“The judge understands statistics, he understands prevalence, but we’re getting him to rule against nearly twice as many people as are actually responsible,” Kamenica said.
The courtroom scene shows the power of being able to control information that is conveyed. If the prosecutor were just spewing cheap talk, he would always simply claim guilt, but such empty claims would never provide any information to a rational judge. “The ability to commit to what type of information will be generated is a powerful tool,” Kamenica said.
Now for some other applications of Bayesian persuasion:
The paper about lying politicians that came out this month is by Florian Ederer of the Yale School of Management and Weicheng Min of Yale’s economics department. Politicians would have no incentive to lie if fact-checkers caught 100 percent of their lies, the authors write, but if the probability of catching a lie is sufficiently low, a politician will compensate for the fact-checking by lying even more. The sender (in this case, a politician) “noises up the information environment,” Ederer said in an interview. (You might think that lying doesn’t fit into a Bayesian persuasion framework, but Ederer says it can fit as long as the politician “commits to sending a truthful or an untruthful message with a certain probability” that can depend on the state of the world.)
“Persuading With Anecdotes,” a working paper issued in April 2021, says that it’s rational for nonexperts to obtain their information from people who are poorly informed but have similar preferences rather than from experts “whose preferences may differ” from their own. Experts have a vast stock of anecdotes and may choose ones that steer people toward their beliefs, says the paper, which has five authors, two from Microsoft, one from the University of California, Berkeley, one from the University of Michigan and one from Princeton. That explains a lot of what you see on social media.
The triumph of the silly anecdote is what you get when talk is cheap. That is not Bayesian persuasion. In the Bayesian case, the authors find, a sender of information won’t cherry-pick anecdotes but rather “will choose an unbiased and maximally informative communication scheme.” That’s reassuring. Unfortunately, the Bayesian situation is rare.
Penélope Hernández of the University of Valencia in Spain and Zvika Neeman of Tel Aviv University in Israel wrote a 2019 paper, “How Bayesian Persuasion Can Help Reduce Illegal Parking and Other Socially Undesirable Behavior.” They assume that if the likelihood of getting a ticket is below some threshold, people will park illegally. Assuming that the budget for enforcing parking rules is fixed, they write, it makes sense to give up on enforcing the rules at certain times and places and focus the budget on others. In an interview, Neeman said drivers could get a red signal on their phones when an enforcement agent is close and either a red or a green signal when the agent is far away. Drivers would be told, in all honesty, that the red signal might be a false alarm, but it would still induce them to park legally.
Bayesian persuasion hasn’t been widely embraced by policymakers. “In practice, people are probably less than fully Bayesian rational, and certainly, probably not as Bayesian rational as assumed in this paper,” the paper by Hernández and Neeman concedes.