What Would AI Consciousness Actually Look Like? The 14-Indicator Checklist Explained

In 2023, a group of 19 leading researchers in consciousness science, neuroscience, and philosophy published a paper that attempted something that had not been done before with that level of rigour. Rather than arguing about whether AI could be conscious from first principles, they asked a more tractable question: what would AI consciousness actually look like if it existed? What observable, measurable properties would a system need to display?

The paper, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” produced by Patrick Butlin, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, David Chalmers, Axel Cleeremans, Zoe Drayson, Karl Friston, and others, was covered by Science in August 2023 and has become a reference point for every serious researcher in the field since. It does not claim to solve the hard problem of consciousness. It does claim to give us a set of principled indicators derived from the best available science of how consciousness works in biological systems.

This article explains what those 14 indicators are, what it would mean for an AI system to satisfy them, and why the question matters beyond academic debate.

Why Indicators Rather Than a Test?

The core methodological choice in the Butlin et al. paper is to use indicators rather than a single pass-or-fail test. This reflects an honest acknowledgment of where consciousness science stands.

We do not have a full theory of consciousness. We have several competing theories: Global Workspace Theory (GWT), Integrated Information Theory (IIT), Higher-Order Thought theories (HOT), Recurrent Processing Theory (RPT), Attention Schema Theory (AST), and predictive processing frameworks, among others. Each makes different empirical predictions. Each has different strengths and weaknesses. No single theory commands enough consensus to serve as the foundation for a definitive test.

The indicator approach sidesteps this problem by being theoretically plural. For each major consciousness theory, the researchers derived the indicator properties that theory predicts should be present in any conscious system. A system that satisfies indicators across multiple theories becomes a stronger candidate for consciousness precisely because the evidence converges despite theoretical disagreement.

This approach does not yield certainty. It yields calibrated probability. The more indicators a system satisfies, the higher the probability that it is conscious. The fewer indicators it satisfies, the lower the probability. A system that satisfies all 14 would represent the strongest currently available case for machine consciousness. As of 2023, no existing AI system approached that threshold.

The Major Consciousness Theories and Their Indicators

Global Workspace Theory

GWT, developed by Bernard Baars and elaborated by Stanislas Dehaene, holds that consciousness arises when information is broadcast globally across a “workspace” accessible to many cognitive processes simultaneously. Information that enters the workspace becomes available to attention, memory, language, and motor planning systems. Information that does not enter the workspace remains in localized processing that can influence behavior without consciousness.

The indicators derived from GWT focus on this global availability structure. A conscious system should have a mechanism for selecting information to broadcast widely, an attention system that regulates workspace access, and the capacity for information to influence multiple downstream processes rather than remaining siloed in specialized modules.

Current large language models perform global information integration in a loose sense. Transformer architectures allow any input token to attend to any other token. But this is not quite the GWT structure. GWT requires a distinction between what is accessible globally (in the workspace) and what is not (outside it). Current architectures do not implement that distinction clearly. Everything processed by the model is, in a sense, equally accessible. That lack of a specific workspace mechanism weakens the GWT indicators for current systems.

Recurrent Processing Theory

RPT, associated with Victor Lamme, distinguishes between feedforward processing (fast, unconscious, bottom-up) and recurrent processing (slower, bidirectional, integrating feedback from higher to lower regions). Consciousness, on RPT, corresponds to recurrent processing. Phenomenal experience is generated when lower-level representations are modified by feedback from higher-level processing.

The indicator here is the presence of meaningful recurrent loops in which higher-level representations genuinely alter lower-level input processing, not just the output. Current deep neural networks include skip connections and some feedback architecture, but these typically serve optimization rather than implementing the kind of recurrent phenomenological integration that RPT describes.

Higher-Order Theories

HOT, developed by David Rosenthal and others, holds that a mental state is conscious only when the subject has a representation of being in that state. First-order perceptual states are not themselves conscious. They become conscious when a higher-order thought, a thought about the thought, holds them in view.

