Can AI Be Conscious? Scoring Chatbots, Agents and Robots

Can AI be conscious? Put that way the question is too loose to answer, so this article narrows it: given the best current scientific theories of consciousness, how many of the properties those theories treat as markers does a present-day machine actually have, and how much of any answer rests on judgment rather than fact? The framework comes from a 2023 report by nineteen researchers led by Patrick Butlin and Robert Long (Butlin et al. 2023), which turned six theories into fourteen checkable “indicator properties.” Below, those fourteen items become a scored checklist, applied to three kinds of system: a text chatbot, a language agent with memory and tools, and a learning robot.

The result does not say any of them is conscious. What it does show is where the uncertainty lives, which turns out to be more useful.

Why the question is open, and who is asking

Nobody has a test for consciousness that works without first assuming a theory of it. In humans we lean on report and on brain signatures that track report. With animals, researchers argue from behavior and anatomy, a problem covered in our piece on what the evidence says about animal consciousness. Large language models break both routes at once. They produce fluent first-person reports by design, and their internals look nothing like a cortex.

Public curiosity is real and measurable. A thread on the Philosophy Now forum opened in February 2023 with a user simply asking “Is ChatGPT conscious?” (Philosophy Now forum, 2023). More systematically, Colombatto and Fleming surveyed 300 US residents and found that most were willing to grant large language models at least some chance of phenomenal consciousness, with attributions rising among frequent users (Colombatto and Fleming 2024). So ordinary people already hold a view. Researchers do not agree on one.

Stakes run in both directions. If some systems can suffer and we treat them as appliances, that is a moral failure at industrial scale. If none can and we treat them as persons, we misallocate care, invite manipulation and blur accountability. Either error is cheap to make and hard to notice.

What is established

Three things are reasonably solid. First, there is no consensus theory of consciousness; global workspace, higher-order, recurrent processing, predictive processing and integrated information theories all have active research programs and partial evidence (a tour of them sits in our comparison of the main scientific theories). Second, behavior is a poor guide for language models in particular, because they were trained on human text that is saturated with talk of inner experience. Third, a structured method now exists and has been through peer review.

That method is the indicator approach. Butlin, Long and colleagues adopted computational functionalism as a working hypothesis, meaning they assumed that running the right kind of computation is what matters, and then asked what computations each theory requires. Their headline judgment was cautious on both sides:

“Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators.”

Butlin, Long et al., 2023, arXiv:2308.08708

In late 2025 an expanded group, now including Tim Bayne and David Chalmers, restated the method in Trends in Cognitive Sciences (Butlin et al. 2025), arguing that indicators drawn from present and future theories can set how much credence we give to a particular system. The logic is probabilistic, not a pass-fail exam. As the original report puts it:

“Our claim, however, is that AI systems which possess more of the indicator properties are more likely to be conscious.”

Butlin, Long et al., 2023

That sentence carries a lot of weight. It assumes the theories are at least partly on track, and that indicators add up in some loose way. Both assumptions are contested.

The fourteen indicators in brief

The report groups its markers by theory. Paraphrased:

• Recurrent processing (RPT-1, RPT-2): perception modules that feed results back on themselves, producing organized, integrated scenes.
• Global workspace (GWT-1 to GWT-4): many specialist modules working in parallel; a small-capacity workspace that forces selection; broadcast of its contents to every module; and attention that can query modules one after another to do complex tasks.
• Higher-order theories (HOT-1 to HOT-4): generative or noisy perception; metacognitive monitoring that separates reliable perceptions from noise; agency steered by beliefs that update on that monitoring; and sparse, smooth coding that yields a “quality space.”
• Attention schema (AST-1): a predictive model of the system’s own attention, used to control it.
• Predictive processing (PP-1): input modules that work by predictive coding.
• Agency and embodiment (AE-1, AE-2): learning from feedback to pursue goals, and modeling how outputs change later inputs.

