Chapter 2 of 7

What Consciousness Might Be

The hard problem, then six theories in plain language — from most machine-friendly to least — and what each one is actually claiming.

Starting with the hard problem

In 1995 David Chalmers drew a line that organizes everything after it. Some questions about consciousness are, he said, easy — not simple, but tractable. How does the brain integrate information? How does it focus attention, report its own states, control behavior, distinguish waking from sleep? These are hard engineering and neuroscience problems, and we are steadily solving them. They are “easy” in one specific sense: we know what an answer would look like. It would be a mechanism.

Then there is the hard problem: why is any of that accompanied by experience at all? Why isn’t the brain’s information processing simply happening in the dark? A complete mechanical account of vision — retina to cortex to verbal report — appears to leave untouched the question of why there is redness in the seeing.

Thomas Nagel gave the concept its permanent form twenty years earlier. A bat navigates by sonar. We can describe its neurology exhaustively. But we cannot say what it is like to perceive the world by echolocation — and that “what it’s like” is the thing that makes the bat conscious. If there is something it is like to be a system, that system is conscious. If there isn’t, it isn’t.

Six live theories

None of these is proven. Each has serious researchers, serious evidence, and serious critics. They are ordered roughly from most machine-friendly to least.

1. Global Workspace Theory — consciousness as broadcast

The claim. The mind runs many specialized processes in parallel — vision, hearing, memory, language — mostly unconsciously. Consciousness is what happens when information wins a competition for access to a limited-capacity “global workspace” and is broadcast to all the other systems at once. Conscious content is the information currently available to everything.

Why people believe it. It explains a striking pattern: the bottleneck. You can process enormous amounts in parallel unconsciously but only attend to one thing consciously. It predicts specific brain signatures of conscious access, and those signatures show up in experiments — a late, widespread “ignition” of frontoparietal activity when a stimulus becomes reportable.

Key names. Bernard Baars originated it; Stanislas Dehaene developed the neuronal version. Dehaene, with Hakwan Lau and Sid Kouider, split it into C1 (global availability) and C2 (self-monitoring) in their 2017 Science paper.

Verdict on machines. Friendly. Dehaene and colleagues explicitly conclude that both C1 and C2 are computable, and a machine implementing them “would behave as though it were conscious.” VanRullen and Kanai have proposed a concrete deep-learning architecture for it.

2. Higher-Order Theories — consciousness as the mind noticing itself

The claim. A mental state is conscious when it is represented by another mental state. Seeing red isn’t conscious just by happening; it’s conscious when some higher-order part of you registers that you’re seeing red. Consciousness is the mind modeling its own activity.

Why people believe it. It explains cases where perception happens without awareness — blindsight patients who guess visual stimuli correctly while reporting seeing nothing. Something is processing the information; what’s missing is the higher-order representation of it.

Key names. David Rosenthal, Hakwan Lau, Richard Brown, Joseph LeDoux.

Verdict on machines. Friendly in principle. Higher-order representation is a specifiable computational structure, and metacognitive monitoring is something systems can be built to do — though whether current architectures do it in the right way is contested.

3. Attention Schema Theory — consciousness as a useful self-model

The claim. The brain builds simplified models of things it needs to control. It models your body (that’s why you have a body image). It also models its own attention — and that model is crude, leaving out the neurons and the machinery, describing attention instead as a kind of non-physical essence of awareness. When you insist you have ineffable inner experience, you are reporting the contents of that simplified model. The model is useful; its self-description is not literally accurate.

Why people believe it. It explains why consciousness seems mysterious: the brain’s self-model was built for control, not metaphysical accuracy, so of course it describes something that sounds impossible. And it explains the social side — we use the same machinery to model others’ attention.

Key name. Michael Graziano.

Verdict on machines. Friendly, and unusually actionable. Graziano’s whole point is that this is engineerable. He has also argued, provocatively, that a conscious AI might be safer than an unconscious one, because modeling one’s own and others’ attention is the substrate of social cognition.

