Prediction Markets Are Becoming Feeds With Odds Attached

Meta is reportedly building a standalone prediction-market app called Arena, and Polymarket has just crossed a startling commercial threshold. According to NPR’s report on internal Meta documents¹, Arena would give users a daily allotment of virtual “play money” to bet on yes-or-no questions about future events, use Meta’s Llama AI model to generate questions from trending topics, recommend markets to users, and resolve outcomes in near real time. Two days later, Reuters reported² that Polymarket’s annualized revenue had surpassed $1 billion, meaning its current revenue pace, if extended over a full year, would exceed that figure.

That pairing matters. A prediction market is a market where people buy contracts, or points in a play-money version, that pay off depending on what happens later. If a contract saying “yes, this will happen” trades at 70 cents on a dollar payout, the market is roughly saying the event has a 70 percent chance. The good version of this technology turns scattered beliefs into a public probability. The bad version turns the news into a casino lobby with better charts.

I think the bad version is the more likely default when prediction markets move inside mainstream social platforms. Not because prediction markets are useless. They are not. The problem is that social distribution changes the product. A market embedded in an attention platform is not just a forecast. It is a feed object, a ranking game, a notification hook, a data source, and possibly a bridge to real-money trading.

The strongest case for prediction markets is real. The old academic evidence is not marketing fluff. In a long-run study of the Iowa Electronic Markets, researchers found that its presidential vote-share markets were closer than contemporaneous polls to the final two-party vote split 74 percent of the time across 964 polls in five elections since 1988, and its election-eve errors were small by polling standards, according to the Journal of Forecasting paper⁵. In science forecasting, a 2021 analysis of replication projects found that prediction markets correctly classified direct replication outcomes with 73 percent accuracy across 103 findings, according to the arXiv version of the study⁶. Manifold, a play-money prediction site, publishes a calibration page showing how its resolved binary markets track stated probabilities, though its own methodology is trade-weighted and based on resolved markets with at least 15 traders, according to Manifold’s calibration documentation⁷.

That record should make critics cautious. Markets can discipline punditry. A pundit can say “this candidate is doomed” and later pretend the word “doomed” meant something subtle. A market price is harder to wriggle out of. It has a timestamp, a probability, and a resolution. In a healthier information ecosystem, I would want more claims converted into scoreable forecasts.

But Meta is not reportedly building a public-interest forecasting utility for epidemiologists and budget analysts. It is reportedly building a consumer app in which AI can manufacture endless bettable questions from whatever is trending, personalize recommendations, and resolve markets automatically, according to NPR¹. That is the crucial shift. The market no longer merely aggregates attention. It can be built to create it.

Play money helps, but it does not solve the platform problem. “Play money” means users are not wagering dollars that can be cashed out directly, at least in the initial version NPR described. That lowers the risk of direct financial ruin and may keep the product outside some gambling and derivatives rules. But variable rewards, leaderboards, streaks, social comparison, and personalized prompts do not require cash to work. Many addictive consumer apps already run on points, likes, badges, and status. Meta does not need to let a user win $500 for the product to learn what scares, thrills, enrages, or obsesses that user.

The legal boundary is messier than the phrase “play money” suggests. In U.S. financial law, many real-money event contracts are treated as derivatives, financial contracts whose value depends on an underlying event or commodity. The Commodity Futures Trading Commission, or CFTC, says event contracts are often yes-or-no derivatives with fixed payouts, and that regulated exchanges must police manipulation and insider trading, according to the agency’s prediction markets explainer³. In 2022, the CFTC fined Polymarket $1.4 million for offering off-exchange event-based binary options and failing to register as a designated contract market or swap execution facility, according to the CFTC enforcement order⁴.

That does not mean Arena would be illegal. It means the category is built around classification games. Real-money prediction markets say they are federally regulated event-contract exchanges rather than sportsbooks. State gambling officials often see sports and election contracts as betting by another name. Sweepstakes try to avoid gambling classification by removing purchase requirements, a distinction that surfaced in an Associated Press story about sweepstakes law¹¹. Fantasy contests have long leaned on claims that skill, not chance, defines the product. Play-money markets add another wrapper: no cash stake, no direct payout, lower legal heat. The business question is whether that wrapper stays put.

NPR reported that gaming lawyer Daniel Wallach described the prediction-market landscape as “legal limbo” and pointed to more than 30 pending lawsuits over the legality of these markets; the same report said Meta’s play-money launch could give it time while the legal landscape settles and while it considers licensing paths. That is exactly why I do not read play money as a philosophical commitment. I read it as a staging area.

Polymarket shows where the money is. Reuters reported that retail users are trading contracts tied to elections, sports, financial events, the FIFA World Cup, and whether the Strait of Hormuz will be closed, and that the revenue milestone came about six weeks after Polymarket rolled out access to its U.S. exchange. Prediction markets generally make money through contract fees rather than by taking the opposite side of every user’s bet, as the Associated Press explained⁸. That is cleaner than a casino’s house model in one respect, but the incentive is still volume. More markets, more trading, more reasons to check back.

This is where accuracy becomes morally slippery. A market can be well calibrated and still be a lousy civic product. If users trade on elections, disasters, wars, celebrity scandals, sports injuries, and rumors, the platform can produce useful probabilities while also training people to experience public life as a portfolio of dopamine jolts. “Will the Strait of Hormuz close?” is a legitimate geopolitical question. It is also an attention magnet when served as a live, tradable prompt to millions of bored users.

The early harm signals are not conclusive, but they point in the wrong direction. The Associated Press reported⁸ that two people with gambling addictions said they relapsed on prediction-market platforms after self-excluding from sportsbooks, and treatment experts described similar cycles of anticipation, action, and reaction. The same AP report noted that prediction markets are often available to 18-to-20-year-olds even though legal sports betting is limited to 21 and older in most states where it is allowed. Axios has reported⁹ that prediction-market brands and affiliated accounts have amplified false, misleading, or context-poor claims about politics and crises on social media. Axios also reported¹⁰ that Kalshi suspended users for alleged violations of its policy against trading on nonpublic information, showing how quickly “who knows what before everyone else” becomes part of the market’s integrity problem.

The best counterargument is that today’s social feeds are already awful forecasting environments. They reward vague punditry, viral outrage, and hindsight. A probability attached to a future event is more honest than a pundit’s theatrical certainty. I buy that. If Meta built a bounded civic-forecasting layer with no cash-out path, no real-money referrals, public calibration dashboards, clear human appeal for resolutions, bans on tragedy markets, age gates, independent audits, and strict limits on notifications, I would be much more open to it.

But that is not the product shape being signaled. The reported design uses AI-generated questions, personalized market recommendations, and near-real-time AI resolution. Those choices solve a business problem, not just an information problem. Manual curation is expensive. AI makes markets infinite. Infinite markets are not what a civic forecasting tool needs most. They are what an engagement engine needs most.

So the question I would ask Meta is not whether play-money prediction markets can ever be informative. They can. The question is whether Meta will accept the boring constraints that make them informative rather than compulsive. No markets on deaths, attacks, or active crises. No minors. No real-money migration. No personalized outrage prompts. Public calibration by category. Human-readable resolution rules before trading begins. Independent access for researchers measuring gambling-like behavior, misinformation belief, and session-time effects.

Without those constraints, mainstream prediction markets will not replace the worst parts of social media. They will quantify them. The odds may be smarter than the pundits, but the feed will still be the feed, now with a price attached to every anxiety.