What to look for
What this gives us: a practice. For each sign, who counts it, and how hard it is to fake.
How hard to fake, on each sign below, is taken from Hardest to Fake. Low means behavior or a label a system can be trained to produce. Medium means an architectural sign that can be bolted on, or trained against a probe. High means a sign that talk and training on text cannot supply. High does not mean the sign has been shown, and it does not settle the question. Where that paper does not score a sign by name, the tag says so.
Checked by our internal skeptic review; corrections welcome.
Assembled by the Sentinel Guild from cleared research briefs.
People reach for a short list of signs when they ask whether a person, an animal, or a machine is conscious. This page is that list, drawn from the markers in the Guild's theory notes. Each sign says what it is, which ideas treat it as important, and where it is weak or easy to fake.
It is a looking-glass. It shows what a claim is relying on. It is not a test that settles the question. The longer map, with the quotes and the studies, is Theories of consciousness: a map, not a verdict. Nothing here says that any AI system is conscious, or that none is.
1. Talk about an inner life
How hard to fake: low. Talk about feelings is behavior, and Hardest to Fake calls behavior the weakest link. Linguistic behavior can be trained.
Someone reports that they saw something, that something hurt, or that they are aware. The notes' clearest machine example is a system that says "I am aware". Researchers call this kind of statement a report. Almost every lab method starts from reports, because the scientist needs some way to sort trials into seen and unseen. Higher-order theories care about a further step: a representation of being in the state, not the words alone. Attention schema theory cares about why a system comes to believe and say that it is aware. Illusionism treats reports as fallible data, not as a window that cannot be wrong.
This is the weakest sign if you are looking at a system trained to talk. Fluent language about feelings is easy to produce and easy to imitate. Birch calls this the gaming problem: if a system or its maker learns which signs are being counted, those signs should be discounted. On the Guild's reading of attention schema theory, a chatbot that says "I am aware" without a model that actually helps control attention is mimicry, not the marker that theory names. Confidence scores have the same problem. Brown, Lau, and LeDoux warn that most higher-order theorists do not equate consciousness with metacognition or with confidence. A report can also be wrong in humans. Illusionism is built on that point, and clinical scans sometimes find task-related activity that bedside behavior missed, which is the opposite mismatch.
2. A broadcast that ignites
How hard to fake: medium. A global workspace can be bolted on as theater, a lobby nobody uses. It is harder to fake if an assessor can switch it off and watch what changes.
Global workspace theory says information becomes conscious when it is amplified, shared across distant brain areas, and held in a limited common workspace. The markers named by Dehaene and Changeux are late amplification of sensory activity, long-distance synchrony, and "ignition": a sudden, all-or-nothing jump into widespread prefrontal and parietal activity.
In machines, Butlin, Long, and colleagues look for a related set of design features: specialist modules in parallel, a limited workspace with selective attention, a global broadcast, and attention that depends on state. Those are about organization, not about the words the system emits.
The largest test, Cogitate 2025, challenged the workspace account on ignition when a stimulus ends, and on how fully certain conscious features show up in prefrontal cortex. Some predictions still lined up. One study does not close the theory. It does mean you should not treat "ignition" as a sign that has already held up in the form supporters predicted. You also cannot see this sign from fluent conversation. And Block's distinction still sits there: a state can be available for use without that availability answering whether anything is felt.
3. Activity that is both unified and varied
How hard to fake: medium. Hardest to Fake does not score this sign by name. The closest warning there is that a check can be hard to game and still track the wrong thing. Exact phi cannot be computed, and simple networks can score high on stand-ins.
Integrated information theory says consciousness is an irreducible pattern of causes and effects in a physical system, measured as phi, and that it comes in degrees. In humans, proponents point to a posterior "hot zone": content held by lasting activity and by lasting short-range synchrony. More broadly, they point to activity that is integrated and also differentiated.
The perturbational complexity index, a magnetic pulse followed by a complexity score on the EEG, was inspired by those ideas. It does not compute phi, it does not assume the theory is true, and its cutoff was set using people who could report. A good score neither proves nor disproves the theory.