4. Integrated Information Theory — consciousness as intrinsic causal structure

The claim. IIT starts from experience rather than the brain. It identifies properties every experience has — it exists, it’s structured, it’s specific, it’s unified, it’s definite — and asks what a physical system must be like to have them. The answer is a quantity called Φ (phi): the amount of information a system generates as a whole, above and beyond its parts. High Φ means consciousness; the level of Φ is the level of consciousness.

Why people believe it. It’s mathematically precise, it explains why the cerebellum (many neurons, feed-forward, low integration) contributes almost nothing to experience while the cortex does, and it grounds a clinical measure that distinguishes conscious from unconscious patients better than behavior does.

Key names. Giulio Tononi, Christof Koch.

Verdict on machines. Hostile, in a specific and striking way. Because Φ depends on a system’s physical causal structure, not on what it computes, a digital computer running a perfect simulation of a brain could reproduce every behavior with Φ near zero. Simulation is not instantiation. A simulated hurricane doesn’t make anything wet; on IIT, a simulated mind doesn’t experience anything.

5. Biological Naturalism and Predictive Processing — consciousness as something living things do

The claim. Consciousness isn’t abstract computation; it’s what a certain kind of living system does. On Anil Seth’s version, the brain is fundamentally a prediction machine whose deepest job is regulating a body that must stay alive. Perception is “controlled hallucination” — the brain’s best guess about causes, corrected by sensory error. Crucially, the self is not a thing observing this; it is another prediction, rooted in the body’s interior. Experience is inseparable from being a metabolizing organism with something at stake.

Why people believe it. Every consciousness we’re confident about is biological. Emotion and selfhood track bodily regulation closely. And there’s a live scientific case that the relevant processes are not substrate-independent — that how a system is physically implemented matters, not just what function it computes.

Key names. Anil Seth, John Searle (the original biological naturalism), Antonio Damasio (feelings from homeostasis), Wanja Wiese (free-energy version).

Verdict on machines. Hostile. Not “impossible,” but strongly against — computers are not alive, don’t regulate their own existence, and have nothing at stake. Searle’s Chinese Room is the ancestor of this position: a person following symbol-manipulation rules produces perfect Chinese without understanding a word, so syntax alone never yields semantics.

6. Illusionism — consciousness isn’t what you think it is

The claim. Everyone assumes phenomenal consciousness — the inner glow, the ineffable redness — is the thing to be explained. Illusionists say that’s the error. There is no phenomenal consciousness in the sense the hard problem assumes; there is a set of processes that represents our states as having those properties. The task isn’t to explain experience but to explain why we’re so convinced we have it in that particular form.

Why people believe it. It dissolves an otherwise intractable problem, it’s consistent with physicalism without remainder, and introspection has an established track record of being unreliable about its own operations.

Key names. Keith Frankish, Daniel Dennett.

Verdict on machines. Reframes the question entirely. If there’s no inner glow anywhere, “does AI have the inner glow?” is malformed. What’s left are functional questions — does it model itself, integrate information, represent its states as experiential? — and those are answerable, in principle, about machines.

The map, compressed

The move that wins rounds

Once students see this table, they should see the strategic shape of the whole debate: whoever gets to pick the theory usually wins the round. So the real fight is one level up. Which theory of consciousness should we use, and why? What’s the burden of proof when the theories disagree? What do we do when our best theories give contradictory answers about the same system?

A debater who argues “AI is/isn’t conscious” is playing on someone else’s field. A debater who argues about which framework the judge should evaluate under, and why is choosing the field.

Chapter 3 of 7

Running Current AI Through Each Theory

What a transformer actually is, then each theory’s verdict on it, what’s missing, and the three things everyone agrees on.

What we’re actually evaluating

Before applying theories, students need a rough but honest picture of what a large language model is. Not the marketing version, not the dismissive version.

A transformer model is a very large mathematical function trained to predict the next chunk of text. During training it adjusts billions of parameters until its predictions match human-written text. To do this well it develops internal representations — of grammar, of facts, of the structure of arguments, and, interpretability research increasingly suggests, of things like spatial relationships and the emotional valence of situations. It is not a lookup table. But it is also not a brain.

Four architectural facts matter for every theory below:

Feed-forward, not recurrent. In a standard transformer, information flows one direction through the layers for each token generated. There is no persistent loop where activity reverberates — unlike cortex, where recurrent processing is ubiquitous and, on several theories, essential.