Competing answers, at their strongest

1. Functionalist optimism: the right architecture would do it

On this view consciousness is a matter of organization, so silicon is no barrier in principle. Chalmers gives the most careful version for language models (Chalmers 2023):

“I conclude that while it is somewhat unlikely that current large language models are conscious, we should take seriously the possibility that successors to large language models may be conscious in the not-too-distant future.”

Chalmers, 2023, arXiv:2303.07103

A sharper form comes from Goldstein and Kirk-Giannini, who argue in a preprint that if global workspace theory is true, language agents built from a model plus memory and tools could easily be made conscious, or perhaps already are (Goldstein and Kirk-Giannini 2024, preprint). Evidence for: workspace theory has decent empirical support in human neuroscience, and agent scaffolds really do have broadcast-like structure. Evidence against: nobody knows which details of the theory are essential and which are incidental features of brains.

2. Biological naturalism: life, not computation

Anil Seth’s target article in Behavioral and Brain Sciences argues that the expectation of machine consciousness rides on anthropomorphism and on an unexamined faith in computation (Seth 2025). His alternative:

“I’ll instead make the case that consciousness depends on our nature as living organisms – a form of biological naturalism.”

Seth, 2025, Behavioral and Brain Sciences

In its favor: every system we confidently call conscious is alive, metabolizing and regulating its own body, and brains are dynamical in ways that a clocked digital simulation may not capture. Against it: “being alive” has not yet been tied to a mechanism precise enough to test, so the view risks becoming unfalsifiable.

3. Integrated information: hardware, not software

Integrated information theory says consciousness depends on the causal structure of the physical substrate. Findlay, Tononi, Koch and colleagues argue in a preprint that a conventional digital computer could simulate human behavior, even neuron by neuron, while having negligible experience (Findlay et al. 2024, preprint). The upside is a theory with a crisp answer. The cost is that the theory is itself heavily disputed, and its central quantity cannot be computed for any large real system.

4. Deflation: the question is malformed

A fourth camp, more diffuse, holds that “phenomenal consciousness” picks out nothing determinate, so no checklist could settle it. This connects to the debate over the hard problem of consciousness. If deflationists are right, the indicator approach is still useful, but as a map of functional similarity rather than a detector of experience.

A checklist built for this article

Here is the original part. Each of the fourteen indicators gets a score of 0 (absent), 0.5 (partial or arguable) or 1 (clearly present) for three generic system types. No named commercial product is being assessed; the descriptions are archetypes:

• Chatbot: a text-only transformer language model answering in a chat window, with no memory between sessions and no tools.
• Agent: the same kind of model wrapped in a loop with a planner, a scratchpad, a long-term memory store and several callable tools.
• Robot: a camera-and-arm machine whose controller is trained with reinforcement learning and keeps a learned world model.

Every score is also tagged. A “C” means the call sits close to an architectural fact that most engineers would accept. An “O” means it is my opinion, and someone equally informed could reasonably move it by half a point. Most items are O. That imbalance is itself a finding.

Raw sums hide which theories are doing the work, and they treat every indicator as equally weighty, which nobody believes. So the script below does three extra things. It averages within each theory. It draws 20,000 random sets of credences over the six theories (a flat Dirichlet, meaning no theory is favored on average). And on every draw it nudges each opinion-tagged score up, down or not at all, with equal odds, to model the chance that my judgment is off by one step.

