Knowledge is fundamentally dispersed across many people and even within individual minds, and the central problem of decision-making is how to aggregate that knowledge across people and across time.

When you ask a large group to estimate an objective quantity (like the weight of an ox or a horse, or beans in a jar), the statistical average of their independent judgments is often uncannily accurate and frequently better than any individual member, even an expert.

The Condorcet Jury Theorem shows that if each member of a group is more likely than not to get the right answer, the probability that the majority gets it right approaches 100% as group size grows; but if the group suffers a systematic bias making them more likely wrong than right, that probability falls to 0% as the group grows.

Crowd wisdom fails when there is a systematic bias—for example, when people systematically underestimate the number of times the Supreme Court invalidated laws because salient examples are few, or systematically overestimate the share of the workforce earning minimum wage.

Crowd-wisdom parlor examples (ox weight, jellybeans) are problematic test cases because they involve unimportant questions and offer no incentive to get it right, whereas the harder, high-stakes questions are matters of judgment where the answer is difficult to know even after the fact.

The important point is the mechanism, not the example: when a large group of people are mostly more likely than not to get it right, statistical aggregation outperforms picking the single expert you trust most, which is why statistical collections of experts typically outperform individual experts in forecasting business and economic outcomes.

Prediction markets improve on surveys by giving participants an incentive to enter only when they have confidence in their judgment, and by allowing prices to dynamically self-correct over time; at Google, prediction-market prices function as accurate probabilities (events priced at 80% occur about 80% of the time).

There is strikingly little difference in predictive power between high-incentive real-money prediction markets and low-incentive 'virtual' markets, suggesting that reputation, honor, and self-conception motivate participants to get it right even without large financial stakes.

A survey or prediction market inside the CIA on whether Iraq had weapons of mass destruction would likely have produced a less confident answer than the official government position, because there was substantial dispersed information within the agency about the difficulty of that question.

After deliberation, people typically end up holding a more extreme version of what they thought before the discussion began (group polarization), and this is close to an iron law of social interactions, demonstrated in controlled experiments such as the Colorado study where Boulder liberals became more pro-affirmative-action and Colorado Springs conservatives more anti.

Deliberating groups often amplify cognitive errors rather than just propagate them—groups can be worse victims of biases like the availability heuristic than individuals—so 'garbage in' leads to even more 'garbage out.'

Deliberating groups tend to emphasize information that everyone already shares while neglecting uniquely-held information that only one member possesses, producing 'hidden profiles' where crucial private knowledge is downplayed and the group marches in the direction indicated by common knowledge.

Federal judicial voting demonstrates group polarization with real (non-experimental) data: Democratic appointees vote much more liberally on all-Democratic three-judge panels than when a Republican is present, and Republicans likewise vote far more conservatively on all-Republican panels, with each becoming more moderate when an opposing appointee is present.

A Republican appointee sitting with two Democratic appointees votes about as liberally as a Democratic appointee sitting with two Republicans, making the two essentially indistinguishable—a striking testimony to the power of social influence over ideology.

The judicial polarization findings might be partly explained by strategic incentives: on a unanimous-party panel the outcome is known, so there is little cost to voting with colleagues, whereas on a mixed panel where the outcome is uncertain a judge has incentive to think harder about what the law actually requires.

The best-functioning companies, boards, and investment clubs deliberately cultivate dissent and disagreement and take institutional steps to counteract group polarization, demonstrating that deliberation can be structured to work better rather than being abandoned.

Deliberation works well on 'Eureka' problems where, once the answer is announced, everyone immediately recognizes it as correct—which is why large groups can solve crossword puzzles better than small groups.

The best way to structure a deliberating group is to have members state their views anonymously and privately before deliberation, so the deliberative process cannot strip members of the independent conclusions they held beforehand.

The Supreme Court's diversity (nine justices) and the strength of will of its members serve as a safeguard against the polarization, hidden-profile, and cascade problems that afflict other government agencies and businesses.

The Supreme Court confirmation process, due to its inherently combative nature, has ruled out many of the most creative and interesting legal thinkers on both left and right who would enhance intellectual diversity on the court.

Hayek's 1945 article 'The Use of Knowledge in Society' established that information is profoundly dispersed—each person holds a bit of it—and that the price system is a marvel because it aggregates this dispersed information into a single signal, an insight whose implications have yet to be adequately exploited.

Sunstein's behavioral critique of Hayek is that the same problems that make deliberation go wrong can make the price system go wrong—producing fads, fashions, and mispricing for non-trivial periods (as Shiller's 'Irrational Exuberance' argues)—but markets have correctives that make such errors less enduring, and prediction markets specifically show no empirical evidence of such failures.

Sunstein has placed Hayek into a 'Samuelsonian Arrow-Debreu box' that Hayek explicitly rejected: Hayek never claimed prices were right or that a correct price was knowable; rather he was concerned with how society solves the problem of order and responds to dispersed change in supply and demand while allowing people to keep making decisions.

There is no evidence of prediction markets seriously failing in domains where dispersed information exists; attempts to inflate prices for candidates like Pat Buchanan and Hillary Clinton on the Iowa Electronic Markets failed because traders saw the inflation and profitably bet against it, straightening prices out.

Government should make much greater use of prediction markets—for example the EPA on climate change and environmental problems, and the FCC on the effects of deregulating cross-ownership rules on program content—because they are the most successful mechanism for compiling dispersed information.

Gambling laws and securities laws are the first and most serious obstacles to prediction markets flourishing: companies fear that an internal market signaling trouble would trigger disclosure obligations, and substantial deregulation is needed to allow beneficial experiments.

Open source software succeeds by exposing source code so a culture of hackers can improve it, but it will not outperform proprietary software in all domains because the economic incentive of ownership can be a real spur; its success surprises economists who underestimated non-monetary incentives like love, passion, and fascination.

Wikipedia achieves surprising accuracy—comparable to Encyclopedia Britannica in a contested study—because it aggregates dispersed information and its contributors outnumber and outwork the vandals.

The CIA's Intellipedia, based on Wikipedia software, attempts to capture the Hayekian process of aggregating dispersed intelligence information without monetary incentives, and works as long as knowledgeable people contribute their bits of information and the system is protected from vandals.

When a small minority does the bulk of the work in a collaborative aggregation system (as with Wikipedia or the Oxford English Dictionary's volunteer contributors), the system still works because the larger 'tail' of participants keeps an eye on it and corrects egregious errors.

Diversity in broadcast programming is a reflection of consumer demand rather than of who owns the media, because owners (even white Republicans) have a lot to gain from serving an audience such as a Hispanic one, so they will locate and serve that audience.

Welcome to EconTalk and details about the website and email address.