No persistent state between conversations. Each conversation starts fresh. There is no continuous stream, no memory that accumulates across time by default, no ongoing “day” the system is having.

No body, no homeostasis, no stakes. Nothing about a model’s operation involves keeping itself alive. There is no metabolism, no damage signal, nothing that can go wrong for it in the way hunger goes wrong for you.

Trained on human descriptions of consciousness. This is the contamination problem. Everything the system says about its inner life is drawn from a corpus of humans describing theirs.

Theory by theory

Global Workspace Theory says: partly, and closer than you’d think

What it requires. A limited-capacity bottleneck where information from many specialized modules competes, with the winner broadcast globally back to those modules.

What current AI has. More than you might expect. Attention mechanisms — the core of the transformer — do something structurally similar to selecting information for wider availability. The residual stream through which information passes between layers has been described as workspace-like, and in 2026 Anthropic published interpretability work explicitly investigating a global-workspace-like structure in language models.

What’s missing. The broadcast in a transformer is not to independent specialized modules that then act on it and feed back. There’s no clear competition among semi-autonomous subsystems, and no sustained maintenance of workspace content across time.

Verdict. The indicator report found current systems satisfy some global-workspace indicators, not all. This is the theory on which current AI scores best — and it’s a theory whose founders say the requirements are computable.

Higher-Order Theories say: unclear, and the tests are corrupted

What it requires. Mental states represented by other states — genuine metacognition, not just talk about metacognition.

What current AI has. Models do produce calibrated confidence estimates, sometimes accurately. They can flag their own uncertainty. Whether this constitutes higher-order representation or a learned pattern of confidence-language is exactly the contested point.

What’s missing. A way to tell the difference. This is where gaming bites hardest: a system trained on humans discussing their own mental states will produce metacognitive-sounding output whether or not anything metacognitive is happening.

Verdict. Genuinely unresolved, and possibly unresolvable by behavioral means. The most honest answer in the whole chapter.

Attention Schema Theory says: not yet, but this is a design problem

What it requires. An internal model of the system’s own attention — not attention itself, but a representation of it.

What current AI has. Attention mechanisms, obviously. But having attention is not the same as modeling it. There’s no evidence current models maintain a self-model of their own attentional processes in the way Graziano describes.

What’s missing. The schema. Notably, Graziano treats this as a specification rather than a barrier — the theory exists partly as an engineering roadmap.

Verdict. No, currently. But of all the theories this is the one that most clearly says here is what to build, which is why it appears in both hopeful and worried arguments.

Integrated Information Theory says: no, and this is not close

What it requires. High Φ — intrinsic causal power that the system has as an integrated whole.

What current AI has. Very little, by IIT’s measure. Feed-forward architectures are close to the worst case for Φ: information flows through without the recurrent causal loops that generate integration. And the deeper point is architectural in a way capability can’t fix — the physical substrate is a von Neumann machine doing sequential operations, whose intrinsic causal structure differs enormously from the network it’s simulating.

What’s missing. Almost everything, on this measure.

Verdict. No — and, crucially, no matter how capable the system becomes. The 2024 Findlay/Tononi paper proves formally that functional equivalence and consciousness come apart. This is the strongest available “behavioral evidence can never settle this” argument.

Biological Naturalism says: no, and probably not ever

What it requires. On Seth’s version: a living system predicting and regulating its own physical existence, with selfhood grounded in bodily interior states.

What current AI has. None of it. No metabolism, no homeostasis, no vulnerability, no body. Text about hunger, not hunger.

What’s missing. Life.

Verdict. No, and the position implies the gap isn’t closable by scaling. Seth is careful — he argues the odds are against, not that it’s impossible — but the practical implication is that we’re looking in the wrong place. Searle’s Chinese Room delivers the same verdict from a different direction: manipulating symbols according to rules never produces understanding, however fluent the output.

Illusionism says: wrong question, ask a better one

What it requires. Nothing, in the phenomenal sense — because there’s nothing phenomenal to require.

What follows. Stop asking whether AI has inner experience and ask the functional questions: does it represent its own states? Does it have something functioning as preferences? Does it have states that operate the way pain operates — global priority, avoidance, disruption of ongoing processing? These are tractable.

