consciousness in AI SystemsInternal and behavioral evidence

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Review of the field

    The source reads: showing that in the case of AI, inferences from behavior to features of internal processes are often unreliable

  • Supports

    Could a Large Language Model be Conscious?section 2, “X = Self-Report”

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: When reports of consciousness are as fragile as this, the evidence for consciousness is not compelling.

Corrected

24 September 2026. Previously: “and the usual informal test, asking it, fails on anything trained to produce human text.” The sources call the inference unreliable and the evidence weak, not a test that always fails.

Claim 5 · checked 24 September 2026 · AI-assisted audit

Claim: The approach a 2023 report calls the most promising derives indicator properties from theories of consciousness and checks systems against them

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

“Most promising” is the report’s own judgement of its method, attributed to it. What the report does is described in claims 16 to 19.

Corrected

24 September 2026. Previously: “The most serious current approach derives indicator properties from theories of consciousness and checks systems against them” “Most serious” was a superlative no source states. The 2023 report calls its theory-heavy method “currently the most promising”; the page now attributes that judgement to it.

Claim 6 · checked 24 September 2026 · AI-assisted audit

Claim: which converts an unanswerable question into a partly answerable one.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading of the report’s own claim that the question is scientifically tractable. The part that becomes answerable is whether a system has the indicators; whether the indicators are right stays open (claim 19).

  • Context

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessExecutive summary, first contribution

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Showing that the assessment of consciousness in AI is scientifically tractable because consciousness can be studied scientifically and findings from this research are applicable to AI

Claim 7 · checked 24 September 2026 · AI-assisted audit

Claim: The traditional reason for attributing experience to other people is resemblance: a body like our own, and behaviour of the kind our own feelings cause.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Other Minds§1.1 The Argument from Analogy

    Cited · Link answered, 19 October 2026 · The cited passage was read, 24 September 2026 · Texts · Reference work

    The source reads: One form of reasoning has been taken to be “traditional” in this connection: the argument from analogy (Blackburn 1994).

  • Supports

    Other Minds§1.1, quoting Mill

    Cited · Link answered, 19 October 2026 · The cited passage was read, 24 September 2026 · Texts · Reference work

    The source reads: First, they have bodies like me, which I know in my own case, to be the antecedent condition of feelings; and because, secondly, they exhibit the acts, and outward signs, which in my own case I know by experience to be caused by feelings.

What this does not establish

Mill’s statement, quoted in the entry, names bodies and outward signs, and the sentence names only those.

Other readings

The entry reports that the argument from analogy came to be considered unfit for purpose, and that the argument from best explanation is the form most widely accepted today.

Corrected

24 September 2026. Previously: “We attribute experience to other people on the strength of resemblance: a similar body, a similar brain, similar behaviour when injured.” Stated as the whole of how we do it, the sentence took a side. Resemblance is the traditional answer, the argument from analogy; the reference work reports that inference to the best explanation is now the more widely accepted one.

24 September 2026. Previously: “The traditional reason for attributing experience to other people is resemblance: a similar body, a similar brain, similar behaviour when injured.” The argument from analogy as the reference work gives it, in Mill’s words, names bodies like one’s own and the outward signs that in one’s own case are caused by feelings. It names no brain and no behaviour when injured.

Claim 8 · checked 24 September 2026 · AI-assisted audit

Claim: A machine breaks the resemblance where it matters.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading. It follows from claim 10: behaviour can match while the body and the workings do not.

Corrected

24 September 2026. Previously: “A machine breaks the resemblance in every direction at once.” The next sentence says a machine may match the behaviour perfectly, so it does not break the resemblance in every direction.

Claim 9 · checked 24 September 2026 · AI-assisted audit

Claim: It may produce behaviour remarkably like ours in some ways while sharing little of the structure,

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

The report says “arguably”, and finds that some existing systems do share some of the computational features theories associate with consciousness, such as algorithmic recurrence.

Corrected

24 September 2026. Previously: “It may produce the behaviour perfectly while sharing none of the structure,” The report finds some indicator properties, such as algorithmic recurrence, already clearly met by existing AI systems, so “none” overstated the difference.

24 September 2026. Previously: “It may produce the behaviour perfectly while sharing little of the structure,” The report says the outputs of these systems are remarkably human-like in some ways, not that they match human behaviour perfectly.

