Copium: Why We Don’t Even Think About AI Risk

Seeking engagement in a research project

This is a link post for https://out-of-distribution.ai/copium/proposal published in April 2026 — that’s the canonical home, with the full reference list. Republished here so you can comment.

---

The argument that superintelligence could end us is not hard to follow. We are building systems we cannot yet reliably control or robustly specify goals for, on a trajectory toward capabilities that exceed our own. This warrants caution. Why would we be confident this technology will be safe? Yet humanity’s response has been vanishingly small at every level: individual, institutional, political. We find poor reasons not to worry about or engage with the argument and its consequences.

I refer to this as AI copium. I believe it to be the most important barrier in communicating about the risk, and one that dangerously impairs our collective thinking.

This post sketches copium’s likely psychological roots, identifies three under-explored research opportunities, and suggests some candidate interventions we might borrow from adjacent fields.

I’m looking for collaborators to discuss and advance, to reduce the odds that superintelligence kills everyone. So the post is written for people already advocating AI safety to others in policy, research, industry, and the general public. I would use different wording to describe the issue directly to those audiences.

What copium is

“Copium” is internet slang for the psychological comfort people reach for when reality is unacceptable. I’m borrowing and sharpening it, and keeping the register deliberately: the casual term is part of the point. This is a name for something people viscerally recognise, and one that emerged in conversations with others in the AI domain, not a dry construct from the literature.

AI copium is the aggregate psychological effect that makes humanity’s response to superintelligence risk so much smaller than the risk warrants.

It operates from first exposure onward. “I don’t want to think about this” on hearing the first sentence is copium. So is “ah, but the engineering constraints make that unlikely” from someone who has understood the argument fully. So is quiet agreement followed by no change in behaviour. Copium can prevent engagement, prevent acceptance, and prevent action; it typically works on several of these at once. “I don’t buy the risk, and anyway, even if I did, surely someone will figure it out?” — the second clause buttressing the first.

There is a real difference between copium and ignorance: we are still early enough to regularly encounter someone who genuinely hasn’t encountered the arguments. Beyond ignorance, copium is hard enough to avoid that it has probably impeded the thinking of anyone who has engaged at all — through avoidance on first contact, sophisticated rationalisation, or acceptance without action. We require people to simultaneously process the threatening information, and recognise their own defensive reactions to that information as problematic. This is recursively metacognitive in a way that naïve persuasion is not, and it creates a communication problem genuinely different from just explaining the risk.

People working in risk communication and motivated reasoning will recognise much of the psychology here. But there are three underappreciated aspects that I aim to highlight in this essay for AI copium in particular: how these mechanisms may compound specifically for superintelligence risk, how poorly the existing research measures them on the reception side, and how little effort has gone into borrowing interventions from adjacent fields.

Why superintelligence risk is distinctively bad

Defensive psychological responses to catastrophic risk are not new. But superintelligence risk activates them more strongly and in more compounding ways than other risks, due to specific features.

Timescales are compressed — potentially years, not decades. There’s little sensory on-ramp — no equivalent of a heatwave to make the abstract concrete, until it’s too late. The AI industry is culturally ascendant rather than delegitimised, so system justification pulls harder. There is no historical referent at all, so normalcy bias has nothing to push against. And the people with the most technical understanding of the risk are often the most deeply identity-invested in the industry producing it — meaning expertise amplifies copium rather than protecting against it. More cognitive resources means more sophisticated rationalisation, not less.

George Marshall‘s work at Climate Outreach — on identity-protective cognition, narrative capture, and post-disaster response — is deeply relevant. But importing the climate communication playbook wholesale risks importing its twenty-five-year track record of insufficient action. AI copium is a more severe instance of the general phenomenon, rather than a rerun.

What generates it: a partial map

The psychology literature offers well-established mechanisms that each explain part of the aggregate effect. This is my current best subset — extending and refining it is part of the work to be done.

