Meta Is Building an 'AI Prediction Market' App Called Arena — and Llama Decides What's True

Meta is jumping into "guess the future" — and the referee is an AI

Here's the deal: leaked Meta internal documents show Mark Zuckerberg told a team to build a prediction-market app called Arena. It's an app where people guess the outcomes of real-world events — but the most provocative part is elsewhere. Generating the questions, recommending them, and even ruling on "did that actually happen or not" are all handled by Meta's own AI model, Llama.

How it works: instead of wagering real money, users get a daily allotment of play money to bet on future outcomes. Llama auto-generates questions from trending topics, recommends "personalized markets" per user, and once outcomes arrive, it resolves the markets. So AI governs both what you bet on and what gets ruled true. The codenames were reportedly "Antwerp" and "FBForecast."

Why is this big? Two reasons. First, prediction markets are exploding — Kalshi and Polymarket grew the category, and some analysts see a potential trillion-dollar industry. Second, Meta enters with an entirely new card: "AI decides the truth." It raises a trust question beyond technology — "should we hand 'what is fact' to an AI?"

So today's story: what Arena is and how it runs, why Meta is entering prediction markets now, why "Llama decides truth" is contentious, and how Kalshi and Polymarket will counter. Start with the cast.

The cast — Meta, Llama, and prediction markets

First, Meta. The social empire behind Facebook, Instagram, and WhatsApp, with billions of users. The key: two of Meta's overwhelming weapons go straight into this app — a vast user base, and its own large model, Llama. Few companies have both the distribution to push a new app to hundreds of millions overnight and the AI to run it.

Next, Llama. Meta's large language model. The point is its role here isn't a sidekick. It handles question generation, personalized recommendation, and the heaviest task — "market resolution" (ruling on what happened). AI moves beyond "creating and recommending" content to sitting in the seat that "makes the final call on what's true." That's the app's most contentious heart.

Third, prediction markets themselves. Markets where people bet on "the outcome of future events," and that collective betting produces a kind of probability forecast. Kalshi and Polymarket are the flag-bearers, trading outcomes of elections, sports, and economic indicators. Meta's Arena differs by using play money instead of real money — sidestepping gambling rules while chasing mass adoption with social + AI.

Tie the three together: a social empire with billions of users (Meta) enters the market Kalshi and Polymarket built, with a play-money prediction app (Arena) where its own AI (Llama) does everything from question generation to truth resolution. That's the spine.

What was actually leaked

Words scatter, so here are the confirmed facts.

Item	Detail
Source	Meta internal documents obtained by NPR (not previously reported)
App name	Arena (codenames: Antwerp, FBForecast)
Directive	Mark Zuckerberg told a team to build it
Betting	Not real money — a daily allotment of play money
Question generation	Llama auto-generates from trending topics
Recommendation	Llama makes personalized market recommendations
Market resolution	Llama makes the final call on "did the event happen"
Competitors	Kalshi, Polymarket
Market outlook	Some analysts cite a potential trillion-dollar industry

Line by line. First, the "play money" choice is shrewd. The moment real money is wagered you sink into gambling regulation; Meta sidesteps much of it with a daily play-money allotment. It also lowers the barrier "like a game" to court mass participation — a social-friendly design that lets people enjoy "guess the future" with no risk of losing money.

Second, "Llama resolves the market" is the core controversy. The most sensitive thing in a prediction market is outcome adjudication — who decides "did that event actually happen" determines the market's trust. Meta wants to hand that ruling not to humans or an external body but to its own AI. What if the AI is wrong or biased? Truth itself could be distorted. That's why Forrester called it "Meta gambling with its trust."

Third, note that the market reacted immediately. When the NYT and CNBC reports landed, existing prediction-market-related stocks wobbled. The mere signal of a giant like Meta entering shook the board — evidence prediction markets are seen as "the next big thing," and Meta's entry instantly raised the stakes.