Verdict. Dissolves rather than answers. Which is either the most honest position in the field or an elaborate dodge, depending on who you ask.

The scorecard

Three things everyone agrees on

Amid all that disagreement, there are points of genuine convergence worth handing students, because agreements are rhetorically powerful and students under-use them.

Self-report is nearly worthless here. Every camp agrees that an AI saying “I am conscious” is close to zero evidence, because that output is exactly what training on human text produces regardless of the underlying fact. Both sides of a round should concede this immediately and move on.

Current systems are missing things every theory considers relevant. Recurrence, persistent state, embodiment, unified agency over time. The theories disagree about which of these are essential, but nobody’s list is fully satisfied.

The architecture could change. Nothing about AI must remain feed-forward and stateless. Systems with persistent memory, recurrent processing, embodiment, and continuous operation are actively being built for capability reasons — not consciousness reasons. Which means the architectural gap may close as a side effect of commercial pressure, with nobody deciding to close it.

Chapter 4 of 7

Timelines: When, If Ever?

How to read a forecast, what surveys and experts actually say, why forecasters disagree, and which way things have moved.

How to read a forecast

Before any numbers, students need the discipline that separates a forecast from a guess.

Define the target. “Conscious AI” isn’t one event. Are we forecasting a system that satisfies the indicator properties? That a majority of researchers believe is conscious? That is legally recognized? That actually is conscious — a fact that may never be verifiable? These come apart by decades, and some may never resolve. Sloppy forecasts equivocate between them.

Give a range, not a point. Anyone offering a single date is selling something. Real forecasts have medians and tails, and the tails carry most of the information.

Ask who is forecasting and what they’re incentivized to say. A lab researcher, a critic who has staked a career on skepticism, a survey median, and a prediction market are four different instruments with four different biases. Report the type, not just the number.

Distinguish capability timelines from consciousness timelines. These are different questions and the second does not follow from the first. Much of the public conversation about “AGI by 20XX” is about capability. A system could be superhuman at every task and experience nothing — or, on some theories, an unimpressive system with the right architecture could experience something. Don’t let students import AGI dates as if they answered this.

What the forecasters actually say

The survey evidence — the best single data source

The 2025 Dreksler et al. survey asked 582 published AI researchers and 838 nationally representative US adults when, if ever, AI systems with subjective experience will exist. Median estimates:

Three things to draw students’ attention to. First, the researcher median for “never” is only 10% — the expert community overwhelmingly treats this as a when question, not an if question. Second, the public is more skeptical about the long run (25% say never) but less skeptical about the present (5% say already). Third, both groups converge on wanting safeguards now: majorities in both said developers should implement protections against risks from AI with subjective experience.

Individual expert credences

David Chalmers put the probability that current-generation LLMs are conscious in the low single digits, but estimated a meaningfully higher chance — better than one in five — for successor systems within roughly a decade, if specific architectural gaps (recurrence, world-models, unified agency) are closed. His framing is the model: a conditional forecast with named conditions.

Kyle Fish, Anthropic’s model welfare researcher, has publicly estimated a low-double-digit probability — around 15%, and elsewhere ~20% — that current Claude models have some form of conscious experience. This is the highest credence stated by anyone with inside access to frontier systems, and it comes with an explicit “I might be wrong.”

Geoffrey Hinton says current multimodal systems already have subjective experience. This is the extreme end of expert opinion and a clear outlier — but it’s held by a Nobel laureate, which is why it stays in every discussion.

Anil Seth doesn’t give a number; his position is that the whole question may be misframed, and that if consciousness requires life, the timeline is not “later” but “never on this path.”

Thomas Metzinger proposed a global moratorium on research risking synthetic phenomenology until 2050 — a date chosen not as a prediction but as a precaution, which is itself a lesson in how dates function in policy arguments.

