Latest / Collective Intelligence: Why groups get smart, strange and stuck / Condorcet's Terrifying Assumption
Transcript
- 0:00Hello, I'm Niccolò Pescetelli and this is Collective Intelligence, why groups get smart,
- 0:05strange and stuck. In the autumn of 1793, there is a man hiding in a boarding house on the Rue
- 0:12Servandoni in Paris, about 200 metres from the Church of Saint-Sulpice. He does not go out,
- 0:19he has been condemned in his absence, and there is a decree with his name on it. And if he is
- 0:25recognised in the street, he will be dead inside the week. The woman who owns the house, Madame
- 0:31Vernet, has agreed to keep him, which is an act of considerable courage given what happens to people
- 0:37who shelter the proscribed. So he sits in a room he cannot leave, and he writes. And what he writes
- 0:45in that room, while a state he helped to build is hunting him, is one of the most staggeringly
- 0:51optimistic documents any human being has ever produced. It is a sketch of the progress of the
- 0:58human mind, and he argues that there is no limit, that errors and superstition and tyranny are
- 1:05temporary, that education and the free circulation of knowledge will go on improving us indefinitely,
- 1:12that the species is perfectible. He is writing this while listening for footsteps on the stairs.
- 1:20His name is Marie-Jean-Antoine Nicolas de Caritat, Marquis de Condorcet. He is a mathematician,
- 1:26one of the very best in Europe. He was permanent secretary of the Academy of Sciences. He was,
- 1:33before all this, a genuine aristocrat who decided that aristocracy was indefensible. He argued for the
- 1:41abolition of slavery. He argued for the rights of Protestants and Jews. He argued in 1790,
- 1:48when essentially nobody was arguing this, that women should have the vote, on the straightforward grounds
- 1:55that either rights come from being a rational being, or they do not. And if they do, then half of the
- 2:03population has been robbed. And eight years before he went into that room, he published a book, which is
- 2:11the reason why I'm talking about him now. It is called, roughly, an essay on applying analysis
- 2:17to the probability of decisions, reached by a plurality of votes. He came out in 1785 from the Royal Printing House,
- 2:25four years before the revolution. In it, he does something nobody has done. He takes the question,
- 2:33should we let the majority decide? And he treats it as a mathematics problem. Not a philosophy
- 2:39problem. Not a question about rights or legitimacy or the consent of the governed. But a calculation. He
- 2:47asks, under what conditions does asking a lot of people and counting the votes actually get you the
- 2:53right answer. And it produces a result which, for 200 years, has been the single strongest argument
- 3:01anyone has for democracy. It is also, and this is the thing I want to spend most of the episode on,
- 3:08one of the most alarming results in the whole of social science. Because the same piece of mathematics
- 3:14that says, a large public is almost certain to be right, also says, with exactly the same confidence,
- 3:22under conditions that are barely different, that a large public is almost certain to be wrong.
- 3:29And the quantity that decides which of those two worlds you are living in is a number that nobody
- 3:35has ever been able to measure. Five months after he finishes writing in that room, he is dead in a cell.
- 3:44Okay, let's do the mathematics first, and then we will come back to the cell.
- 3:50To understand why a mathematician was writing about voting at all, you have to understand what the
- 3:55enlightenment was, or, like, thought mathematics was for. Condorcet opens the 1785 book by talking
- 4:03about a dead friend. He does not name him in the sentence, but he means Turgo, the economist and
- 4:10reforming minister who had been his mentor and who had died in 1781. And what he says about him is that
- 4:18Turgot had been persuaded of something: that the truths of the moral and political sciences are capable of the
- 4:27the same certainty as the truth of the physical sciences. That a claim about how a society should
- 4:36be arranged could, in principle, be as solid as a claim about where Jupiter will be next Thursday.
