The Bigger Party
My father has spent decades working on the Large Hadron Collider at CERN — he was back there again just two weeks ago. Years earlier, while he was in the thick of the search for the theoretical Higgs boson, I asked him what would happen if they found it. He said there'd be a party. I asked what would happen if they didn't. He said there'd be an even bigger one.
The best scientists — and the best thinkers — learn to want refutation over confirmation. It isn't only about certainty; it's about how much information is actually sitting in front of you. David MacKay's Information Theory, Inference, and Learning Algorithms formalizes exactly this: information is surprise, and there's no value in an input that only confirms what you already know. You can never know for sure that you're right — the next observation might overturn you. But when the evidence squarely contradicts your prediction, you learn something, with as much confidence as empirical inquiry ever offers — and it's worth more than any pile of confirmations, precisely because it was the least expected outcome in the room, not the most.
That asymmetry isn't a curiosity of physics. It's one of the first things the Computational Deliberation Lab has to understand if its larger mission is going to work: building the first deliberation calculator — a machine that maps an open-ended problem, tests what the map contains, and helps a human see what survives. Refutation is not the whole machine. It is the first method under test: the first thing such a machine has to be able to do well.
The Conversation Has Stalled
Political conversation optimizes for engagement, scientific conversation is drowning in slop, and philosophical conversation hasn't kept pace with how fast everything else is changing — and all three failures share the same root cause. Our political discourse has become two sides shouting past each other, producing claims engineered for social media reach rather than resolution, which only escalates the rhetoric instead of settling anything. Several friends of mine who are working scientists and professors have independently told me the same thing: peer review, the mechanism by which science is supposed to refute its own bad ideas, is buckling under its own volume. One recent analysis of submissions to the journal Organization Science — a single field, but a telling one — found volume up 42 percent since ChatGPT's release, with more than three in ten submitted peer reviews showing detectable signs of AI authorship. But the deeper problem isn't the machines. In my experience talking to people inside that system, serious review has increasingly become a high-ticket, socially expensive thing — scrutiny that's easiest to get if you already arrive with standing. An idea that doesn't come from the right department, journal, or name too often gets waved through or ignored, rather than actually tested.
Meanwhile philosophy — the discipline whose entire job is to clarify what we should believe and do — hasn't scaled to meet a civilization that's changing faster than at any point in its history. We're generating plenty of ideas. What we've lost is the capacity to kill the bad ones fast enough for it to matter: our conversation hasn't scaled to meet the sheer volume of voices now in it, and the incentive structures that used to force real scrutiny have collapsed under that pressure. The result isn't just slower deliberation — in a real sense, there's no effective conversation happening at all, let alone the capacity to kill a bad idea, which means the actual work is putting the conversation back on track at the scale this moment demands. I was told as a kid that the only way to find one good idea is to find ten. The lesson wasn't about creativity. It was about being prolific enough to survive the pruning that reality was always going to do anyway. What CDL is betting on is that same lesson, scaled roughly a hundredfold: generate many more good ideas, and build a way to refute the weak ones quickly, cheaply, and with the consequences kept small — instead of learning the hard way, by living through the failures ourselves.
Refutation Is the Opposite of Limiting
Being told you're wrong doesn't close off your options — being told you're right does. Karl Popper made the underlying point cleanly: "It is easy to obtain confirmations, or verifications, for nearly every theory — if we look for confirmations." Confirmation is cheap and always available unless it's the risky kind Popper meant — the kind that could have failed and didn't — which is why most of what passes for confirmation is weak evidence at best. For centuries, every swan Europeans had ever seen was white, and that was treated as confirmation that all swans were white — until Dutch sailors began reporting black ones from Western Australia in the 1600s, culminating in Willem de Vlamingh's well-documented 1697 expedition up the Swan River — though it took the better part of a century for the news to fully sink in. Centuries of confirmation, undone by one bird.
This is where the psychology runs backward from how we expect it to. A refutation feels like a slap, but it doesn't tell you what to do — it only tells you what not to do, which leaves every other path still open. A confirmation feels good, but it does something quietly coercive: it tells you to stop looking, commit to the one path you were on, and give up the rest. The feedback that actually limits you is "yes, proceed." The feedback that keeps you maximally free is "no, not that." Thinkers with real imagination and ambition learn to want to be told no as often and as cheaply as possible — not because it's comfortable, but because it's the only feedback that doesn't narrow the field. That's also why CDL isn't building toward consensus. Consensus is tactical — useful, but shallow, and it doesn't catalyze imagination. Refutation is the deeper lever, and it's the one that's been neglected.