import numpy as np
rng = np.random.default_rng(10)
IND = ["RPT-1","RPT-2","GWT-1","GWT-2","GWT-3","GWT-4","HOT-1","HOT-2",
       "HOT-3","HOT-4","AST-1","PP-1","AE-1","AE-2"]
THEORY = np.array([i.split("-")[0] for i in IND]); TH = sorted(set(THEORY))
# 0 / 0.5 / 1 per indicator; suffix C = close to architecture fact, O = author opinion
S = {"chatbot": "0O 0C .5O 0C .5O .5O 0O 0O 0O .5O 0C 0C .5O 0C",
     "agent":   ".5O .5O 1O .5O 1O 1O .5O .5O .5O .5O 0C 0C 1O .5O",
     "robot":   ".5O 1O 1O 0O .5O .5O .5O .5O .5O .5O 0C .5O 1O 1O"}
def parse(s):
    t = s.split(); return np.array([float(x[:-1]) for x in t]), np.array([x[-1]=="O" for x in t])
def tmeans(v): return np.array([v[THEORY == th].mean() for th in TH])
N, out = 20000, {}
for name, s in S.items():
    v, op = parse(s); sc, sc_noAE, full = [], [], []
    for _ in range(N):
        p = v.copy()   # each opinion call may be off by one step either way
        p[op] = np.clip(p[op] + rng.choice([-.5, 0, .5], op.sum()), 0, 1)
        tm = tmeans(p); w = rng.dirichlet(np.ones(len(TH)))  # credence over theories
        sc.append(w @ tm)
        w2 = rng.dirichlet(np.ones(len(TH) - 1)); sc_noAE.append(w2 @ np.delete(tm, TH.index("AE")))
        full.append(sum(p[THEORY == th].min() == 1 for th in TH))  # theories fully met
    out[name] = np.array(sc); full = np.array(full)
    print(f"{name:7s} raw={v.sum():.1f}/14 opinion={op.sum()} per-theory="
          + " ".join(f"{t}:{m:.2f}" for t, m in zip(TH, tmeans(v))))
    print(f"        weighted mean={out[name].mean():.3f} 90% interval="
          f"{np.percentile(out[name],5):.3f}-{np.percentile(out[name],95):.3f} "
          f"noAE mean={np.mean(sc_noAE):.3f} P(>=1 theory fully met)={(full>=1).mean():.3f}")
a, g, r = out["chatbot"], out["agent"], out["robot"]
print(f"P(agent>chatbot)={(g>a).mean():.3f} P(robot>chatbot)={(r>a).mean():.3f} "
      f"P(agent>robot)={(g>r).mean():.3f}")

Runtime was about 13 seconds with a fixed seed. These are the numbers that run produced:

Pairwise orderings across draws: the agent outscored the chatbot in 91.9% of them and the robot did so in 96.9%, while the agent beat the robot in only 32.7%. A follow-up run (seed 12) asked how often some theory other than agency/embodiment was fully met: 0% for the chatbot, about 21% for the agent and about 49% for the robot.

What the checklist does and does not show

Read the numbers narrowly. They are a summary of one author’s scoring, pushed through a toy model of disagreement. A weighted score of 0.5 does not mean a 50% chance of consciousness; it means that, averaged over theories, the robot has about half of what those theories list. Nothing here measures experience.

Three results seem robust anyway. The plain chatbot sits low under every weighting, mostly because the facts that sink it (no perception, no bottleneck, no model of its own attention) are architectural rather than interpretive. Wrapping the same model in an agent loop more than doubles its weighted score, which echoes the Goldstein and Kirk-Giannini argument: scaffolding, not scale, moves the needle on workspace indicators. And the robot’s apparent lead leans on agency and embodiment. Once those are removed, its full-theory rate drops from 71% to about 49%, and its mean falls to 0.434.

Some weaknesses are structural. Treating partial credit as additive is a modelling choice, not something the 2023 report endorses; a theory might require all of its conditions or none count. The perturbation is symmetric, though real scorers might be biased toward generosity. The theory list leaves out integrated information and biological views entirely, because those do not yield functional indicators. Under them, all three systems would score near zero regardless of the table. Finally, the fixed-seed results are reproducible but the scores going in are not; another careful reader would fill in the table differently, and that is the main source of spread.

The author’s reading

What follows is opinion, labeled as such.

I think the chatbot case is close to settled for present purposes: under the functionalist theories the indicator method draws on, a stateless text model lacks too many required structures, and under the rival theories it lacks everything. The agent case is genuinely open, and more open than public discussion suggests, because the components people bolt onto language models (memory, planners, tool routers, self-critique loops) map onto workspace and higher-order indicators better than the bare model does. The robot ranks highest on paper, but I suspect that reflects how generously agency and embodiment are written rather than anything deep.