- 4:43Now, sit with how strange and how enormous that is. Astronomy in the 18th century was the miracle
- 4:50science. Newton had taken the movements of the heavens, which had been the property of gods and omens for
- 4:58the whole of human history, and reduced them to a few lines of algebra that worked. You could predict
- 5:05an eclipse. You could predict it to the minute, decades out, and be right. And a certain kind of mind looked
- 5:15at that and thought: why should this stop at the sky? Why should the motion of a planet be calculable
- 5:21and the arrangement of a state not be? We do not have a science of moral things yet, but there is
- 5:28no reason in principle that we cannot have one. We just have not done the work. That project has a name
- 5:36in French: mathématique sociale, social mathematics, and Condorcet is its most serious advocate. In 1785,
- 5:47the essay is an installment of this mathématique sociale. He's not writing about juries because he's
- 5:55interested in juries. He's writing about juries because a jury is the simplest possible case of the thing he
- 6:02actually cares about, which is whether a society can reason. It is worth knowing who was making this
- 6:10argument because the biography is not just context here. It explains why he cared. He was born in 1743 in
- 6:20Picardy into the minor nobility, and he was a prodigy. By his mid-twenties, he was in the Academy of
- 6:27Sciences. D'Alembert, one of the editors of the great encyclopedia, took him up. And he could have
- 6:35had that life, you know, the life that came with that: silence, pensions, a reputation, an easy path,
- 6:43and instead he kept using it on things that could only cost him. In 1781, eight years before the
- 6:52Revolution, he published An Attack on Slavery. He published it in Switzerland, and he published it
- 6:59under a false name, a fictious Swiss pastor called Joachim Schwartz, which tells you what he thought he
- 7:08would do to do to him at home. The argument is not economic, and it is not sentimental. It is that
- 7:14black people have the same mind, the same reason, the same virtues as white people, and that owning
- 7:22them is therefore a violation of right, full stop. He wanted emancipation to be gradual, which is, you know,
- 7:30a real limitation, and I'm not going to paper over it. But he was arguing that the thing was indefensible
- 7:37in principle, in 1781, in a France that was getting extremely rich from it. In 1790, he wrote an essay
- 7:46arguing that women should have full political rights, not improved conditions, but the votes and citizenship
- 7:55on the same terms as men. His reasoning is almost brutally simple: rights follow from being a creature
- 8:04with reason and a moral sense. Women are such creatures, therefore, excluding them is not a
- 8:12tradition or a convention, but it is a straightforward breach of the rights the Revolution claimed to be
- 8:19founded on. Almost nobody was saying this. He said it in print under his own name. And he married Sophie
- 8:28de Grouchy, who translated Adam Smith and ran one of the serious intellectual salons in Paris.
- 8:36And who is the reason? We have the manuscript I mentioned at the start. So, a man who believed that
- 8:44human reason was general, that it did not come in grades according to race or sex, and that a society
- 8:51organised around it would keep improving. Now, you can see why the question of whether a group of
- 8:58ordinary people can reason their way to the truth is not an abstract exercise for him. It is the main claim
- 9:06of his entire life. And there was an argument going on that he wanted to win. The other side of it was
- 9:15Rousseau. Now, Rousseau had an idea called the general will, which is not the same as what, you know,
- 9:21the majority happens to want. The general will is what the people would want if they were thinking
- 9:28about the common good, rather than their own interests. It is a real thing out there, discoverable,
- 9:37and it is the only legitimate basis for law. Which sounds wonderful, until you ask the obvious
- 9:46question: who gets to say what the general will is? And Rousseau's answer is a figure he calls the
- 9:53legislator, a kind of super wise outsider who perceives the common good and explains it to a
- 9:59people who cannot quite see it themselves. And I want you to notice what a dangerous piece of machinery
- 10:05that is, because Lex Paulson, in his history of collective intelligence, puts it blindly:
- 10:11this is a mega technocrat, and the idea has underwritten authoritarian projects ever since.
- 10:18If the general will is a fact, and you are the one who knows it, then people who disagree with you are
- 10:25not opponents. They are just wrong. They are obstacles to the truth, and you can do things to
- 10:32obstacles that you cannot do to opponents. Now, Condorcet wants the same destination, but refuses that
- 10:39approach. He agrees with Rousseau that collective reason is real, and it is the legitimate basis of
- 10:47authority. But he will not accept an oracle. What he wants is a procedure, something mechanical, public,
- 10:56checkable, that produces the right answer without anybody having to be trusted to know it in advance.
- 11:03And because he is a mathematician, when he says procedure, what it really means is
- 11:09something you can calculate. So here is the question he sets himself, and it is worth stating precisely,
- 11:15because, you know, the mathematics, the precision is the whole point. Now, suppose you have a group of
- 11:21people, suppose there is a question with a genuine right answer, and they each give an opinion,
- 11:30and you go with the majority. Now, the question is: how likely is the majority to be right?