But refutation only matters if it leaves something behind. A contradiction that happens inside a conversation and disappears is useful for the moment, then gone. The artifact worth building is the record: what was considered, what was attacked and killed and on what evidence, what survived, and what nobody examined at all. That is the tested map.
A verdict without a map is a verdict without a trial.
Inspiration Is the Scarce Resource
In an age of cheap, abundant computation, the truly scarce resource isn't compute, or even information — it's inspiration. Good ideas still have to come from somewhere, and no amount of parallel API calls generates the spark itself. What we can do is stop wasting that spark once it exists: offload everything laborious around it, starting with refutation, so an inspired person isn't the one burning months finding out whether their own idea survives contact with reality. If good ideas and bad ideas alike get killed quickly and cheaply, inspired people spend their limited attention only on what's actually still standing — which means that scarce resource lands where it's likely to survive, instead of being spent proving what a machine could have told them in an afternoon.
The Real Baseline
The obvious objection is not a distraction. Why not just tell a strong model to push back?
Maybe that is enough. If one strong model, properly instructed, can test an open-ended problem as well as anything more elaborate, that is worth proving. The lab is not committed to complexity. It is committed to finding out what it actually takes to make open-ended reasoning inspectable, measurable, and scalable.
Large language models already give us reason to be careful with their judgments. A 2025 study in PNAS found that large language models don't merely inherit our cognitive biases in decision-making — in some contexts they amplify them. These are not neutral computers whose outputs can simply be trusted the way we trust a calculator. Ask two or three different models the same question with the same context, and you'll often get different recommendations. Picking the one that's supposedly "smarter" doesn't resolve anything, because it doesn't tell you which answer is actually right.
For genuinely high-stakes decisions, a model's confident answer has to be treated as a hypothesis, not a verdict — which is exactly the discipline that intelligence analysts had to formalize when the cost of being confidently wrong was catastrophic. The CIA's own training manual on the subject, Richards Heuer's Psychology of Intelligence Analysis, is built around that same idea: structured techniques for testing a judgment, not just trusting the analyst's confidence in it. CDL is drawing on that same discipline and applying it to machine-generated judgment.
The fair distinction is not that one architecture has already been shown to reason better than another. It has not. The distinction is about what comes out. An instruction to push back produces disagreement inside a conversation; it leaves no durable artifact you can inspect for what it considered, what it declined to attack, or what it never looked at. A tested map is an artifact. Whether building that artifact needs more machinery than a strong model instructed to argue back is exactly the question.
The Machinery That Makes This Possible Now
Refuting an idea properly used to require a research team's week; now it costs a coffee and a coffee break, and that changes what's actually worth attempting. Four shifts in model capability, each unglamorous on its own, add up to a new opening.
Depth. Reasoning now unfolds instead of just completing. Models can hold an entire argument, its evidence, and their own prior reasoning in view across contexts that have gone from a few thousand tokens to millions, instead of losing the thread after a few steps.
Breadth. Scrutiny has become nearly free. Cost per token has fallen by orders of magnitude while capability has risen, and dozens or hundreds of independent calls can run in parallel — scrutinizing one claim from every angle is now a computational non-event, not a staffing problem an intelligence agency would need to solve.
Groundedness. Reasoning can close the loop with the world. Reliable tool use lets a system fetch and incorporate current, dated, sourced evidence, instead of reasoning only from frozen training data.
Composability. The pieces can finally be chained. Models follow narrow instructions reliably enough, and produce structured output reliably enough, that it is now plausible to try turning one large, unreliable guess into many smaller claims that can be checked, attacked, and recorded separately.
The closing piece is the one that actually explains why this is a new attempt rather than a faster version of an old idea: a single general-purpose model now generates credible claims across law, business, science, and philosophy without domain-specific tuning — credible enough to be worth refuting, which is a different bar than credible enough to be believed. That's the difference between "an expert system per domain" — the narrower architecture AI research spent decades on, with mixed and mostly disappointing results — and one architecture that might generalize. It is what makes a refutation engine worth attempting once, instead of assuming every field needs a separate machine. Whether that architecture earns its cost is not settled. It is the open question the project exists to answer.
Civilization needs to increase the speed of its deliberation now — and for the first time, doing that at the pace and scale this moment demands is a computational problem, not only a social or philosophical one. Refutation, industrialized, may be buildable. What that actually looks like in practice, whether we can really build it, and what it takes for the resulting maps to deserve anyone's trust: that's the open question worth spending the rest of this project answering.