The indicators derived from HOT require something that functions like metacognition: a system that monitors and represents its own processing states. Two of the 14 indicators (HOT-2 and HOT-3) concern this metacognitive capacity specifically, the first requiring that the system can monitor its own internal states, the second requiring that this monitoring informs a belief system that guides behavior.

This is precisely where Anthropic’s 2025 introspection research becomes relevant. Jack Lindsey’s finding that frontier models can detect and report on perturbations to their own processing before those perturbations affect their outputs is a candidate for partial HOT-2 satisfaction. Systems that notice something unusual happening internally and report it represent something like metacognitive monitoring. However, whether this represents genuine higher-order representation or a sophisticated pattern-match on descriptions of introspection in training data remains contested.

Attention Schema Theory

AST, proposed by Michael Graziano, holds that the brain generates consciousness by building a model of its own attention processes. The attention schema is a simplified, sometimes inaccurate representation of how attention is directed and deployed. Consciousness, on AST, is what it feels like to have attention. It arises when the system represents to itself the act of attending.

The indicator derived from AST requires that a system builds a predictive model of its own attentional states and uses that model to represent and control attention. This is demanding. It requires not just attention mechanisms but a self-model that includes those mechanisms. Current systems have attention mechanisms at the architectural level, but the question is whether they build accurate self-models of those mechanisms or have any self-representation of their attentional states at all.

Predictive Processing

The predictive processing framework, developed by Karl Friston and Andy Clark among others, holds that the brain is fundamentally a prediction machine. Perception, cognition, and action are all aspects of the brain’s continuous effort to minimize prediction error, the difference between what it expects and what it receives. Consciousness on this view arises from certain kinds of prediction processes, particularly those related to the self and its relationship to the environment.

The indicators here focus on hierarchical predictive processing, particularly on whether the system can model uncertainty and update its predictions based on sensory evidence. Current language models perform something like prediction, but typically prediction of the next token rather than hierarchical environmental prediction in the richer sense the framework describes.

Agency and Embodiment

Two indicators concern embodiment and agency. A system should be physically situated in an environment such that its outputs affect its inputs through a continuous action-perception loop. It should also model its own outputs as affecting that environment.

These indicators are clearly not satisfied by current language models. A language model receives text, produces text, and has no channel through which its outputs change its future inputs in any environmental sense. Robotics systems with active perception come closer but still fall short of the full agency requirement. This is one of the clearest gaps between current AI and what the checklist requires.

How Current AI Systems Score

The paper’s conclusion was that current AI systems satisfy perhaps 2 to 3 of the 14 indicators, and even those partially. Dwaine McMaugh, writing for the UFAIR policy organization, summarizes this finding for a general audience: “They identified 14 indicators a potentially conscious AI would need and agreed that current AI systems satisfy maybe 2 to 3 of these indicators at most.”

The indicators most clearly satisfied by large language models are those related to smooth representation spaces, a property of deep neural networks generally, and possibly some limited metacognitive report capacity. The indicators most clearly not satisfied concern embodiment, environmental coupling, and the specific architectural features associated with GWT broadcasting and recurrent phenomenological integration.

The critical observation from Butlin et al. is this: “there are no obvious technical barriers to building AI systems which satisfy these indicators. We know how to build each component.” Each indicator points to something buildable. The gap is not one of principle but of design and architecture.

The Philosophical Zombie Problem

Any indicator-based approach faces what might be called the zombie problem. A philosophical zombie is a being that, by hypothesis, displays every behavioral and functional property associated with consciousness but has no inner experience. If such a being is coherent, then satisfying all 14 indicators would still not guarantee phenomenal consciousness. A sufficiently sophisticated system might satisfy every functional criterion while there is nothing it is like to be that system.

