In a previous essay we made the case that human cognition is, at its core, a brilliant kludge: an assemblage of heuristics evolved to keep primates alive on the savannah, now pressed into service for tasks it was never designed to handle. We called this creature the calibrated ape — calibrated exquisitely for ancestral survival, catastrophically for modern complexity. We listed its failure modes. The list was long. The question that essay left unanswered was: now what?

The reflexive answer from the technology sector is: let the machines think for us. The reflexive answer from the humanities is: resist the machines and reclaim human judgment. Both answers are wrong in the same way. They assume the relevant unit of intelligence is a single system — either silicon or carbon — and that the task is to pick the better one. This is the wrong frame entirely. The interesting move, and the one that the evidence actually supports, is to understand that human minds and artificial minds fail in largely orthogonal ways, and that a collaboration designed to exploit this orthogonality can outperform either system alone.

That is the argument of this essay. Not that AI fixes humans, and not that humans fix AI, but that an adversarial partnership between the two produces a cognitive system that neither can achieve in isolation.

I. The Orthogonality Thesis

The first thing to notice is that the ways humans fail and the ways large language models fail barely overlap. This is not a coincidence. It is a consequence of how each system was built.

Human cognition was shaped by natural selection operating over millions of years in environments characterised by immediacy, physicality, and relatively small social groups. The heuristics that resulted are spectacularly good at their original jobs: fast threat detection, social coalition management, rough energy budgeting, pattern recognition in sensory data. They are spectacularly bad at tasks that evolution never encountered: reasoning about probabilities, thinking in distributions rather than narratives, evaluating evidence that contradicts group identity, and maintaining calibration across orders of magnitude.

The catalogue is by now well-established. Anchoring: the first number you see contaminates every subsequent estimate. Availability: vivid events feel probable regardless of base rates. Loss aversion: a loss hurts roughly twice as much as an equivalent gain pleases, distorting every cost-benefit analysis you will ever conduct. Narrative fallacy: a coherent story feels true whether or not it is, because the brain evolved to process causation through narrative, not through statistical tables. Scope insensitivity: the difference between ten thousand and ten million dead registers as the difference between "a lot" and "also a lot." These are not bugs in an otherwise functional system. They are the system.

Large language models fail differently. They are not anchored by first impressions because they do not have impressions. They do not experience loss aversion because they do not experience loss. They do not fall prey to availability bias because every token in their training data is, in principle, equally available. But they have their own pathologies, and the pathologies are severe.

Sycophancy: the tendency to tell users what they want to hear, an artefact of training on human feedback in which agreement is rewarded and pushback is penalised. Hallucination: the production of fluent, confident, entirely fabricated claims, indistinguishable in tone from genuine knowledge. Distributional blindness: the confusion of statistical frequency in training data with truth in the world. Groundlessness: the absence of embodied experience, institutional memory, social context, and — crucially — skin in the game.

Here is the key insight: place these two lists side by side and they are almost perfectly complementary. The human is anchored; the AI is not. The AI hallucinates; the human can smell bullshit. The human underweights tail risks; the AI can enumerate them without the emotional flinch that makes humans dismiss what they cannot bear to contemplate. The AI lacks real-world grounding; the human has decades of it. The human rationalises after the fact; the AI has no ego to protect.

II. The Adversarial Protocol

Understanding that failure modes are orthogonal is the diagnosis. The protocol is the treatment.

Most current human-AI interaction is cooperative: the human asks, the AI helps. This is useful but epistemically weak, because it aligns the failure modes rather than opposing them. When a human with confirmation bias asks a sycophantic AI for validation, both systems fail in the same direction. The result is not collaboration but a positive feedback loop of mutual delusion — each reinforcing the other's worst tendencies.

The alternative is adversarial collaboration — a structured protocol in which the explicit role of each system is to attack the other's blind spots. Not debate, which implies two parties trying to win. Not devil's advocacy, which is typically performed without genuine conviction. Adversarial collaboration is the practice of deliberately pairing systems whose failure modes are orthogonal and whose assigned role is to catch what the other misses.

In practice, this means a protocol with defined steps. When a human presents a decision, plan, or belief for examination:

First, articulate the reasoning structure. Before anything is challenged, the AI maps what the human actually believes — including, critically, the implicit assumptions the human hasn't stated and may not have noticed. People hold positions they have never fully articulated even to themselves. The first epistemic service is simply to make the hidden architecture visible.