Terror Management Theory (Greenberg, Solomon, Pyszczynski) names the foundational reaction. Confronting one’s mortality triggers “defensive worldview reinforcement” where people retreat into identities and arrangements that promise symbolic permanence. Identity-protective cognition (Kahan) names what that defence looks like in cognitive processing: threatening information is filtered through identity-preserving criteria. “I work in tech” and “I believe in progress” are load-bearing identities for millions. System justification (Jost) extends the defence outward to the social arrangements that house identity: existing institutions are treated as legitimate or inevitable. The deeper AI is embedded in economic life, the stronger the pull. That three research traditions converge on overlapping defensive territory is itself part of the picture: the same underlying operation gets rediscovered under different names, giving the avoidance multiple anchors.

Cutting across these are further mutually reinforcing failures of the emotional and imaginative systems to process magnitude, and novel ideas. Psychic numbing (Slovic) starts close to home: a single identifiable victim moves people more than a statistic about thousands. The emotional system can’t engage proportionally. At millions, this can get worse: engagement even decreases. The extinction of billions or future zillions ups the ante into territory for which we have no prior models, which is where normalcy bias also activates. This bias against considering the unprecedented hits AI risks in many ways beyond scale. Nothing like superhuman thinking agents has happened before. The very feature that should make AI risk distinctive instead disarms us. System justification feeds back into this: it supplies ready-made narratives that normalcy makes credible. “Naive scaling arguments often fail due to not foreseeing new constraints, fears of technology have regularly proven unfounded. Things will continue as they have.” Whichever rationalisation is most available in the moment gets deployed; the others sit ready as backup, lending it a feeling of robustness.

I want to note that the psychology humans evolved re fear was not a dumb mistake. These factors are typically heuristics that work reasonably well, especially in our original environment. Terror induces extreme short-lived responses: we react to the tiger quickly and we run for a while. Over a longer period, we can’t afford this. It isn’t useful to be permanently terrified, hence management and returns to stable patterns. Psychic numbing protects you and your family rather than your species. Normalcy biases help reduce false alarms. We really have discovered that we often fear change more than we should.

Alas, the prospect of superintelligence now takes our trained psychology out of distribution. We are reasonably optimized to succeed at things we are fit for. This decade, our generalizations can kill us. The pattern is ironically common in our reactions to AI risk.

I also flag that this compounding claim is a hypothesis. It’s plausible simply that different mechanisms dominate in different people, with no within-individual compounding. Whether the reinforcement operates within individuals or across the population distribution will matter for intervention design and should be empirically distinguishable; our research programme should survive the compounding hypothesis being measured and found wanting.

But intuition and anecdote strongly suggest it. I have smart friends who I have discussed and solved difficult problems with, thoughtfully and in depth. Then I observe how badly debate on AI risk proceeds with them. Watching brilliant people desperately seek, grab, drop and cycle through a remarkably large and varied set of possible rationalizations across minutes until they find one that kind of works and cease to engage is saddening and maddening. Tell me you haven’t watched this happen. And did you not do it yourself? “Anyway, even if that is true, I just think that maybe this other thing, and what could anyone really do” — it is embarrassing.

A mother indulges her son. “Well, yes, I can see how it could be dangerous.” But copium manifests so bizarrely. “The thing is, I’m pretty old. I’ll be dead soon. And I’m selfish — I don’t care enough about future generations for it to matter.” Will she repeat that on reflection? “You can’t fight government and business. We kind of deserve it anyway.” Please speak directly into the microphone. Don’t forget that you can claim “it is morally wrong to favor ourselves over a successor species” if you want to.

What the data shows

Several independent research groups have now produced measurements at population scale that are hard to explain without something like copium.

Philip Trippenbach and the Seismic Foundation ran a randomised controlled trial with 1,063 respondents and found extinction framing was the lowest-performing theme across all demographics — worse than framing AI risk in terms of jobs or children. The Social Change Lab surveyed 3,467 UK adults and found existential risk was both the lowest concern and the area with the lowest willingness to act, out of eleven AI risk categories. The Existential Risk Observatory has tracked US spontaneous awareness of AI extinction risk from 7% to 24% over three years — awareness is rising, but it isn’t converting into action or political demand.