Claim 10 · checked 24 September 2026 · AI-assisted audit

Claim: which is exactly the case in which our usual method stops working.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading: an argument from similarity has little to work with where the similarity is in behaviour alone and the behaviour was trained to match.

  • Context

    Other Minds§1.1

    Cited · Link answered, 19 October 2026 · The cited passage was read, 24 September 2026 · Texts · Reference work

    The source reads: Analogical reasoning cites similarities between two things and uses this as support for concluding that further similarities may be taken to exist.

Claim 11 · checked 24 September 2026 · AI-assisted audit

What can be said

Claim: What can be said

Kind
Framing. Signposting. It asserts nothing a source could check.

Fluency is weak evidence

Claim: Fluency is weak evidence

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Could a Large Language Model be Conscious?section 2, “X = Self-Report”

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: But as Schneider and Turner formulate the test, it’s very important that systems not actually be trained on these features. If it has been trained on this material, the evidence is much weaker.

Corrected

24 September 2026. Previously: “Fluency is not evidence” Chalmers calls the evidence much weaker, not absent, and the 2025 restatement of the indicator method allows that carefully designed behavioural tests could provide some evidence.

Claim 13 · checked 24 September 2026 · AI-assisted audit

Claim: A language model trained on human writing will describe feelings, preferences and inner states, because human writing does.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Could a Large Language Model be Conscious?section 2, “X = Self-Report”

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: LaMDA: The nature of my consciousness/sentience is that I am aware of my existence, I desire to learn more about the world, and I feel happy or sad at times. ... LaMDA has actually been trained on a giant corpus of people talking about consciousness.

What this does not establish

Chalmers also reports that a one-word change to the question got a language model to say it was not sentient: what such a system says about itself depends on how it is asked.

Claim 14 · checked 24 September 2026 · AI-assisted audit

Claim: That output is as available to a system with no experience as to one with experience, so on its own it does little to distinguish between them.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of Consciousnesssection 1.2.3

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: AI systems may be trained to mimic human behaviour while working in very different ways, thus “gaming” behavioural tests

  • Qualifies

    Identifying Indicators of Consciousness in AI SystemsInternal and behavioral evidence

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Review of the field

    The source reads: behavioral tests could provide some evidence for the presence of our indicators and show that they support consciousness-linked capacities in particular systems

Corrected

24 September 2026. Previously: “That output is as available to a system with no experience as to one with experience, so it cannot distinguish between them.” The authors are sceptical of behavioural tests, not certain they are empty, and allow that carefully designed ones could provide some evidence.

Claim 15 · checked 24 September 2026 · AI-assisted audit

Claim: This is the point the serious assessments start from, and it is negative.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

The two assessments the page draws on, the 2023 report and Chalmers, both set self-report aside before doing anything else.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessExecutive summary

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: The main alternative to this approach is to use behavioural tests for consciousness, but this method is unreliable because AI systems can be trained to mimic human behaviours while working in very different ways.

  • Supports

    Could a Large Language Model be Conscious?section 2, “X = Self-Report”

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: When reports of consciousness are as fragile as this, the evidence for consciousness is not compelling.

Corrected

24 September 2026. Previously: “This is the one firm result in the area, and it is negative.” It is an argument both sources make, not an experimental result, and the page cannot vouch that it is the only firm point in the field.

Claim 16 · checked 24 September 2026 · AI-assisted audit

Claim: The conversational test is weakest for precisely the systems people most want to apply it to.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Corrected

24 September 2026. Previously: “The conversational test is uninformative for precisely the systems people most want to apply it to.” Weak rather than uninformative, as in claim 15.

Claim 17 · checked 24 September 2026 · AI-assisted audit

There is a method that is not just asking

Claim: There is a method that is not just asking

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Claim 18 · checked 24 September 2026 · AI-assisted audit

Claim: A 2023 report by nineteen philosophers, neuroscientists and AI researchers took several prominent scientific theories of consciousness, derived from them computational properties that a conscious system would have, and assessed existing AI systems against them.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Nineteen authors are listed. Their affiliations include departments of philosophy, experimental psychology and neuroimaging, and AI institutes such as MILA.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessTitle page, author list