- 11:36Not, you know, is it legitimate, not do they have the right to decide, but how likely are they to be
- 11:43correct? Now, nobody had asked it that way before. Okay, so let's build it from the bottom up, because
- 11:53the whole thing rests on one small fact that you can check on your fingers. Picture three people deciding
- 12:00something. So it's a jury of three. The question has a right answer, and they do not know it.
- 12:09But they are not, you know, guessing blindly either. Say each of them, on their own, gets it right 60%
- 12:15of the time. Better than a coin, but you know, not much better. Four times in ten, each one of them is wrong.
- 12:24If you picked one of them at random and let them decide, you know, by themselves, you would be right 60%
- 12:31of the time. So that is your baseline. Now, let all three vote and take the majority. Now, what happens?
- 12:40For the majority to be right, you need at least two of the three to be right. Either all three or exactly two.
- 12:50And for the majority to be wrong, you need at least two of them to be wrong at the same time.
- 12:55And here is the fact that everything else in the episode
- 13:00grows out of two people being wrong at the same time is rarer than one person being wrong.
- 13:08Each of them is wrong four times in ten. Both of them being wrong together, if they are not influencing each other, happens 60 times in a hundred.
- 13:19Being wrong is an event, and events stack up against each other.
- 13:24Now, do the arithmetic, and the three-person jury is right 64.8% of the time. So almost 65%.
- 13:35That is not much. If I had opened with that number, you know, you would have probably shrugged.
- 13:40But look at what just happened, because the direction of change matters more than the size of that step.
- 13:47Three people, each individually, let's say mediocre, arranged into a majority, are better than any one of them.
- 13:58Not because they talked, not because they pulled their wisdom, but purely because for them to fail collectively,
- 14:06more than one of them has to fail at the same moment, and that is harder.
- 14:13Now, at people, five jurors, still 60% accurate individually, collectively, they are 68% accurate.
- 14:25Nine jurors, 73.
- 14:27Eleven jurors, 75.
- 14:30In 101 jurors, each of them individually, no better than, you know, slightly informed amateur, each of them wronged two times in five.
- 14:44The majority is right 98% of the time.
- 14:48If you had a thousand and one jurors, it rounds to certainty, 100%.
- 14:55That is Condorcet's result.
- 14:57Take people who are individually barely adequate.
- 15:01Do not train them.
- 15:02Do not improve them.
- 15:03Do not have them talk to each other.
- 15:06Just add more of them and count and the accuracy of the group climbs towards one.
- 15:13And it does not stop, right?
- 15:15Climbs to certainty.
- 15:16And you can feel why, if you think about what it would take for a thousand person majority to be wrong, right?
- 15:24You would need 501 separate people to make a mistake in the same direction, on the same question, and at the same time.
- 15:33Any individual mistake is common, but 500 simultaneous mistakes, all pointing the same way, all independent of each other, is a coincidence so absurd that it effectively never occurs.
- 15:48So, the majority is not smart.
- 15:50The majority is safe.
- 15:52Those are different things, and the difference is going to matter here.
- 15:57I want to slow down on the mechanism, because if you take one idea out of this episode,
- 16:02I would like it to be this one, and it is easy to nod, you know, without really seeing this.
- 16:10But picture three doctors looking at the same scan.
- 16:14Each of them is right 60% of the time, right?
- 16:17This is not slander.
- 16:18It's roughly, you know, what diagnostic accuracy on a hard case could actually look like.
- 16:24Put them in three separate rooms, so they cannot see each other's notes.
- 16:29Each makes a call.
- 16:30Each makes a call, and you go with two out of three.
- 16:34So, for that procedure to fail, it is not enough for one of them to be wrong.
- 16:39Doctors are wrong all the time, but you need two of them to be wrong on the same scan, in the same direction, and at the same moment, after having never spoken.
- 16:51And that is much rarer than one of them being wrong, not because they are better than they were, but because you have changed what counts as failure.
- 17:01You have moved the target from a mistake to a coincidence of mistakes.
- 17:08That is the entire engine.
- 17:09You know, everything else is kind of a footnote.
- 17:13Now, hold that image and make the room bigger.
- 17:17Five doctors, nine, 101.