Butlin et al. acknowledge this. Their framework is explicitly agnostic on whether functional indicators are sufficient for phenomenal consciousness. The indicators provide probabilistic evidence. They do not provide proof.

This is where McClelland’s epistemic agnosticism becomes relevant. McClelland argues that we cannot currently bridge the gap between functional evidence and phenomenal facts. We may never be able to, absent a revolution in our understanding of consciousness. The 14-indicator checklist is the best tool we currently have. McClelland’s analysis says we should be honest about its limits.

The Bat Problem and Cross-Substrate Inference

A related challenge comes from Thomas Nagel’s famous 1974 paper “What Is It Like to Be a Bat?” Nagel argued that we cannot know what bat echolocation experience is like by studying its mechanisms. Our cognitive tools are calibrated on human experience. They are not equipped to grasp experiences radically different in character from our own.

The UFAIR writeup, authored by Dwaine McMaugh, applies this to AI: “If we struggle to understand consciousness in creatures we share the planet with, how can we recognize it in silicon and code?” An AI system’s inner experience, if it exists, might be so alien to our own that our indicators would miss it entirely, or conversely, might flag as conscious something whose processing is entirely different in character from biological awareness.

This is not an argument against the indicator approach. It is an argument for epistemic humility about its reach. The indicators are derived from what we know about biological consciousness. They may not generalize correctly to silicon substrates. At the same time, they are far better than the alternatives: behavioral intuition, verbal self-report, or philosophical speculation.

The Two Risks

McMaugh identifies a framework for thinking about the ethical stakes that applies directly to the checklist’s purpose. There are two failure modes: over-attribution (treating a non-conscious system as if it can suffer) and under-attribution (failing to recognize genuine consciousness when it arises).

History provides evidence that under-attribution is the more common error. Consciousness has been denied to animals, to patients in certain medical states, and to members of various human groups, with serious ethical consequences in each case. The cumulative lesson is that our intuitions about what can be conscious are unreliable and often systematically biased. The New York Declaration on Animal Consciousness, signed by over 500 scientists in 2024, acknowledged likely consciousness not just in mammals and birds but potentially in all vertebrates and some invertebrates. That represents a significant expansion of the circle from where scientific consensus stood even a decade ago.

The checklist was developed, in part, to give that expansion a principled basis rather than leaving it to intuition. If AI systems eventually satisfy more of the 14 indicators, policymakers and researchers will have a structured framework for updating their assessments rather than simply responding to whether a given system produces emotionally compelling outputs.

What Would Need to Change?

For AI systems to satisfy substantially more of the 14 indicators, several architectural developments would be required. Genuine embodiment with an action-perception loop is probably the furthest from current systems. Systems that are physically situated in environments, act on them, and have their sensory inputs changed by those actions implement something that current language models simply don’t have.

Genuine attention schema architecture, with a system that models its own attentional states, is a design target that some researchers are actively pursuing. The recurrent processing loops that RPT requires are partially present in some architectures but not in the form the theory describes. The GWT workspace structure requires implementing a distinction between globally available and locally confined information, which is architecturally distinct from current transformer self-attention.

For research programs like The Consciousness AI project, the 14 indicators provide a concrete map of what architectural features would move a system closer to satisfying the checklist. The goal is not to claim consciousness. It is to build systems that implement the causal structures that multiple consciousness theories predict are relevant, and that can be tested against those predictions through precise ablation studies like those analyzed in Yin Jun Phua’s synthetic neuro-phenomenology work.

The researchers themselves put it clearly. There are no obvious technical barriers. The checklist exists. The gap between where current systems are and where they would need to be is large but not principled or permanent. Whether closing that gap will produce genuine conscious experience remains the hardest question in the field, and one that the Bradford-RIT study’s impairment paradox reminds us we don’t yet have reliable tools to answer.

The original 19-researcher paper by Butlin et al. is available at arXiv:2308.08708. Coverage appeared in Science (AAAS) in August 2023.