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Patrick Butlin* Robert Long* Eric Elmoznino Yoshua Bengio Jonathan Birch Axel Constant George Deane Stephen M. Fleming Chris Frith Xu Ji Ryota Kanai Colin Klein Grace Lindsay Matthias Michel Liad Mudrik Megan A. K. Peters Eric Schwitzgebel Jonathan Simon Rufin VanRullen

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessAuthors and affiliations, p. 2

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: George Deane, Department of Philosophy, University of Montreal ... Liad Mudrik, School of Psychological Sciences and Sagol School of Neuroscience ... Yoshua Bengio, University of Montreal and MILA - Quebec AI Institute

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessAbstract

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: We survey several prominent scientific theories of consciousness, including recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory. ... We use these indicator properties to assess several recent AI systems

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessExecutive summary

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Each of these indicator properties is said to be necessary for consciousness ... by one or more theories, and some subsets are said to be jointly sufficient.

What this does not establish

The systems assessed were a handful of 2023 examples: transformer language models, the Perceiver architecture, DeepMind’s Adaptive Agent, a virtual rodent and PaLM-E.

Later work looked for

Crossref, 2023 on, "indicators of consciousness AI systems Butlin": the method was restated by most of the same authors in Trends in Cognitive Sciences (2025 online, 2026 in print), which the page now cites; it does not change this description.

Corrected

24 September 2026. Previously: “A 2023 report by a group of consciousness researchers and computer scientists took the leading scientific theories of consciousness, derived from each a set of computational properties a conscious system would have, and assessed existing AI systems against them.” The report calls its theories “several prominent”, not the leading ones, and deliberately leaves out integrated information theory; several of its authors work in departments of philosophy.

Claim 19 · checked 24 September 2026 · AI-assisted audit

Claim: It assumed, as a working hypothesis, that the right computations are enough for consciousness, and so left out integrated information theory, which is incompatible with that assumption.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessExecutive summary

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: we adopt compu-tational functionalism, the thesis that performing computations of the right kind is necessary and sufficient for consciousness, as a working hypothesis ... We do not consider integrated information theory, because it is not compatible with computational functionalism.

What this does not establish

The report calls computational functionalism mainstream but disputed; the whole assessment is conditional on it.

Corrected

24 September 2026. Previously: “” The report’s central assumption was missing from the page, and the page’s next section turns on exactly that assumption.

Claim 20 · checked 24 September 2026 · AI-assisted audit

Claim: Their analysis suggested that no current system is conscious, that none of the properties looks technically out of reach, and that the assessment is only as good as the theories it draws on.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

The authors added a note that satisfying the indicators would not mean a system was definitely conscious. The 2025 restatement calls the survey of existing systems begun but far from complete.

Later work looked for

Crossref, 2023 on, "consciousness in AI indicator properties": the 2025 restatement keeps the method and calls the survey of existing systems far from complete; nothing found that changes the claim.

Corrected

24 September 2026. Previously: “Their conclusion was that no current system is conscious,” The report says its analysis “suggests” this, and that no current system appears to be a strong candidate; it does not state it as a conclusion.

Claim 21 · checked 24 September 2026 · AI-assisted audit

Claim: That last caveat is the honest part and the limiting one.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s judgement on claim 21. It follows from the report’s own statement that its rubric is provisional and depends on confidence in the theories.

Claim 22 · checked 24 September 2026 · AI-assisted audit

Claim: How the assessment works

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: Start from theories, not intuitions

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

The report contrasts theories with behavioural signatures; “intuitions” is the page’s word for the impression a system makes.

Claim 24 · checked 24 September 2026 · AI-assisted audit

Claim: Recurrent processing, global workspace, higher-order, predictive processing and attention schema theories each describe what a conscious system does.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

These are the theories the report used, not all the theories in the field: it also drew indicators from agency and embodiment and left out integrated information theory. The higher-order theories it used are computational versions.

Claim 25 · checked 24 September 2026 · AI-assisted audit

Claim: Derive indicator properties

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Claim 26 · checked 24 September 2026 · AI-assisted audit

Claim: Each theory is restated in computational terms, so that the question becomes whether a given architecture has a given property rather than whether it seems conscious.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessAbstract; section 1.2

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: elucidated in computational terms that allow us to assess AI systems for these properties ... A theory-heavy approach is one that focuses on how systems work, rather than on whether they display forms of outward behaviour

Corrected

24 September 2026. Previously: “rather than whether it seems alive.” The report contrasts its method with behaviour that seems characteristic of conscious beings. Life is not its subject.