- 17:21With 101 doctors, failure means 51 of them going wrong together independently on the same scan.
- 17:32The individual competence has not moved at all.
- 17:35Every one of them is still the same sort of fallible 60% accuracy that they were at the start.
- 17:43What has changed is what the failure condition has become now absurd.
- 17:50This is why Condorcet's result feels like a trick the first time you meet it.
- 17:56It looks as though information is being created out of nothing, but it is not, right?
- 18:01Like, the information was always there, distributed across the group.
- 18:0660% of it in each head.
- 18:10What the majority rule does is throw away the part of each judgement that is noise.
- 18:16And the more heads you have, the more thoroughly it can throw it away.
- 18:21The first refinement is about how much competence you need.
- 18:25And it is more interesting than, you know, it sounds.
- 18:29Everything above assumed 60% accuracy.
- 18:33What if your citizens are worse than that?
- 18:35What if they are barely better than a coin?
- 18:39Well, the mathematics still works.
- 18:40It just gets expensive.
- 18:44If people are right 51% of the time, 1% better than random, then a majority still converges on the truth.
- 18:52But to get to that 99% confidence, you need about thirteen and a half thousand of them.
- 19:00At 52%, you need about 3 and a half thousand.
- 19:04At 55%, 539.
- 19:08At 60%, 131.
- 19:12At 70%, 31 people.
- 19:17Now look at the shape of that.
- 19:19Competence and numbers trade against each other.
- 19:23And the exchange rate is brutal at the bottom end, right?
- 19:26A barely informed public needs to be a city.
- 19:30A moderate or moderately informed public needs a village hall.
- 19:36Which is, you know, if you squint, an argument for mass education dressed up as an argument about probability.
- 19:44And we will actually come back to that because Condorcet absolutely noticed that.
- 19:49The second refinement is the three conditions.
- 19:53And I want to name them explicitly because the popular version of these results usually mentions one of them, but tends to forget the other two.
- 20:04Condition 1: Each person has to be better than random, better than a coin flip, however slyly.
- 20:13Condition 2: They have to vote independently.
- 20:16Condition 3: Their errors cannot be connected to each other.
- 20:18Condition 3: And this one often gets dropped sometimes.
- 20:23They have to vote sincerely.
- 20:26They have to say what they actually think is true, rather than what they think will produce a good outcome.
- 20:34Or what their side needs or, I don't know, what will make them look sensible afterwards.
- 20:40So, three conditions.
- 20:41Now, the trapdoor.
- 20:44Go back to Condition 1 and try to turn it around.
- 20:50What happens if people are slightly worse than random?
- 20:54Not catastrophically bad, just under the line.
- 20:58Let's say right 49% of the time instead of 51.
- 21:03Run exactly the same arithmetic.
- 21:05And now three jurors, 48.5%, slightly worse than one person.
- 21:13A 101 jurors, 42%.
- 21:16A 1,001, 26%.
- 21:2110,001 jurors, 2%.
- 21:25A 100,000 people, 0%.
- 21:30Not, you know, approximately zero, but zero to four decimal places.
- 21:33A public of 100,000 individually almost coin-flipping citizens voting independently and sincerely is certain to reach the wrong answer.
- 21:47Now, read those two columns next to each other, because this is the thing I want to get to.
- 21:51At 51% competence, 100,000 voters are certain to be right.
- 21:56At 49% competence, 100,000 voters are certain to be wrong.
- 22:04The gap between those two worlds is two percentage points in the ability of the average person.
- 22:11And the machine does not care which side of the line you are on.
- 22:14It just amplifies with exactly the same ferocity in both directions.
- 22:20So Condorcet theorem is not a theorem about the wisdom of crowds.
- 22:25It is a theorem about amplification.
- 22:28Scale does not make a public wise.
- 22:30Scale makes a public certain.
- 22:34Whether it is certainly right or certainly wrong depends on a quantity sitting just under one half or just over it.
- 22:43And nobody in the 240 years since he published has ever managed to measure that quantity for any real electorate or any real question.
- 22:55That is the terrifying assumption.
- 22:57Not independence, which, you know, we are coming to, but competence.
- 23:03The theorem hands you a machine of enormous power and declines to tell you which way it is pointed.
- 23:12Now, the second condition, right?
- 23:14Independence.