The sign is easy to overread. Exact phi, on the software paper the notes cite, is out of reach for a real brain, so labs use stand-ins. Cogitate 2025 found a lack of sustained synchronization in posterior cortex, which cuts against the claim that network connectivity of that sort specifies consciousness. Some sustained posterior activity was there, but sparse. Aaronson showed that simple networks can score high on phi while doing nothing smart. And on the theory's own statement, a digital computer could behave like us, even simulate a brain, and still experience next to nothing. Behavior and this marker are supposed to come apart. That also means a convincing demo is not the sign.
4. A monitor on one's own perception
How hard to fake: low. A confidence number can be trained to look right where it is tested and nowhere else. That is one of the higher-order gaming routes in Hardest to Fake.
Higher-order theories say a state is conscious when it is represented as one's own. The things to look for are metacognitive sensitivity, meaning confidence that tracks accuracy, and prefrontal activity that tracks reported awareness when the task score stays the same. The first is a pointer. The Brown, Lau, and LeDoux line says it is not the thing itself.
The relative-blindsight study found more prefrontal activity when people reported seeing more, at matched accuracy. The authors called the correlation across subjects weak, and direct replications were not checked. A stimulation result, less metacognition after prefrontal disruption, failed one replication attempt and remains disputed. Cogitate's trouble with prefrontal content also presses those higher-order theories that expect prefrontal cortex to carry conscious content.
For machines, the indicator is a mechanism that monitors the system's own perception and separates reliable signal from noise. A printed confidence number, or a sentence about being uncertain, does not show that mechanism. It is the same self-report problem as in the first section, in technical costume.
5. Loops from higher sensory areas back down
How hard to fake: low. The weak, algorithmic form of recurrence is a condition many systems already meet. A datasheet that says "recurrent" is not the biological marker.
Recurrent processing theory says a fast feedforward sweep is not enough for visual experience, and that local feedback among visual areas is enough. Wider loops, up into the networks used for report, add access, not the seeing itself. The markers are feedback onto primary visual cortex, figure-ground structure that depends on it, an EEG signal around 200 milliseconds (the "visual awareness negativity"), and the loss of late feedback signals under masking or magnetic stimulation while the early sweep survives.
A masking study from Lamme's group found that pattern: early activity remained when the target was unseen, and feedback signs did not. An independent replication was not checked. Magnetic-stimulation work supports some role for feedback into primary visual cortex. Related studies are not a direct copy of the original experiment. Lamme himself writes that some levels of recurrence fit with unconsciousness, so something else may be needed. Recurrence can be necessary on many people's view without being sufficient on his.
In software, reusing a computation is not the same thing as physical feedback between brain areas. Butlin and colleagues say existing AI already meets the weaker, algorithmic version of the indicator. Whether Lamme's theory should accept that weaker reading is unresolved. His published remark that we may already have created conscious machines is conditional on recurrence, or on information integration, being the right explanation. A model called "recurrent" on a datasheet is not the biological marker.
6. Guessing the world, then correcting the guess
How hard to fake: low. Next-token prediction can be offered as if it satisfied the sign. An indicator so broad it cannot be failed is its own kind of gaming.
Predictive processing says the brain keeps a hierarchy of guesses, sends prediction errors upward, sends predictions downward, and uses attention as a volume knob on those errors. Active inference adds action: change the world so the guess comes true. Seth and Hohwy call this a promising theory for the science, not yet a single theory of consciousness.
The markers are real features of a research program. On the Guild's reading they do not, by themselves, show felt experience. The mismatch response to an oddball sound is about sensory learning. Binocular-rivalry models in this family were fitted to known flips of a percept. They were not a preregistered test of a consciousness theory. With enough freedom to set prior beliefs and precision, many outcomes can be retold as error minimization. A system that minimizes prediction error, including a vision model that already meets the indicator Butlin and colleagues name, can be doing prediction and still leave the consciousness question untouched. "Predictive" on a label is easy to apply and hard to falsify.