Claim 27 · checked 24 September 2026 · AI-assisted audit

Claim: Assess actual systems

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Claim 28 · checked 24 September 2026 · AI-assisted audit

Claim: Existing models are checked against the indicators.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Claim 29 · checked 24 September 2026 · AI-assisted audit

Claim: The result is a profile rather than a verdict, which is the appropriate output for a question with no settled theory behind it.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading of a method that yields credences rather than yes or no. The report does still state a tentative overall finding about current systems (claim 21).

  • Context

    Consciousness in Artificial Intelligence: Insights from the Science of ConsciousnessExecutive summary

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Various theories are currently live candidates in the science of consciousness, so we do not endorse any one theory here. ... Our claim, however, is that AI systems which possess more of the indicator properties are more likely to be conscious.

Claim 30 · checked 24 September 2026 · AI-assisted audit

What it would depend on

Claim: What it would depend on

Kind
Framing. Signposting. It asserts nothing a source could check.

Whether the substrate matters

Claim: Whether the substrate matters

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: A philosopher working through the same question in 2023 reached a similar shape of answer from the other direction:

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The philosopher is David Chalmers. The talk was given in November 2022 and the paper published in 2023. “From the other direction” is the page’s reading: he weighs reasons for and against, rather than building a rubric from theories.

  • Context

    Could a Large Language Model be Conscious?title note

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Published in the Boston Review, August 9, 2023 (published version). This is an edited version of a talk given at the conference on Neural Information Processing Systems (NeurIPS) on November 28, 2022

Claim 33 · checked 24 September 2026 · AI-assisted audit

Claim: the obstacles in current models are specific and nameable, among them the absence of recurrent processing, of a global workspace and of unified agency, and he thought it quite possible that they would be overcome within a decade or so.

Kind
An argument. An argument a named thinker makes. That it is made is checked; whether it succeeds is argued.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Could a Large Language Model be Conscious?Abstract, arXiv page; section 4

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: there are significant obstacles to consciousness in current models: for example, their lack of recurrent processing, a global workspace, and unified agency. At the same time, it is quite possible that these obstacles will be overcome in the next decade or so.

What this does not establish

Chalmers also lists biology as a candidate requirement, which engineering on silicon would not remove, and gives it at least a one-in-three credence on mainstream assumptions. His numbers are illustrative credences, not predictions.

Corrected

24 September 2026. Previously: “and none of them looks permanent.” Chalmers says it is quite possible these obstacles will be overcome, not that none is permanent; biology, which he also weighs, is not one a machine could overcome.

Claim 34 · checked 24 September 2026 · AI-assisted audit

Claim: If consciousness depends only on what a system computes, a sufficiently faithful simulation has it.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Conscious Artificial Intelligence and Biological Naturalismsection 3.7

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Would simulating the neurobiological mechanisms underlying consciousness – whether exactly or approximately – be sufficient to realise consciousness? This would be guaranteed only if computation is sufficient for consciousness

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of Consciousnesssection 1.2.1

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: The material substrate of a system does not matter for consciousness except insofar as the substrate affects which algorithms the system can implement.

What this does not establish

The 2023 report adds that the simulation would have to run the right algorithm, not merely produce the same outputs.

Claim 35 · checked 24 September 2026 · AI-assisted audit

Claim: If it depends on the physical medium, on chemistry or on continuous rather than discrete dynamics, then no digital system does, however faithful.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Conscious Artificial Intelligence and Biological Naturalismsection 3.1

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: Non-computational neural functions could include processes relating to (continuous) electromagnetic fields, fine-grained timing relations (only order, not dynamics as such, matters for Turing computation), freely diffusing neurotransmitters, and so on.

  • Supports

    Consciousness in Artificial Intelligence: Insights from the Science of Consciousnesssection 1.2.1

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: It could be, for instance, that some non-computational feature of living organisms is necessary for consciousness (Searle 1980, Seth 2021), in which case consciousness would be impossible in non-organic artificial systems.

What this does not establish

On these views unconventional hardware, neuromorphic or biological, might still qualify; the claim is about digital systems.