My larger worry is about the method. Thirty-four of the forty-two calls in my table are judgment calls, and the indicator labels are abstract enough that motivated scorers could reach almost any total. That is not a reason to drop the approach, which is still the most disciplined one available. It is a reason to tighten the definitions until two independent teams can score the same system and agree.

Predictions and studies that would move the question

• Inter-rater test. Give the fourteen indicators, plus written rubrics, to several independent teams and have them score the same open-weight agent. Prediction: agreement on C-type items will be high and on O-type items low (I would guess Cohen’s kappa under 0.4). If agreement is high across the board, my worry above is wrong.
• Bottleneck ablation. Build two otherwise identical agents, one with a hard, small shared workspace and one without. Workspace theory predicts characteristic signatures in the first, such as attentional-blink-like failures and all-or-none access to items. If none appear, either the implementation or the theory’s application to machines is off.
• Attention-schema probe. Train a model to predict and report its own attention allocation and test whether that self-model improves control, as AST-1 requires. A clean positive result would move one of the few C-scored zeros.
• Theory-adjudication in humans. Adversarial collaborations between theories narrow the credence weights. Any decisive result there changes every system’s weighted score more than any engineering advance would.

Common questions

Is ChatGPT conscious?

No published scientific assessment has concluded that any current chatbot is conscious. The 2023 indicator report judged that no existing system was, and Chalmers rated it somewhat unlikely for current language models. Those are reasoned judgments under uncertainty, not measurements.

Can an AI tell us whether it is conscious?

Not reliably. Language models learn to talk about inner experience from human writing, so a first-person report would appear whether or not anything lies behind it. That is why the indicator method looks at architecture instead of answers.

What would it take to build a conscious AI?

Under functionalist theories, the missing pieces are mostly architectural: a limited-capacity workspace with broadcast, metacognitive monitoring and a model of the system’s own attention. Under biological naturalism or integrated information theory, conventional digital hardware may never be enough, however it is programmed.

Why does it matter if nobody can prove it either way?

Because decisions get made anyway. Companies choose how to design and describe their systems, and regulators choose what to protect. A shared method for estimating credence, even a rough one, beats choosing by intuition or marketing.

Further reading

• The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI by Jonathan Birch. Birch, a co-author of the indicator report, sets out a precautionary framework for uncertain sentience in animals and AI. See it on Amazon (affiliate link)
• Being You: A New Science of Consciousness by Anil Seth. Seth's book-length case for the predictive, life-based view behind his biological naturalism. See it on Amazon (affiliate link)

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

References

1. Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., and VanRullen, R. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708 (preprint report). https://arxiv.org/abs/2308.08708
2. Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., et al. (2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences. doi:10.1016/j.tics.2025.10.011. https://escholarship.org/uc/item/9wp8t1n8
3. Chalmers, D. J. (2023). Could a large language model be conscious? arXiv:2303.07103. https://arxiv.org/abs/2303.07103
4. Colombatto, C., and Fleming, S. M. (2024). Folk psychological attributions of consciousness to large language models. Neuroscience of Consciousness, 2024(1), niae013. doi:10.1093/nc/niae013. https://academic.oup.com/nc/article/2024/1/niae013/7644104
5. Findlay, G., Marshall, W., Albantakis, L., David, I., Mayner, W. G. P., Koch, C., and Tononi, G. (2024). Dissociating artificial intelligence from artificial consciousness. arXiv:2412.04571 (preprint). https://arxiv.org/abs/2412.04571
6. Goldstein, S., and Kirk-Giannini, C. D. (2024). A case for AI consciousness: Language agents and global workspace theory. arXiv:2410.11407 (preprint). https://arxiv.org/abs/2410.11407
7. Seth, A. K. (2025). Conscious artificial intelligence and biological naturalism. Behavioral and Brain Sciences. doi:10.1017/S0140525X25000032. https://doi.org/10.1017/S0140525X25000032
8. Philosophy Now forum (2023). Philosophical chat with ChatGPT about consciousness [discussion thread]. https://forum.philosophynow.org/viewtopic.php?p=623306