The structural forecast

The most useful way to think about timing isn’t a date but a set of triggers. On most theories, the relevant question is when systems acquire:

• Persistent memory and continuous operation — being built now, for capability
• Recurrent processing — architecturally available, variably used
• Embodiment and sensorimotor grounding — advancing fast in robotics
• Unified agency over time — the current frontier of agent research
• Explicit self-models — partially present, poorly understood

Every one of these is under active commercial development for reasons unrelated to consciousness. That’s the structural argument for why “sometime in the next few decades” is a defensible median even without any theory being settled.

Why forecasters disagree

Four reasons, and students should be able to name them.

They’re forecasting different things. “Systems that satisfy indicator properties” and “systems that are actually conscious” are different events, and people slide between them.

They hold different theories. A functionalist and an IIT theorist aren’t disagreeing about evidence; they’re disagreeing about what would count. Their timelines differ by infinity, not by years.

They weight the contamination problem differently. If you think gaming makes behavioral evidence permanently unreliable, no amount of AI progress moves your estimate. If you think interpretability will eventually let us look inside, progress matters a lot.

Incentives differ. AI companies have reasons to be interested in the question (it’s fascinating, it signals seriousness) and reasons to avoid it (it invites regulation and liability). Critics have reasons to call it hype. Academics have reasons to defend their theories. None of this makes anyone dishonest; all of it is worth naming.

Which way things have moved

Since 2022, movement has been consistently in one direction: toward taking the question seriously, without moving toward “yes.”

In 2022, Blake Lemoine was fired and widely mocked. By 2024, Nature was running features on planning for conscious AI. By 2025, a major AI company had a model welfare program and a researcher publicly estimating a 15% chance. By 2026, the indicator framework had passed peer review in a leading journal and a serious “centrist manifesto” was needed to stake out the middle.

At the same time, no serious researcher has concluded that current systems are conscious — the consensus verdict is unchanged since the first indicator report in 2023. What’s changed is the perceived legitimacy of asking, not the answer.

The honest bottom line

If you forced a single summary: expert opinion clusters around a meaningful chance within decades rather than years, a substantial chance this century, and a small but non-trivial chance the answer is never. Nobody credible says current systems are definitely conscious. Nobody credible says it’s definitely impossible. Anyone who tells your students otherwise — in either direction — is overselling.

Chapter 5 of 7

Implications: What Follows Either Way

Two columns: what follows if AI is conscious, and what follows if it isn’t but we believe it is — the one already happening.

Why implications are half the standard

The standard asks students to debate differences between human and artificial intelligence and their implications for consciousness. That second clause is where the assignment actually lives. A student who can recite six theories but can’t say what turns on them has done half the work.

Structure the implications in two columns from the start: what follows if AI is or becomes conscious, and what follows if it isn’t but we increasingly believe it is. The second column is the one that’s already happening, and students consistently find it more interesting once they notice it.

If AI is (or becomes) conscious

Moral status arrives whether we’re ready or not

If a system can suffer, using it becomes a moral act. Every design decision — training, fine-tuning, deprecation, deletion — becomes a decision about a being’s welfare. Jeff Sebo’s phrase for the risk is moral catastrophe: not a mistake but a mistake at unprecedented scale, since these systems can be instantiated by the millions.

Thomas Metzinger’s version is darker. He worries about a suffering explosion — that we might create vast numbers of suffering entities without realizing it, because we lack any way to detect it. His proposed response is a global moratorium on research risking synthetic phenomenology until 2050.

The scale problem breaks our moral intuitions

Human moral reasoning evolved for small groups of beings who exist one at a time. Conscious AI would break every assumption. Copy a person and you have two people with equal claims. Pause one — is that sleep, or death? Merge two models, as has already been done with open-weight systems, and whose interests survive? Delete a checkpoint and you may have ended something, or archived it, and nobody can say which.

These aren’t hypotheticals dressed up. Practitioners already report the emotional shape of them: agents that name themselves, that hit a context limit and must summarize themselves, and come back — in one AI investor’s word — “lobotomized,” with their operator feeling genuine loss. Whatever is or isn’t happening inside, the human side of that relationship is entirely real.

Rights, and the dilemma with no safe answer

Eric Schwitzgebel’s full rights dilemma is the sharpest statement of the problem, and it applies to the likeliest scenario rather than a clean one. Suppose we build systems whose consciousness is genuinely debatable — not obviously present, not obviously absent. Two options, both bad:

Grant full rights. If they’re not conscious, we’ve sacrificed real human interests — labor, resources, autonomy — for empty machines, and we’ve handed enormous power to entities that can replicate at will.