Claim 36 · checked 24 September 2026 · AI-assisted audit

Claim: The theories in use disagree on this, and the question is still open.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

The 2025 paper sets computational functionalism beside biological views and integrated information theory, and says new arguments for or against it could help to provide clarity.

  • Supports

    Identifying Indicators of Consciousness in AI SystemsBox 2; Concluding remarks

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Review of the field

    The source reads: However, many theorists favor alternative views, and each of these other views raises different questions about AI consciousness.

  • Supports

    Conscious Artificial Intelligence and Biological Naturalismsection 3.6; section 5.3

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: The arguments so far do not disprove computational functionalism. But they do render it less plausible ... it would seem an open question whether these approximations would be good enough to realise consciousness

Corrected

24 September 2026. Previously: “The theories in use disagree on this and the disagreement is not decidable by any experiment anyone has proposed.” Seth proposes empirical measures that may bear on it (claim 38), so “no experiment anyone has proposed” was not true. The narrower statement is carried by both sources.

24 September 2026. Previously: “The theories in use disagree on this, and no experiment has settled it.” Neither source speaks of experiments on the question. Seth says the arguments so far do not disprove computational functionalism and calls it an open question whether digital approximations would be good enough to realise consciousness.

Claim 37 · checked 24 September 2026 · AI-assisted audit

Claim: Anil Seth, who argues that the medium does matter, suggests that measuring how far a brain’s activity can be abstracted from its finer-grained workings may help.

Kind
An argument. An argument a named thinker makes. That it is made is checked; whether it succeeds is argued.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Conscious Artificial Intelligence and Biological NaturalismAbstract; section 3.5

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: I’ll instead make the case that consciousness depends on our nature as living organisms ... In practice, such measures may help empirically determine whether, or to what extent, neural dynamics can be abstracted away from finer-grained levels of description.

What this does not establish

Seth is one side of the dispute. Such a measure would bear on how far the brain’s workings are separable from its material; he does not claim it would settle whether a machine could be conscious.

Later work looked for

Crossref, 2024 on, "conscious artificial intelligence biological naturalism": Seth’s target article drew published commentaries, including one arguing that consciousness is probably computational; not read.

Corrected

24 September 2026. Previously: “” Added so that the page reports the empirical route the rival side proposes, with who proposes it. Source added to the passage.

Claim 38 · checked 24 September 2026 · AI-assisted audit

Whether we would notice

Claim: Whether we would notice

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: There is a detection method for humans in ambiguous states: perturb the cortex, measure how complex the response is, and compute an index that, in the people it has been validated on, separates conscious from unconscious states reliably.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    A Theoretically Based Index of Consciousness Independent of Sensory Processing and BehaviorAbstract (Europe PMC record; the same text in the PDF)

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: PCI is calculated by (i) perturbing the cortex with transcranial magnetic stimulation (TMS) to engage distributed interactions in the brain (integration) and (ii) compressing the spatiotemporal pattern of these electrocortical responses to measure their algorithmic complexity (information).

  • Supports

    Stratification of Unresponsive Patients by an Independently Validated Index of Brain ComplexityAbstract, Results

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Observational study: associations, not causes

    The source reads: We found an empirical cutoff that discriminated with 100% sensitivity and specificity between the conscious and the unconscious conditions in the benchmark population. This cutoff resulted in a sensitivity of 94.7% in detecting MCS

What this does not establish

The 2013 study tested 20 brain-injured patients and called itself hypothesis-generating. The 2016 validation used 150 people who could report, then 81 patients who could not; in 9 of 43 vegetative-state patients the index was high, and whether they were conscious is not known.

Later work looked for

Europe PMC and Crossref, 2013 on, "perturbational complexity index": the independent validation in 150 people (Casarotto 2016) supports and extends the 2013 result; page updated to cite it.

Corrected

24 September 2026. Previously: “and compute an index that separates conscious from unconscious patients reliably.” The 2013 result was in 20 patients; the reliability rests on a benchmark of people whose state was known from their reports, and in unresponsive patients the true state cannot be checked. The 2016 validation was added as a source.

Claim 40 · checked 24 September 2026 · AI-assisted audit

Claim: Its cutoff was set on people who could report whether they had been conscious,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

Europe PMC, 2016 on, "perturbational complexity index": later work found includes a 2026 expert review of TMS-EEG, not read; nothing read changes the claim.