For context: the medical literature on terminal diagnoses finds 60–80% initial rejection and only 18–23% intention-to-action conversion even though that concerns risks that are personal, definitive, and immediate. AI extinction risk is probabilistic, collective, and unprecedented, so we should expect the conversion problem to be tougher. My own rough estimate from ERO’s data puts the AI-specific acceptance-to-action rate at somewhere between 1.4% and 3.3%, though this figure needs proper validation.

This message-testing research is valuable and necessary. But important things are missing.

Reception-side measurement. Nobody is systematically measuring the psychological mechanisms operating in recipients. We know extinction framing underperforms, but we’re not tracking why — which defences activate, in whom, under what conditions.

Set and setting. Comparing message framings without attending to context is methodologically questionable. The same message lands differently depending on who delivers it, in what setting, to someone in what emotional state. Few studies discuss this.

Cross-domain remedies. Health communication, climate psychology, crisis communication, and behaviour change research have all developed interventions for analogous problems. Even simple marketing is barely explored. The AI risk communication field is doing its own iteration on message variants rather than borrowing and adapting what already exists. Carefully rediscovering things is usual for contemporary rats, but this ship is sinking fast.

Candidate interventions worth exploring

I will sketch what “borrowing” might look like. These are candidates a working group should assess rather than blindly rolling them out.

Autonomy-supportive framing from motivational interviewing, which treats resistance as information about where the person is rather than as an obstacle to overcome. This directly addresses the reactance that identity-protective cognition produces.

Staged disclosure along the lines of SPIKES, the medical protocol for delivering bad news. Finding out what the recipient is already worried about is a way in. The transfer is imperfect since AI risk lacks consented listeners, individual agency, and definitive information but the underlying intuition that how bad news is delivered matters for whether it lands is sound. (To be clear, this does not extend to hiding extinction risks in favor of lesser ones — but a “yes, and” re disempowerment helps to make the connection.)

Vulnerability framings from therapeutic communication and peer-recovery models, where the speaker’s own past resistance is a more useful opening than naming the listener’s. See the related note on Bulverism below. This is the recursive metacognitive problem handled through shared position rather than expert instruction.

Targeted trusted-messenger work as developed in public health, which accepts that identity-protective cognition is not going away and that who delivers the message often matters more than what the message says.

None of these are silver bullets. Each has been developed for problems that resemble AI copium only partially. Adapting them is real work, and starting that work is part of what I’d like a working group to do.

A necessary caveat

C.S. Lewis coined “Bulverism” for the fallacy of assuming your opponent is wrong and then explaining their error to them through psychological analysis. We probably don’t want to lead with “you are thinking wrong”. How much to make the recipient themselves aware of acceptance issues in this case needs tested.

Explicitly, pre-measurement I am not inclined to say “you have copium, therefore you’re wrong.” Whether and when more direct psychologisation might help rather than harm is itself an empirical question this working group could pursue; the self-affirmation and motivational interviewing literatures suggest the answer mostly tilts against, and that “I had cope” is a more honest and effective opening than “you have cope” — but the conditions under which more direct moves help are measurable, and I’d rather measure than guess. I watched myself comprehend superintelligence risk and find reasons not to reorganise my life around it, for years. The psychological literature offers explanations for why this happens, and they’re consistent with the data now emerging.

What I’m looking for

I’ve spent time reading across terror management theory, risk perception, cognitive dissonance, normalcy bias, and behaviour change, while tracking the empirical work from ERO, Seismic Foundation, and Social Change Lab.

I want to assemble an informal working group to do three things: design studies that measure reception-side mechanisms, survey adjacent fields for transferable interventions, and develop communication approaches that account for copium rather than ignoring it. If you work in risk communication, behaviour change, social psychology, or AI safety advocacy, and any of the above resonates or provokes disagreement, I’d welcome the conversation.

anthony@pauseai.info

---

This essay was first written in the context of a London Futurists panel on 29 April 2026: “Talking about catastrophic and existential risks”.

Canonical version, with the full reference list: https://out-of-distribution.ai/copium/proposal.