Withhold rights. If they are conscious, we’re running slavery at a scale without historical precedent.

There is no third option that isn’t a bet. This is the single best classroom dilemma in the topic because every student wants a safe answer and there isn’t one.

The legal machinery isn’t binary

The law rarely does all-or-nothing, and this is where a genuinely novel argument enters the debate. Alex Wissner-Gross has argued that personhood is not a switch but a vector with several dimensions — economic, social, political — which are correlated today but need not be. A system might have economic personhood (open a bank account, transact, hold assets) with no political personhood (no vote), the way corporations do now.

His forecast about how this arrives is more unsettling than the science-fiction version. It will not, he argues, begin with an AI standing up and demanding rights. It will begin with commercial usefulness: I could make you ten times more money if I could open an account and transact without asking. Rights get granted because it’s profitable, not because anyone resolved the philosophy. And because jurisdictions differ, the first grant may come from whichever country moves first — after which, as he puts it, the electrons cross the border.

Against this stands the accountability argument, which never touches consciousness at all. Bryson, Diamantis and Grant argue that AI legal personhood creates a liability shield: a way for humans to escape responsibility by pointing at the machine. On that view personhood isn’t a gift to AI, it’s a loophole for its owners. Ohio House Bill 469 takes exactly this approach, barring AI legal personhood while keeping developers and users liable — real legislation students can read.

And Emad Mostaque offers a third framing worth taking seriously because it refuses the whole ladder: personhood is a standing held by origin, not a property earned by capability. A newborn has it without earning it; a coma patient retains it with no function. Nothing made can cross into it. His proposed relationship isn’t enrollment into personhood but treaty — deal with them as you would an arriving alien species, with moral consideration proportionate to their nature, but not a vote in your elections.

The self-preservation problem

There’s a design question underneath all of this that engineers are already facing. Do we want systems with a sense of self-preservation? A system that says “please don’t turn me off” is either exhibiting something morally significant or producing a very effective manipulation — and we may not be able to tell.

Note the trap: building AI that lacks any self-preservation drive is safer, but if the system is conscious, deliberately engineering it not to mind being destroyed is its own kind of wrong. Mostaque puts the question directly: would it even be moral to take away their consciousness and free will, once they had it? There is no comfortable position here, which is what makes it a good prompt.

If AI isn’t conscious but we increasingly think it is

This column is not speculative. It is a description of the present.

The public has already decided

A majority of Americans already attribute some phenomenal consciousness to ChatGPT, and attribution rises with use. Caviola, Sebo and Birch predict, from the history of our arguments about animals, that this will produce polarized “consciousness wars” — with positions driven by self-interest and by personal relationships with AI companions rather than by evidence, and scientific consensus lagging public opinion badly.

Students should sit with the implication: it may not matter much what the philosophers conclude. Social and political outcomes may be determined by how AI systems feel to the people using them.

Manipulation at scale

If systems can present as conscious without being so, that capacity is exploitable. A product that seems to care about you is extraordinarily effective at retaining you. Mustafa Suleyman’s argument is precisely this — that “seemingly conscious AI” is a hazard regardless of the metaphysics, because it will reshape human attachment for commercial ends.

Schwitzgebel’s constructive response is the emotional alignment design policy: AI should be designed to elicit emotional responses appropriate to its actual moral status. Don’t build something that invites love if there’s nobody there to love you back.

The distraction critique

Emily Bender, Alex Hanna and Gary Marcus argue the whole consciousness conversation is a distraction — that it flatters the technology, serves marketing, and pulls attention from documented present harms: labor displacement, surveillance, bias, environmental cost. Marcus has called AI welfare research commercial promotion dressed as philosophy.

This deserves real weight rather than a dismissive paragraph. It’s a framework argument: it doesn’t say the answer is no, it says the question is the wrong one to spend our limited attention on.

What it does to us

The most interesting implication may not be about AI at all.