- 23:15Everything I have just shown you depends on people's mistakes being unrelated.
- 23:20The whole engine is the improbability of 500 simultaneous errors all pointing the same way.
- 23:27Take that improbability away and the engine stops.
- 23:31So, what makes errors point the same way?
- 23:35Well, someone speaking first.
- 23:38A newspaper everybody reads.
- 23:40A phrase that catches on.
- 23:42A respected person committing to a position in public before the rest of us have decided.
- 23:48A shared education that taught everyone the same wrong rule of thumb.
- 23:54An expert consensus that is confident and mistaken.
- 23:59Anything at all that puts the same thought into more than one head.
- 24:03In other words, a society.
- 24:04In other words, a society.
- 24:07Now, let me put a number on how badly this goes.
- 24:10Because I think the number is genuinely shocking.
- 24:14And I did not expect it to be that bad before I worked it out.
- 24:19Take a simple model of shared influence.
- 24:22Most of the time, everybody reasons for themselves.
- 24:25But some of the time, with some probability, the whole public is following a single common signal.
- 24:32Maybe one story, one frame, one confident voice.
- 24:37And their answers move together.
- 24:40Call that probability the correlation.
- 24:46Set the correlation at 30%.
- 24:48Keep everybody at 60% competence, which was our sort of comfortable case.
- 24:54The one where 101 jurors go to 98%.
- 25:00101 voters now, 86.5%.
- 25:071,000 voters, 88%.
- 25:11100,000, 88%.
- 25:1410,000, 88%.
- 25:16So it basically stops.
- 25:17It reaches 88% and it stops.
- 25:20It does not creep towards that certainly like before.
- 25:24It walks up to 88% accuracy and then it sits.
- 25:30The last 12% points are not for sale.
- 25:33You cannot buy them with more people.
- 25:35Because the thing that was making extra people valuable was the independence of their errors.
- 25:41And you have spent 30% of that.
- 25:44Even a light coupling costs you.
- 25:47A 5% correlation, your ceiling is 98%.
- 25:51A 10% correlation, 96%.
- 25:55Those, you know, sound tolerable until you notice that no quantity of additional citizens will ever get you past that.
- 26:03So here is the shape of Condorcet's results seen as a whole.
- 26:08Democracy as an error-correcting machine works beautifully and it works on one condition,
- 26:15which is that the citizens are not talking to each other.
- 26:19Which brings me to the part of the story I find almost unbearable.
- 26:23Because of all the people in the 18th century, the one who did the most to build the machinery that breaks his own assumption was Condorcet.
- 26:32In 1792 he is elected to the national convention.
- 26:36He is given the job of chairing the committee to write a constitution for the new republic.
- 26:44And what he produces is, in Paulson's judgement, the most radical design for collective intelligence since ancient Athens.
- 26:52Attempted at the scale of millions of people.
- 26:55It is not a representative system, as we would understand it.
- 27:00Citizens participate directly in assemblies that nest inside each other.
- 27:04Local, then regional, then national.
- 27:07The people who sit in the higher assemblies are not representatives and are not meant to be a wiser class of people.
- 27:13He calls them mandate holders, delegates operating under constraint.
- 27:21And Paulson notes that the constraints resemble those in the Iroquois or Gikuyu systems.
- 27:32Which is not a comparison I expected to be making about a French marquise.
- 27:37And in his speech to the convention, he goes further.
- 27:42And he departs from Rousseau in a way that matters enormously for this episode.
- 27:47He insists on the value of deliberation.
- 27:51Citizens should argue publicly and form their judgements in the course of arguing.
- 27:57He insists on the free flow of knowledge through newspapers and through scientific associations.
- 28:03And echoing Aristotle, he says none of this work without a system of free public education for everybody.
- 28:12Now read that list again with now the mathematics in your hands.
- 28:18The liberation is a machine for correlating opinions.
- 28:22That is what the liberation is for.
- 28:24People come in with different views and live with more similar ones.
- 28:29And if it did not do that, we would just not bother.
- 28:35Newspapers are a machine for correlating opinions and a spectacularly efficient one.
- 28:41The whole point of a newspaper is that a large number of people receive the same account of events on the same morning.
- 28:49Scientific associations correlate opinions deliberately by design among exactly the people whose judgements carries most weight.
