If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
The most selectively believed-in verse in the conservative catechism is the idea that "incentives matter."
Sure, "incentives matter" if you're seeking healthcare. That's why you're nibbled to death by co-pays and deductibles – if you could get healthcare whenever you felt like it, you might get too much healthcare. "Incentives matter," so we have to make sure that you only seek care when you really need it:
But rich people don't need to be disciplined by incentives. They can get no-bid contracts with Uncle Sucker without being tempted to rip off the USA. They can force their workers into nondisparagement clauses without being tempted to act like a colossal asshole, secure in the knowledge that they can sue workers who tattle on them. They can force their workers into noncompete clauses without being tempted to underpay and abuse their workers, secure in the knowledge that they can sue workers who take their labor elsewhere. They can force their workers into binding arbitration clauses without being tempted into maiming or killing them, secure in the knowledge that the workers can't sue them.
So incentives matter…when you're fucking over working people. But incentives don't matter, when you're gilding the Epstein class's lilies.
But incentives really do matter. That's the premise of Goodhart's law: "When a measure becomes a target, it ceases to be a good measure." This comes up all the time. Google got its start by observing that people who made websites linked to other websites that they found important or worthy or informative. With this insight, Google repurposed the academic practice of "citation analysis" to predict which pages on the internet were most authoritative, calling it Pagerank.
Google Search, powered by Pagerank, was vastly superior to any search engine in history. But as soon as Google became the most popular search engine, people started making links to bad websites – sites filled with spam and malware and junk – in order to game the results. The metric – inbound links – became a target – get inbound links – and stopped being a useful metric.
There is something quite wonderful and life affirming about the idea of Pagerank: the idea that people are, on average, pretty good at figuring out what's good. Rather than taking Yahoo's approach of having experts rank and categorize every website on earth, Google trusted "the wisdom of crowds" and it worked (until they created an incentive to subvert it).
"The wisdom of crowds" was in the air in those days. James Surowiecki had a massive bestseller with that title in 2004, expounding on the idea that people were, in aggregate, good at figuring stuff out:
Surowiecki's book revolved around a famous anecdote from 1906, when 800 people at the Plymouth county fair were invited to guess at the weight of a slaughtered and dressed ox. Statistician (and eugenicist creep) Francis Galton noted that the average guess of 1207 lbs was within 1% of the actual weight, 1198 lbs. This turns out to be a repeatable phenomenon: if you get a lot of people – non-experts, experts, people paying close attention, people who barely think about it – to guess about something, the average is surprisingly accurate. Importantly, it's often more accurate than the best guess of experts.
This idea of the wisdom of crowds inspired a lot of 2000s-era internet projects. Some of them (Yahoo Answers) were pretty bad. Others (Wikipedia) were astounding. Of course, economists observed that "the wisdom of crowds" sounds a lot like the idea of "price discovery" – the idea that markets are a way of processing widely diffused information about desires and capacity in order to derive and emit signals about what should be produced.
Economists have long spoken of future events being "priced in" to markets – for example, the price of oil today reflects more than the diminished supply resulting from Trump's military blunders, it also reflects "the market's" belief that oil production capacity will be disrupted for a long time to come. Add up all the different buyers' and sellers' guesses about the future of oil (incorporating diffuse knowledge about damage to infrastructure, capacity to rebuild, and intentions of the actors) and (we're told) we'll get a number that accurately reflects the real situation.
And, unlike Pagerank, this number can't be manipulated by flooding the system with spurious, self-serving inputs. If you want to move this price, you have to buy or sell something, which costs money. And because the market is "deep" (with a lot of participants), the sums you'd have to inject into the system to alter its consensus is incredibly large – more than you could possibly stand to make by manipulating the price itself. Incentives matter.
Put "markets," "the wisdom of crowds" and "incentives matter" together and you get "prediction markets." Just create a market where people can bet real money on the outcomes of events and you can recreate Galton's ox-guessing miracle, but for everything – how much new solar capacity will come online in Pakistan next year; the likelihood that the Toronto Transit Commission will finish the Ontario Line this year; whether a biotech firm will ship an AIDS vaccine before 2040.
This is where Goodhart's law comes in. The idea that betting markets improve the wisdom of crowds because participants have "skin in the game" only works if the cheapest way to win a bet is to be right. If it's cheaper to win by cheating, well, "incentives matter," and you'll get cheating.
Any prediction market needs an "oracle" – a decisive source of truth about how an event turned out. "How much new solar capacity came online in Pakistan" this year sounds like an empirical question, but unless every bettor agrees to travel to Pakistan together and walk the land, counting solar panels and checking proof of their installation dates, these bettors need to agree on some third party assessor as authoritative and trust whatever they say.
Which means that the single most important factor in any prediction market is the quality of the oracle. If you let Trump be your oracle, he'll insist (on a daily basis) that his war in Iran is over, and that he had bigger crowds for his inauguration than anyone in history, and that every criminal is Somali, and on and on and on.
So you need to get someone trustworthy and diligent to serve as your oracle. But that person also has to be incorruptible, because otherwise a bettor will offer them a bribe to lie about the outcome of a bet. And if the oracle can't be bribed, they can be coerced.