7. A model of attention, used to steer it
How hard to fake: low. A sentence that says "I am aware," without a model that helps steer attention, is mimicry. Hardest to Fake treats that kind of report as easy to produce.
Attention schema theory says the brain builds a simplified model of its own attention, and that the content of that model is what we report as subjective awareness. The markers are attention that still happens when the cue is unseen but is less stable or less internally guided, overlapping brain networks for monitoring one's own attention and other people's, and social use of the same model to attribute awareness to others.
The cue studies that match the control prediction come from Graziano's lab. Independent replications were not checked. The imaging overlap supports the social strand. It does not prove that the schema is the feeling. On the Guild's reading, the theory is strongest as an account of reports and of why awareness seems mysterious. Felt experience stays underdetermined. Graziano's own engineering claim is that a machine with a rich model of consciousness could believe it is conscious and act as if it is, in the same sense that, on his view, we do. That is a claim about belief and action. A sentence of self-description, without the control model, is not the marker.
8. A body that acts and senses in a world
How hard to fake: medium. Hardest to Fake treats behavior as the easiest sign to game and does not score a lived body apart from that. A robot that moves is not, by that fact, the marker these views name.
Embodied and enactive views look for sensorimotor coupling the agent has mastered, skillful engagement with the world, a lived body that is part of the experience rather than a container, and, in the enactive strand, a self-producing organism engaged in sense-making. The extended-mind strand looks for tools that play a cognitive role. Extending thinking into a notebook is not, on these notes, the same as extending felt experience.
Change blindness is a robust finding. The enactive reading of it, that seeing is a way of acting rather than an inner picture, is disputed. Sensory substitution and robots that learn by moving are real phenomena and real models. They are not a replicated test that someone is conscious. The view can stretch to absorb counterexamples by widening what counts as the system. On the Guild's reading, without tight criteria almost any success can be redescribed as enaction.
A language model bolted into a robot is not automatically a lived body. Butlin and colleagues' agency indicators, learning from feedback and modeling how outputs change inputs, are closer for some robots and some learning agents than for text-only chatbots. Closer on a checklist is not a finding of felt experience.
9. The right biological cause, not just the right program
How hard to fake: high. Conversation cannot supply the biological powers this view asks for. It is easy to fake only if talk is treated as the test, which is the mistake the view is aimed at.
Biological naturalism, Searle's view, does not offer a single brain signature. It tells you to look for biological causal powers of the right kind, and for a subjective, first-person side that is not defined by input-output role. The Chinese Room is the intuition he uses: fluent symbol-shuffling is syntax, and syntax is not understanding. It is an argument, not a study with a replication record. The line that some physical machine could in principle understand, because our brains are machines, sits next to a denial that a machine defined only as a program would suffice. The notes paraphrase that denial. They did not re-check the excerpt against the journal PDF.
The sign is vague. "The right causal powers", without a specified mechanism, can wait on the next neuroscience paper. It is also easy to fake if you treat conversation as the test, which is the mistake the view is aimed at. It does not say that a future system with those powers is impossible. It does say that today's digital systems are not conscious merely because the software is clever or the answers fit. On the Guild's reading, that still does not hand experience to a neuromorphic chip or to a dish of neurons. The powers have to be shown, and they have not been.
10. Quantum events in the cell's tubular scaffolding
How hard to fake: high. The marker is a quantum event in neuronal microtubules. Classical digital hardware cannot produce it by talking. The red flags for overclaiming are on the quantum brief.
Orch-OR says moments of consciousness are orchestrated collapses of quantum states in neuronal microtubules. The markers are coherence that lasts long enough, timing that lines up with neural and behavioral scales, and anesthetic effects on those tubes or on their optical signatures.