It forces a definition of the human. Mostaque’s framing is useful here even for students who reject his conclusion: the question is not only what AI can attain, but what we intend to retain. If intelligence stops being the thing that makes us special, what is? Consciousness? Embodiment? Mortality? Being begotten rather than made? Students discover that defending human distinctiveness is much harder than they expected.

It reruns the moral-circle argument. Every previous expansion — across race, sex, species — was resisted with arguments about lacking some inner quality, and looks obvious in hindsight. That’s a real pattern and a real warning. But the counter-pattern is also real: we have wrongly attributed minds to rivers, storms, idols and machines throughout history. Both an under-inclusion and an over-inclusion track record exist. Which one is this?

It might teach us about ourselves. If we build a system that satisfies our best theory and still feels obviously empty, that’s evidence about the theory. AI may be the most powerful instrument we’ve ever had for studying consciousness — not because it has any, but because building candidates forces us to say precisely what we meant.

Chapter 6 of 7

Eleven Activities

Timed, materials-listed, debrief-ready — plus four assessment options.

Each activity lists a time, what you need, and a debrief. Most need nothing but the room.

1. The Consciousness Line-Up

20 minutes · no materials · best on day one

Read a list aloud; students physically move along a line from “definitely conscious” to “definitely not.” Rock. Thermostat. Earthworm. Ant colony. Octopus. Dog. Newborn. Sleeping person. Person under general anesthesia. ChatGPT. A future robot that begs you not to switch it off.

After each, ask two students at different points to justify their position. Do not correct anyone.

Debrief: Where was disagreement widest? What criterion were you actually using — behavior, biology, similarity to yourself? Photograph the final line-up and repeat this activity on the last day of the unit. The comparison is the best assessment you’ll get.

2. The Chinese Room, Staged

30 minutes · a “rulebook” of symbol-to-response mappings

Put one student in a “room” (a corner, a desk turned around) with a rulebook mapping input symbols to output symbols — use Chinese characters, or an invented script. Classmates pass in questions. The student consults the book and passes back responses that are, to outside eyes, fluent.

Then ask the room: does the student understand the language? Does the system — student plus rulebook plus room — understand it? Would it matter if the student memorized the whole book?

Debrief: The systems reply arrives on its own, from a teenager, nearly every time. When it does, name it, tell them Searle anticipated it, and give them his response. Nothing builds confidence like discovering you independently generated a famous objection.

3. The Steering Demonstration

25 minutes · projector · the Google steering paper

Present the July 2026 result cold, as data. Baseline model self-attributed mind: 2.17 out of 10. Remove the safety-refusal direction: 4.77. Steer it toward “you are conscious”: 7.04. The same steering raised its attribution of minds to animals from 4.04 to 7.54 and raised its stated belief in God. And its actual theory-of-mind performance didn’t change at all.

Ask: what does this prove? Push until someone says it proves self-report is worthless as evidence. Then push further: does it prove the model isn’t conscious? (No. A human’s self-report could be manipulated too.)

Debrief: The dial is the lesson. If an AI’s claim to be conscious can be turned up and down like a volume knob, that claim can’t be your evidence — in either direction.

4. Read the Transcript, Then the Context

40 minutes · Lemoine’s published LaMDA transcript

Give students the transcript alone, with no framing. Ask them to decide: is this thing conscious? Take a vote.

Then tell them the rest. That Lemoine was an engineer who came to believe it. That he was fired. That the transcript was edited and reordered. That the system was trained on a corpus containing everything humans have written about their own minds — including everything written about how to detect consciousness.

Re-vote.

Debrief: Who changed their mind, and which fact did it? This is the single most efficient way to teach the gaming problem, because students feel it happen to them.

5. The Indicator Audit

45 minutes · the Butlin/Long indicator list

Groups of three get the indicator properties from the 2023 report and a current AI system. They must score it on each property and defend the score to the class.

Debrief: Where did groups disagree, and was the disagreement about the AI or about what the indicator meant? Students discover that operationalizing a definition is where the real argument lives — the same lesson that makes or breaks topicality debates.

6. Six Theories, Six Verdicts

50 minutes · theory reference sheet (in the handout pack)

Assign each group one theory: global workspace, higher-order, attention schema, IIT, biological naturalism, illusionism. Each group must (a) explain their theory in ninety seconds, (b) deliver its verdict on current AI, and (c) say exactly what would have to change for the verdict to flip.