- 28:58And mass public education is the most powerful correlating device ever invented.
- 29:02It takes an entire generation and teaches them the same facts, the same methods, the same habits of mind.
- 29:09When they are right, they are right together.
- 29:12But when the curriculum is wrong, they are wrong together for a generation.
- 29:18Condorcet spent the last years of his life fighting for the political programme that most efficiently destroys the assumption his own most famous result depends on.
- 29:29I don't think, you know, he was being stupid.
- 29:31I think he was doing something more interesting.
- 29:35And to see what, you need a result he could not have had at the time.
- 29:40In 1992, a political scientist called Krishna Ladha published a paper in the American Journal of Political Science, whose title contains the entire problem.
- 29:49The Condorcet Jury Theorem, Free Speech and Correlated Votes.
- 29:54Free speech in that title is the thing doing the correlation.
- 29:58What Ladha shows is that the theorem does not simply die when votes correlate.
- 30:03It survives, but conditionally.
- 30:05He derives a bound.
- 30:08Given how competent people are on average, there is a maximum average correlation the system can tolerate and still do better than its individual members.
- 30:18And still improve as it grows.
- 30:21Above that line, majority rule the grades.
- 30:24Below it, you're fine.
- 30:26Which turns the whole thing from a contradiction into something more useful.
- 30:31Shared influence is not poison.
- 30:33It is a budget.
- 30:35A society can afford some amount of people listening to each other.
- 30:39And the amount it can afford depends on how good its citizens are at judging.
- 30:46Raise competence and you can afford more conversation.
- 30:49That is a coherent position and it is more or less exactly the position Condorcet's political programme takes.
- 30:56Which is educate everybody and then let them argue.
- 31:01But notice what he could not do and what we will still cannot do.
- 31:07To know whether a society is inside its budget, you would have to measure two things.
- 31:12The average competence of citizens on the question at hand and the average correlation between their errors.
- 31:19Nobody has ever measured either of those for a real electorate on a real issue.
- 31:25Condorcet's defence of deliberation is a bet that the budget is large enough.
- 31:31It is a reasonable bet.
- 31:33It is not a calculation.
- 31:35And he, of all people, would have known the difference.
- 31:41And there is a deeper objection, which I should give you at full strength, because it's the best reason to doubt the entire edifice.
- 31:50In 2008, the philosopher Franz Dietrich published a paper arguing that the premises of the jury theorem cannot be simultaneously justified.
- 32:00Not that they are empirically unlikely, but that they cannot both be defended at once.
- 32:08His argument turns on what you take probability to mean.
- 32:13The interpretation of probability under which the competence assumption looks reasonable is not the interpretation under which the independence assumption looks reasonable.
- 32:24And under most readings, independence has to be weakened.
- 32:30If he is right, then the theorem has never supported democracy in the way people have been claiming it does for two centuries.
- 32:37And the problem is not that the world fails to match the model.
- 32:41The problem is inside the model.
- 32:44I do not think that, you know, that just by itself ends the argument.
- 32:49But anyone who quotes the jury theorem at you, and I include myself, you know, an hour ago, owes you that objection.
- 32:58And I have not even mentioned the third condition, which is sincerity.
- 33:02Everybody voting for what they believe rather than for what helps their side.
- 33:09I invite you to hold that thought against any election you have watched in your lifetime.
- 33:15So, what do you do with this?
- 33:18Here is the practical residue, and it is uncomfortable, because it takes the thing we most want to believe about large-scale democracy and puts two conditions on it that nobody can check.
- 33:31Whenever you are shown a large majority and invited to be impressed by its size, there are two questions that decide whether the size means anything at all.
- 33:42How good are these people, on average, at judging this particular question, and how much of their errors are moving together?
- 33:52Notice that neither question is about the number.
- 33:55The number is the thing everybody quotes and the thing that carries the least information.
- 34:01Take a referendum.
- 34:02Now, one question, one national electorate, tens of millions of votes, and the result is treated as an oracle precisely because so many people spoke.
- 34:12But a referendum is the worst case the jury theorem has.
- 34:16The electorate has spent months inside a campaign explicitly designed to make people's errors move together.
- 34:23That is what campaigns are, right?
- 34:26Two well-funded operations spend a year pushing the correlation as high as they can get it.