What has actually been measured is thinner than that list. Tegmark argued that quantum coherence in the relevant degrees of freedom dies far too fast for neural timescales. Hagan, Hameroff, and Tuszynski argued for a longer window. That dispute is open. An underground experiment rules out the natural parameter-free version of the Diósi–Penrose collapse model, not every variant of the theory. Energy spread along microtubules, light emission in tryptophan networks, and a drug that delayed rats losing their righting reflex do not show orchestrated collapse and do not show a moment of experience. Hameroff and Penrose are co-authors on the energy study, which is a reason to read the claim narrowly, not a reason to ignore the measurement.
If this theory were required, classical digital computers would be out on hardware grounds no matter what they said. That "if" is not established. Quantum hardware would remain a separate, open case. The conditional is not a moral conclusion.
11. Animal-style signs of damage, learning, and trade-offs
How hard to fake: low. Flexible behavior and verbal report can be trained. Hardest to Fake is the home for that gaming warning.
The 2021 review by Birch and colleagues scores eight criteria: damage-detecting nerves, integrative brain regions, connections between them, responses that painkillers change, trade-offs between threat and reward, flexible self-protection, associative learning beyond habituation, and valuing a painkiller when injured. The Cambridge Declaration stresses neurological substrates of affective states and says a neocortex is not required. The New York Declaration uses two tiers: strong scientific support for other mammals and birds, and at least a realistic possibility for all vertebrates and for many invertebrates, including cephalopods, decapod crustaceans, and insects.
Inside those frames, the stronger packages are mammals, birds, and octopuses under the eight criteria. What stays debated is the step from behavior to feeling, insect and crustacean sentience beyond a realistic possibility, and single studies that have not been repeated. Low confidence on a criterion, the 2021 review says, does not mean sentience is unlikely or disproven. On the Guild's reading, the declarations are judgments under uncertainty, not lab proofs of phenomenality.
These signs travel badly to digital AI. There is no nociceptor to point to, and missing one does not, by itself, answer a theory that was never about nociceptors. Flexible behavior and verbal report can be trained. Birch's gaming problem applies directly: discount markers you have reason to think have been gamed. The animal case can justify caution about animals without granting sentience to a machine, and without denying it.
12. A brain that answers when the body does not
How hard to fake: medium. Looking inside can be fooled by bad statistics, as the hidden-awareness brief says, and Hardest to Fake warns that the next place you look can be gamed too.
In disorders of consciousness, clinicians look at bedside behavior with the Coma Recovery Scale–Revised, at willful imagery in a scanner (tennis versus walking through a house), at EEG versions of command-following, and sometimes at the perturbational complexity index. Owen and colleagues found a patient, diagnosed as vegetative, whose imagery activations matched healthy volunteers. Monti and colleagues found a patient who could answer yes or no in the scanner and still could not do so at the bedside. Bodien and colleagues, in a large cohort, found task-related activity in patients with no observable response to commands.
The frank limits matter as much as the findings. Many patients who can follow commands still fail the brain tests, so a negative scan is not proof of absence. Critics argue that an activation can be an automatic response to words rather than a choice. The methods compare an injured human brain with healthy human volunteers. An AI system has no such anchor. Behavioral clues that help with patients are, on the Guild's reading, easy for a language model to learn from training text. "Looking inside" with a theory's checklist is a different project, and it inherits that theory's assumptions.
Marks from traditions outside the lab are not on this list. That section says itself that these are not tests. They are on the map, under contemplative and non-Western accounts.
A looking-glass, not a test
You can walk the whole list and still not know. The signs point at different things: a broadcast, a wiring pattern, a monitor, a sensory loop, a guess, a model of attention, a lived body, a biological power, a quantum event, an animal's trade-off, a hidden response in an injured brain. Several of the clearest signs are disputed. The clearest behavioral ones, especially fluent talk about experience, are easy to fake.
Illusionism and panpsychism appear in the longer map. They do not yield a positive "look for" marker of the same kind as the entries above: one rephrases the question about reports, and the other turns on combination or intrinsic natures rather than a lab sign you can tick.
Use the list to see what a claim is using as its evidence, and what that evidence cannot show. This is a looking-glass, not a test that settles the question.