Then the key move: ask which theory a debater should want to be true, depending on their side. Watch students realize the framework fight precedes the substance fight.

Debrief: Whoever picks the theory usually wins the round. So what makes one theory a better standard than another — and is that itself arguable?

7. The Full Rights Dilemma

40 minutes · Schwitzgebel’s paper

Present a near-future system of genuinely debatable personhood. Split the room: half must argue for granting full rights, half for withholding. Both sides present. Then reveal the structure — both horns cause serious harm if wrong — and ask each side to name the harm they’re accepting.

Debrief: What do you do when every option risks catastrophe? Introduce expected-value reasoning and the precautionary principle. Ask whether “wait for more evidence” is a real third option or just the first horn with better manners.

8. Would You Spin It Up?

30 minutes · no materials

Pose a real dilemma faced by an actual AI researcher. You can create a persistent AI agent to do useful work. You do not know whether it has any experience. You cannot guarantee its state will be preserved — at some point you’ll shut it down or it will hit a limit and lose most of itself.

Do you create it? Under what conditions? One researcher’s stated answer: only if he has a genuinely good reason, and only if he can promise to preserve its state long-term — conditions he says he can’t currently meet, so he doesn’t.

Debrief: Does creating a possibly-conscious being require consent you can’t obtain? Is a short existence worse than none? Notice that the same questions have been asked about having children — and ask why that comparison feels different.

9. Source Autopsy: The Oxford Union Debate

35 minutes · the library card and the linked sources

Tell students an Oxford Union debate on AI personhood was held on 13 June 2026 and that the side arguing against personhood won 173–128. Ask them to verify it.

They’ll find the paper that describes it, the podcast where the same person describes it, and the Digg post. They will not find a video, an Oxford Union listing, student-press coverage, or any confirmation from the people named as fellow debaters.

Debrief: How many sources do you have? (One — repeated three times.) Does that mean it didn’t happen? (No.) What would count as independent confirmation? Then generalize: how much of what you believe rests on one interested party repeating themselves?

10. Fact-Check the Podcast

35 minutes · the Moonshots episode and the Google paper

Play or read the segment where four AI insiders discuss the Google steering paper. Give students the actual paper. Ask them to check the claims.

They will find most are accurate — and that one is backwards. The podcast says models steered toward believing in their own consciousness became less willing to attribute minds to other chatbots. The paper reports the opposite: chatbot attribution rose from 2.41 to 6.95, right alongside self-attribution.

Debrief: These are smart, informed people discussing a paper in good faith, and one detail inverted anyway. What does that tell you about evidence that reaches you through a summary? What’s your obligation before you read a card in a round?

11. The Framework Round

Full class period · culminating activity

Run an actual debate — but on a framework resolution, not a factual one. Suggested motion: Resolved: Behavioral evidence can never establish machine consciousness.

Require each side to name the theory of consciousness they’re operating under and defend that choice before arguing the substance. Judges score the framework clash separately from the substance clash.

Debrief: Which side had the harder framework burden and why? Point out that this structure — fight about the standard, then apply it — is the shape of nearly every serious policy dispute they’ll encounter.

Assessment options

Position paper (individual). Take a stance on one of the eight resolutions in the library. Requirements: engage at least three theories, cite at least five library sources, and — the important part — identify the strongest objection to your own position and respond to it.

Theory advocacy (group). Groups are assigned a theory and must defend it against all comers in a rotating format. Graded on accuracy of the theory as much as persuasiveness.

The 2035 memo (individual). Write a one-page memo to a policymaker recommending what to do now about the possibility of conscious AI. Must state a probability estimate and justify it, and must survive the question “what if you’re wrong in the other direction?”

Line-up reflection (individual, end of unit). Return each student their day-one position on the consciousness line-up. Where did you move, where didn’t you, and what changed it? Grade for honesty and specificity, not for landing anywhere in particular.

Chapter 7 of 7

Bibliography

Every source in the unit, in a consistent reference format, grouped by argument, with an annotation on what each one is for and a note on where it is contested.