- 34:33And then we treat the output as though it were the independent judgement of 40 million separate minds.
- 34:41On the mathematics, the guarantee is simply gone.
- 34:44Not just weakened, just gone, in the sense that the theorem no longer applies and it has nothing to say.
- 34:51The theorem's best case, on the other hand, looks nothing like an election.
- 34:56It looks like a lot of people who do not know each other, quietly forming a view on a question where being wrong in one direction is no more tempting than being wrong in the other.
- 35:07And then being counted.
- 35:09It looks, in fact, like the queue in Plymouth from the last episode.
- 35:14And that is not a coincidence.
- 35:15Condorcet was one of the early contributors to the law of large numbers that Galton's Ox was a demonstration of.
- 35:23Two men, a century apart, who found the same machine and wrote down the same condition for running it.
- 35:32And now think about what we have built since.
- 35:35Every correlating device Condorcet wanted has been constructed at a scale he could not have pictured.
- 35:43Universal education, achieved.
- 35:45Newspapers, superseded by something far more powerful.
- 35:49Scientific associations, global.
- 35:51And on top of all of that, in the last 15 years, an information architecture whose explicit optimization target is to show a large number of people the same thing at the same time.
- 36:05And to reward them for agreeing about it in public.
- 36:10And now, in the last few years, something with no precedent at all, a handful of very large models that a substantial fraction of the species is beginning to consult before forming a view.
- 36:23If you want a device for driving the correlation between human errors towards one, it is difficult to improve on everybody asking the same oracle.
- 36:34Now, I'm not saying that that is bad.
- 36:37Ladha's point stands.
- 36:39Correlation is a budget, not poison.
- 36:41And a better informed public can afford more of it.
- 36:45It might be that we are inside our budget.
- 36:49It might be that raising competence has outrun the rise in correlation.
- 36:53And Condorcet's bet is, you know, coming in.
- 36:56What I'm saying is that nobody knows.
- 36:59And that the most famous argument for trusting large numbers has a condition attached, which we have spent two centuries systematically undermining while quoting that result.
- 37:12Which brings me back to the man in the boarding house.
- 37:16In June 1793, the convention debates his constitution.
- 37:22It is, by any measure, the most ambitious attempt anyone had made to build his mathematics into an actual state.
- 37:30And the timing could hardly be worse.
- 37:33The spring of 1793 is when the revolution turns.
- 37:37There is a war on several fronts.
- 37:39There is a revolt in the Vendée.
- 37:40There is bread rationing.
- 37:42And the convention is splitting into factions that are beginning to understand that losing an argument there will soon mean losing your head.
- 37:53Condorcet belongs loosely to the Girondins, who are about to lose.
- 37:57On the 2nd of June, an armed crowd surrounds the convention and the leading Girondin deputies are arrested.
- 38:05The faction Condorcet is associated with has just been removed from politics by siege.
- 38:11His constitution is blocked at the verge of adoption.
- 38:15Robespierre and the Montagnards produce their own document, faster and more centralized, and push it through in a matter of weeks.
- 38:24And Condorcet, who has just watched his colleagues taken away, does the thing that a man who genuinely believes in public reason would do.
- 38:32He writes a pamphlet explaining why the new constitution is bad, and he arranges for it to be distributed.
- 38:39He is denounced within days, and a warrant goes out for his arrest.
- 38:44That is when he goes into Madame Vernet's house and starts writing about the perfectibility of the species.
- 38:52There is a version of this story where the irony is so much, I want to resist it slightly, because I do not think he was a fool.
- 39:00He had spent his life arguing that reason would win because it was reason.
- 39:05What he was watching was a revolutionary assembly, following a procedure, reaching decisions by majority, and using them to kill people, including, shortly, him.
- 39:17If you wanted a demonstration that a large group can converge with total confidence on a catastrophic answer, the terror is sitting right there in his own biography,
- 39:28being run by his own colleagues, under voting rules he had helped design.
- 39:33In March 1794, he leaves the house, apparently because he fears he is endangering the woman sheltering him.
- 39:40He walks out into a country where being unrecognised is the only thing keeping him alive.
- 39:46He is arrested at Clamart, and he is taken to a cell in Bourg-Larenne, which the revolutionaries had renamed Bourg-Egalité.
- 39:55Two days later, on the 29th of March, he is found dead on the floor.