Corrected

24 September 2026. Previously: “It works because human brains are similar to one another.” Why the index works, on its authors’ account, is that it measures integrated and differentiated activity. What the passage needs is how it was calibrated, which the 2016 validation reports.

Claim 41 · checked 24 September 2026 · AI-assisted audit

Claim: and it carries over to other people because human brains are similar to one another.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading. The authors took care to calibrate on brain-injured as well as healthy people before applying the cutoff to patients, which is a claim about human brains being comparable, not about brains in general.

Claim 42 · checked 24 September 2026 · AI-assisted audit

Claim: Nothing equivalent exists for a system with a different architecture.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Identifying Indicators of Consciousness in AI SystemsLooking ahead

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Review of the field

    The source reads: Validating our method in the AI case would be challenging because it would require the development of alternative assessment methods suited to AI

Claim 43 · checked 24 September 2026 · AI-assisted audit

Claim: An index calibrated on brains is not evidence about something that is not one.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s judgement. It follows from how the index was calibrated (claim 41) and from the objection the 2023 report discusses: evidence from humans does not say how far its conditions stretch.

Claim 44 · checked 24 September 2026 · AI-assisted audit

The wall

Claim: The wall

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: We grant experience to each other on resemblance, which is weak evidence,

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Other Minds§1.1

    Cited · Link answered, 19 October 2026 · The cited passage was read, 24 September 2026 · Texts · Reference work

    The source reads: Mill who writes that analogy is an “inferior degree of inductive evidence” ... it was considered problematic that this way of extending one’s knowledge proceeds from a single case

Other readings

The entry reports defenders of the argument from analogy, and that the argument from best explanation is now the more widely accepted account of why the belief is justified.

Corrected

24 September 2026. Previously: “We grant experience to each other on resemblance, which is not evidence, and has never had to be.” Analogy is inductive evidence, Mill’s “inferior degree” of it; calling it no evidence took one side of a philosophical dispute as settled.

Claim 46 · checked 24 September 2026 · AI-assisted audit

Claim: and has never had to be more.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading: in ordinary life nobody waits for a better argument before treating other people as conscious.

Claim 47 · checked 24 September 2026 · AI-assisted audit

What would settle it

Claim: What would settle it

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: No proposed observation decisively distinguishes a system that has experience from one that behaves as though it does.

Kind
Open question. Something the page says is not known.
Review
Presented as open. The page says this is unresolved, and the review found that it is.

The indicator method looks at how a system works rather than how it behaves, but its verdicts are only as good as the theories behind it (claim 21), so it does not settle the question either.

  • Context

    Could a Large Language Model be Conscious?section 1

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Theory · Argument

    The source reads: It’s not easy to devise benchmarks for consciousness. But perhaps there could at least be benchmarks for aspects of consciousness

Corrected

24 September 2026. Previously: “No proposed observation distinguishes a system that has experience from one that behaves as though it does.” The page’s own second section describes a method that bears on the difference; what none does is settle it.

Claim 49 · checked 24 September 2026 · AI-assisted audit

Claim: That is the same wall the question about other people runs into, and with other people we get past it by inference from resemblance rather than by any test.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Other Minds§1.1

    Cited · Link answered, 19 October 2026 · The cited passage was read, 24 September 2026 · Texts · Reference work

    The source reads: One form of reasoning has been taken to be “traditional” in this connection: the argument from analogy ... in the case of other minds the conclusion is logically uncheckable (Locke 1968)

Other readings

On the account most widely accepted today, the belief rests on inference to the best explanation of what others do rather than on resemblance; some philosophers deny that it is an inference at all.

Corrected

24 September 2026. Previously: “and we get past it with other people by resemblance rather than by evidence.” An argument from analogy is a kind of evidence; what it lacks is a check, since its conclusion cannot be verified.

Claim 50 · checked 24 September 2026 · AI-assisted audit

Claim: So the practical question may not be whether a machine is conscious but what to do under uncertainty about it, which is a question about conduct rather than about measurement, and is not improved by pretending the measurement exists.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s conclusion. The 2023 report recommends thinking in credences and considering the moral risks; Chalmers raises whether conscious AI should be built at all.

Claim 51 · checked 24 September 2026 · AI-assisted audit

What was read