That's just what's happened. Times of Israel war correspondent Emanuel Fabian didn't know that he was serving as an oracle for a bunch of degenerate gamblers on Polymarket – until he wrote a 150 word blog post that made a bunch of bettors in a $14m wager very, very angry:
The $14m was riding on a bet about when Iran would successfully strike Israel, with "success" defined as a missile getting through without being intercepted. Fabian filed a routine report that a missile had struck an open area in Jerusalem without hurting anyone. That's when the degenerate gamblers found him.
At first, they sent thinly veiled threats, demanding that Fabian revise his reporting to say that the missile had been intercepted and that the impact was just wreckage from the interception. When Fabian did not revise his article, the gamblers tracked down his messaging IDs – Whatsapp, Discord, X – and bombarded him with escalating threats. A journalistic colleague contacted Fabian with the lie that his boss wanted Fabian to change the story, then admitted that he was actually invested in the wager, and offered to split the money with Fabian.
Then, a gambler calling himself "Haim" sent Fabian a new series of blood-curdling threats, including a promise to spend at least $900,000 (the money Haim said he stood to lose) on a hit-man to kill Fabian. Haim threatened Fabian's "lovely parents" and "brothers and sisters" too. The threats continued until Fabian published his article about the threats, then Haim disappeared.
Speaking to Charlie Warzel, Fabian said that he would never be able to report the same way again, because from now on, he'd be worried that some gambler would threaten to kill him if they didn't like what he wrote:
It's sadly not unusual for journalists to receive death threats for reporting the truth, and Israel is the most dangerous country in the world to be a journalist. The IDF has murdered at least 274 journalists to date:
But those journalists are being murdered for political reasons, because someone has an ideological stake in suppressing the truth. Fabian's talking about an entirely novel – and far less predictable – threat; namely, that you will piss off someone who guessed wrong about the outcome of some arbitrary event and who thinks that they can salvage their bet by intimidating you.
Writing for Techdirt, Mike Masnick talks about the sheer perversity of this: that prediction markets, far from being a means of surfacing hidden information, have become a system for distorting information:
As Masnick says, this is no routine proof of Goodhart's law, where a metric becomes a target. In this case, participants can "put a gun to the metric's head." And of course, not every journalist is as incorruptible as Fabian – think about Fabian's colleague who offered to split the take if Fabian would lie about the missile strike. So there's plenty of incentive to publish lies – and incentives matter, right?
Now, "prediction markets" are big business and they have plenty of apologists (incentives matter). These apologists will say that the corruption is a feature, not a bug, because prediction markets will attract insiders who cheat on the bets by using their insider knowledge, and that means that looking at the moving odds of an event can help everyone else figure out what's about to happen. If military insiders who know that Trump is about to kidnap the president of Venezuela and steal its oil start laying big bets that this is going to happen, the shifting odds are a signal about a true future event.
But even if you buy this perverse argument, it doesn't offset the even more perverse effect – that prediction markets create an incentive to corrupt our best sources of information, the oracles that every prediction market absolutely requires if it is going to hope to function.
Meanwhile, Polymarket and Kalshi suck at predicting things. As Molly White points out, the predictions in the recent Illinois 2nd District Congressional race weren't just incredibly wrong, they also precisely tracked the sums flooded into the election by cryptocurrency Super PACs, who tried (unsuccessfully) to buy the race. Polymarket and Kalshi are heavily crypto-coded (the only things you can do with crypto is buy other kinds of crypto, launder money, and make wagers) so these demonic freaks flush nearly as much money into the betting markets as they do into the elections they seek to corrupt:
Prediction markets aren't good at producing information, but they're amazing at producing corruption. Polymarket and Kalshi have at last realized the unhinged fantasy of "assassination markets" – where you stochastically murder someone by putting up huge wagers at favorable odds that your target will be killed. Anyone can collect the wager by putting up a small counterwager and then bumping off the victim. But, as Protos's Cas Piancey and Mark Toon note, Polymarket and Kalshi know what side their bread is buttered on – they have banned bets on Trump's death (Trump's sons are heavily invested in both Polymarket and Kalshi):
Incentives do matter. These are the foreseeable and foreseen outcomes of prediction markets. Many science fiction writers (Charlie Stross, Ted Chiang, me, and others!) have noted that long before the current AI bubble, our society was dominated by artificial life forms: the limited liability corporation, a "slow AI" that is an immortal colony organism that uses human beings as a form of inconvenient gut flora:
Anyone who's worked with machine learning systems knows that they're prone to "reward hacking," like the ML-guided Roomba that was programmed to avoid collisions with walls and furniture as it found the quickest path around the room. The Roomba's collision sensor was on its front face, so the Roomba started moving around the room in reverse, smashing the hell out of the furnishings and walls, but never registering a hit:
Markets are absolutely capable of inducing reward hacking in participants. The metric becomes a target. You think you're betting on the outcome of an event, but what you're really betting on is what an oracle will say the outcome was. No matter what the outcome is or how robust it is against outside influence, the oracle can be influenced with a gun to the temple. Sure, we all want "number go up," but why bother increasing the thing the number measures, when it's so much easier to threaten to dismember the person who publishes the number if they don't publish a higher number?
Anya is live and ready to show you everything. Watch her strip, dance, and perform exclusive shows just for you. Interact in real-time and make your fantasies come true.
✓ Live Streaming✓ Interactive Chat✓ Private Shows✓ HD Quality✓ Free Actions
Free to watch • No registration required • HD streaming