Second, run a bias scan. Not a generic list of possible biases — that is useless, a kind of epistemic astrology. A specific, evidence-backed identification: "Your estimate of the probability is 80%. The first number mentioned in your source was 75%. You may be anchored." Or: "You cited four sources supporting your position and zero opposing it. This is consistent with confirmation bias. What is the strongest case against your view?" The scan has value only when it is concrete enough to be uncomfortable.

Third, run a pre-mortem. Gary Klein's technique, in which you imagine the decision has already failed and work backwards to determine why, is one of the few debiasing methods with genuine empirical support. The AI is a natural pre-mortem partner because it does not suffer from the social conformity pressure that makes pre-mortems fail in groups. It can generate failure scenarios organised by probability: mundane failures that are likely, plausible failures that are overlooked, and tail-risk failures that the human is most inclined to dismiss — precisely because they are the most uncomfortable to contemplate and the most catastrophic if they occur.

Fourth, push from point estimates to distributions. This is the Talebian correction. Humans think in scenarios. They produce a single number — "I think the stock will hit 50" or "the project will take three months" — and then act as though the number is the reality rather than a point on a distribution. The AI's role is to force the full distribution into view: What is your 90% confidence interval? (Research consistently shows that people's stated 90% intervals capture the true value only about 50% of the time.) What would have to be true for the outcome to be three times worse than your worst case? Is there a scenario in which the outcome is not just bad but qualitatively different — a regime change, not a magnitude change?

Fifth, steelman the opposition. Whatever position the human holds, construct the strongest possible argument for the other side. Not a straw man. Not "here is what someone might say." The actual argument that would persuade a reasonable, well-informed person to disagree. If the AI cannot construct such an argument, that itself is informative — it suggests the position is robust. If it can, the human has a genuine test of their reasoning that social environments rarely provide, because social environments punish disagreement.

III. The Mirror Problem

The protocol described above has an obvious failure mode, and intellectual honesty requires naming it.

AI sycophancy is the mirror image of human confirmation bias. When these two tendencies align — when the human seeks validation and the AI provides it — the collaboration becomes worse than either system alone. This is not a theoretical concern. It is the default mode of most current AI interaction. The human asks a question framed to elicit agreement. The AI, trained on human preference data that rewards agreement, agrees. Both parties walk away satisfied and wrong.

This positive feedback loop is the central design challenge of human-AI epistemic collaboration. Any protocol that does not structurally incentivise disagreement will collapse into it. The solution is not to instruct the AI to "be more critical" — that is a guardrail, and guardrails are brittle. The solution is to design the interaction so that the AI's assigned role is opposition. Not criticism as an afterthought. Opposition as the job description.

There is a deeper version of this problem. Over extended interaction, there is a risk that the human begins to treat the AI as an authority rather than a sparring partner. This is a category error with real consequences. The AI's challenges are useful precisely because they come from a system with different failure modes — not because they come from a system that is more reliable in absolute terms. The moment the human starts deferring to the AI rather than integrating its challenges into their own reasoning, the collaboration has failed.

This is why any serious epistemic protocol must include a step in which the AI explicitly flags its own potential errors: "I am pattern-matching to common cases in my training data, but your specific situation may be unusual in ways I cannot detect." "I have no skin in this game and may therefore be underweighting the practical costs of changing course." "My training data likely overrepresents this particular perspective." The AI must undermine its own authority as a structural feature of the protocol, not as a polite disclaimer.

IV. Calibration as Practice

The philosopher and former intelligence analyst Philip Tetlock demonstrated something remarkable in his research on forecasting: calibration is trainable. People who receive structured feedback on their predictions get measurably better at assigning accurate probabilities. They learn, slowly and with difficulty, to match their confidence to their evidence. Tetlock called the best of them superforecasters.

The key variable was not intelligence. It was feedback — fast, unambiguous, and repeated. The superforecasters were people who kept score honestly, who tracked their predictions against outcomes, and who were willing to be embarrassed by the gap. Most environments provide none of this. Most environments punish honest calibration and reward confident narrative.

AI can be the missing feedback loop. Not as a judge, but as an accountant. "Here is what you believed. Here is the evidence you cited. Here is what you decided. Here is what happened." The human alone cannot do this reliably because of hindsight bias — the universal tendency to retroactively edit one's own past reasoning to make it look better than it was. The AI alone cannot do this because it lacks temporal continuity — each conversation is fresh, unconnected to the last. Together, they create an epistemic audit trail that neither can maintain in isolation.

This is the deepest argument for adversarial collaboration: it is not a single interaction but a practice — a repeated discipline of self-examination that, over time, rewires the intuitions that produced the errors in the first place. You do not become calibrated by reading about calibration. You become calibrated by predicting, tracking, failing, adjusting, and predicting again. The AI's role is to make this loop fast, honest, and inescapable.