- 40:01He is 50.
- 40:04And nobody knows what killed him.
- 40:05The most common account is poison, supplied earlier by his friend, the physician Cabanis, taken to avoid the guillotine.
- 40:13There is a competing account in which his guards killed him on instruction because the Committee of Public Safety did not want the public spectacle of executing a famous mathematician with admirers across Europe.
- 40:27And there is the plain possibility that a 50-year-old man, who had spent months in hiding and then walked a long way and slept rough,
- 40:34simply died of exhaustion and illness in a cold room.
- 40:38He was found with nothing obvious on him.
- 40:42I have seen all three asserted confidently.
- 40:46The honest position is that the cause of his death is unknown.
- 40:51And I would rather leave it unknown than hand you, you know, a tidy ending.
- 40:56The manuscript survived.
- 40:58His widow, Sophie, published it the following year.
- 41:00And the sketch for a historical picture of the progress of the human mind became one of the founding documents of the idea that things get better.
- 41:09Written by a man in hiding from a state he built who died in a cell, arguing that humanity has no ceiling.
- 41:18So here is where I want to leave you.
- 41:21Condorcet's answer to the correlation problem, though he never phrased it that way, was education.
- 41:29Make people better at judging and you can afford to let them talk to each other.
- 41:35Competence buys you conversation.
- 41:38That is the bet.
- 41:39And the whole of his politics follows from it.
- 41:42But the last two episodes have both ended in the same place, from opposite directions.
- 41:47Galton's ox worked because nobody spoke.
- 41:51Condorcet's theorem works because nobody listens.
- 41:54And every institution we actually value, every parliament and jury and committee, peer review, public argument,
- 42:03is built on the opposite conviction, that talking to each other makes us better.
- 42:10But both of those cannot be simply true, right?
- 42:13Now you might expect the next step to be an experiment.
- 42:16Somebody puts people in a room, lets them talk, and measure what happens to their judgement.
- 42:21That does eventually happen, and we will get there.
- 42:25But for most of the century after Condorcet died, almost nobody thought the question needed an experiment.
- 42:32They thought the answer was obvious, and the answer they had in mind was the opposite of his.
- 42:38In 1841, a journalist called Charles Mackay published a long and rather gleeful catalogue of collective insanity.
- 42:47Dutch tulip bulbs, trading for fortunes, the South Sea bubble, the Mississippi scheme, witch hunts, crusades.
- 42:54The title is the part that everybody remembers.
- 42:57Extraordinary popular delusions and the madness of crowds.
- 43:02And in 1895, a French writer called Gustave Le Bon went even further.
- 43:07He argued that when individuals gather in a crowd, a new thing appears.
- 43:11A single creature that is scredulous, impulsive, and capable of violence that none of its members would commit alone.
- 43:20Here is the detail that made me stop when I found it.
- 43:24When Le Bon wants an example of ordinary, decent people turned savage by a crowd,
- 43:29the example he reaches for is the French Convention.
- 43:33The same assembly that blocked Condorcet's constitution and issued the warrant and sent him into hiding.
- 43:40Le Bon describes its most ferocious members as men who, in calmer times, would have been harmless notaries and upright magistrates,
- 43:50and who went quietly back to being exactly that once the storm had passed.
- 43:56So, the man who proved that a large public converges on the truth was destroyed by the institution
- 44:03that became the textbook case for the opposite view.
- 44:07Over the next two episodes, I want to take Mackay and Le Bon seriously, one at a time.
- 44:13They are easy to dismiss, and Le Bon in particular writes with open contempt for the people he's describing.
- 44:19But I think they saw something real, and I think the mathematics of the last two episodes let us say precisely what it was.
- 44:29They were describing a crowd whose errors had stopped being independent.
- 44:33We start with Mackay and with Tulips.
- 44:36That is next time.
- 44:38Thanks for listening.
- 44:39The essay is Condorcet, 1785, and the readable part is the preliminary discourse.
- 44:44Lex Paulson's History of Collective Intelligence is the opening chapter of the Routledge Handbook of Collective Intelligence for Democracy and Governance,
- 44:52and it is open access.
- 44:54Ladha on correlated votes is in the American Journal of Political Science, 1992.
- 45:00Franz Dietrich's Objection is in Episteme in 2008, all in the notes.