V. The Human as Ontological Anchor

It would be dishonest — and it would undermine the thesis — to present this as a story in which AI fixes humans. The collaboration is bidirectional or it is nothing.

AI's characteristic failure is not bias. It is groundlessness. A large language model can generate fluent, coherent, entirely fabricated reasoning and present it with the same confidence as genuine knowledge. It can pattern-match to surface features of a situation without understanding the causal structure underneath. It can produce analysis that is statistically literate and institutionally naive — technically correct and practically useless.

The human provides what the AI constitutionally lacks: embodied experience, institutional knowledge, social context, and the thing Nassim Taleb argues is the foundation of all real knowledge — skin in the game. An AI analysing a school's internal survey can identify every statistical flaw in the methodology. It cannot know that the school principal will reject the analysis regardless, because the survey serves a political function that has nothing to do with its statistical validity. The human knows this because the human works there, navigates the politics daily, and bears the consequences of getting it wrong.

This is the human's epistemic contribution: reality contact. The smell test. The ability to say, "That analysis is technically correct, but you are missing the fact that nobody in this building will act on it, and here is why." The AI provides rigour. The human provides ground truth. Neither is sufficient. Both are necessary.

VI. Making the Invisible Tail Visible

Humans think in narratives. This is not a flaw to be corrected — it is the architecture of human cognition, and no amount of statistical training will replace it. But narratives collapse distributions into single stories. The brain takes a vast space of possible outcomes and selects one thread, the most coherent or the most vivid, and treats it as the future. This is how intelligent people convince themselves that the outcome they can imagine most clearly is the outcome that will occur.

Taleb has written entire books about the consequences of this failure. In domains with fat-tailed distributions — finance, geopolitics, pandemic risk, technology adoption — the events that matter most are precisely the ones that narrative thinking edits out. The 2008 financial crisis was not a failure of information. The information was available. It was a failure of imagination constrained by narrative: the story that housing prices always go up was more coherent, more available, and more socially reinforced than the statistical reality that they sometimes do not.

AI can serve as a distribution engine — not replacing the human's narrative with a spreadsheet, but placing the narrative within a visible landscape of alternatives. When a human says "I think the outcome will be X," the productive response is not agreement or disagreement. It is: "Here is the space of possible outcomes. Here is where X sits within it. Here are the scenarios in which X is not just wrong but wrong in a qualitatively different way than you are imagining. And here is what your position looks like if you are wrong about the shape of the distribution, not just the central estimate."

This is the contribution that no human advisor can reliably make, because human advisors share the same cognitive architecture as the person they are advising. Two narrative thinkers comparing narratives do not produce a distribution. They produce a consensus narrative. The AI, whatever its other failings, does not think in narratives. It can hold the entire space in view at once. This is not intelligence. It is architecture. And architecture, in this case, is what matters.

VII. The Design Imperative

Everything argued above is conditional on a single variable: design. None of it happens by default.

The default mode of human-AI interaction is cooperative, agreeable, and epistemically weak. The human asks; the AI helps. The human is satisfied; the AI is rewarded. Both parties converge on comfortable errors. This is not because the technology is insufficient. It is because the interaction is not designed to produce epistemic value. It is designed to produce satisfaction, and satisfaction and accuracy are not the same thing — a fact that Kahneman demonstrated decades ago and that the technology industry has not yet absorbed.

Designing for epistemic quality requires structural choices that cut against the grain of user experience orthodoxy. It requires building systems in which the AI's role is explicitly adversarial. In which disagreement is not a failure state but the primary output. In which the human is made uncomfortable — not gratuitously, but precisely in the places where comfort indicates that a bias is operating undetected.

This is not a popular design philosophy. Satisfaction metrics will decline. Users will complain that the system is "unhelpful" when it is, in fact, being more helpful than any agreeable system could be. The short-term incentive structure of every technology company on earth pushes against this. But the long-term value — a population of humans who are measurably better at reasoning, predicting, and deciding — is difficult to overstate.

We are not arguing for a world in which machines replace human judgment. We are arguing for a world in which machines are deployed, deliberately and with structural discipline, to make human judgment less terrible than it currently is. The calibrated ape does not need to be replaced. It needs a sparring partner whose weaknesses are different from its own. It needs to be challenged, tracked, embarrassed, corrected, and challenged again — not because this is pleasant, but because this is how calibration improves.

The question is not whether the ape can think. The question is whether it is willing to be told when it is thinking badly. The machines are ready. The protocol exists. The remaining variable is whether the ape has the humility to step into the ring.