What each side gets

Meta's win first. One, a new engagement engine — "guess the future" creates addictive participation and brings people back daily for play money. Two, expanding Llama's real-world use — embedding its model as the core engine for generation, recommendation, and adjudication, a new path to recoup AI investment. Three, data — how people predict is itself enormous behavioral and psychological data, a goldmine for the ad empire that is Meta.

Users' win exists too. They get light fun "competing with friends over the future" with no risk of losing money. With AI auto-generating markets from trending issues, there's always something new to bet on, and personalized recommendations surface markets that fit one's interests. Done well, it could become a new way to consume information — "following the news like a game."

Conversely, there are clear risks and costs. AI deciding truth is double-edged. If Llama mis-resolves an ambiguous event or is subtly biased, that ruling becomes "the truth Meta set." And even with play money, a "prediction-stoking" mechanism risks amplifying misinformation or inflammatory topics. Meta could face the content-trust criticism of its past again — this time in the more sensitive form of "AI adjudication."

Past parallels — wins and failures

There's already a success showing prediction markets' power: Polymarket and Kalshi demonstrated that "collective prediction is sometimes more accurate than experts" on elections, sports, and economic events. Aggregate many people's bets and you get a kind of probability sensor. Meta is adding its own distribution and AI to this proven mechanism, chasing the next step: mass adoption.

Another to recall is the explosion of social games. A light, participatory feature bolted onto a giant platform can reach hundreds of millions in no time. Meta's distribution is exactly the force that can trigger that explosion. Fuse "play-money prediction" with the social feed, and it has the potential to pull in mass audiences existing prediction markets never reached.

Conversely, the shadow of failure is dark. Meta has ambitiously launched then quietly shelved various new apps before (standalone metaverse apps, some social experiments). And this time adds the regulatory-ethical minefield of "truth adjudication." If disputes erupt over AI-resolved outcomes, or regulators decide "play money is still gambling," the app could be blocked or sharply scaled back before launch.

Competitor counter-plays

Polymarket and Kalshi counter by emphasizing "the real thing" — genuine markets with real money at stake, plus transparency and neutrality in outcome resolution, drawing a line: "we're not Meta's play-money game." If Meta gathers the masses, it could even create an on-ramp — "entry → move to real markets" — making it a curiously mixed threat and opportunity.

Other Big Tech, watching this, may eye similar territory. As prediction markets emerge as "the next big board," the AI + social + betting combo is an attractive new-business model. But few have both Meta's distribution and an in-house model, a barrier that's hard to match quickly.

And regulators are effectively the strongest counter. They'll scrutinize how far the "play money" workaround dodges gambling rules, and whether "AI adjudication" is sound on consumer protection and information trust. Meta's real opponent may be regulation, not Kalshi — this app's fate may be decided in court more than in code.

So what actually changes

If you're a regular user, you may soon see a "today's predictions" feature inside Facebook or Instagram. Enjoying it lightly is fine, but start knowing one thing: AI decides what you predict and what counts as "fact." Use it for fun, but it's healthy not to take its rulings as absolute truth.

If you're in prediction markets or fintech, this is "the flare of mass adoption." Meta entering exposes the very concept to hundreds of millions. The total pie could grow, but a trust race will also unfold around the new standard of "AI adjudication" — who resolves outcomes more transparently and fairly becomes the differentiator.

If you follow AI ethics or policy, this is a real-world case of the weighty question "should we let AI decide truth?" Beyond content recommendation, AI sits in the seat that finally determines "what happened." How it runs and what disputes it spawns will be an important reference for future policy debates on "AI and truth."

One more layer — the Pandora's box of "an AI that adjudicates truth"

This app's real issue isn't prediction markets but the design itself — "handing truth adjudication to AI." The heaviest task in a prediction market isn't the bet but the step of "finalizing the outcome." Who decides "did that event happen" governs the whole market's trust. Existing platforms handled this via clear data sources or human review. Meta wants to hand that seat wholesale to its own AI. Clear-cut events are fine, but on ambiguous events open to interpretation, the AI's ruling becomes "the truth Meta set."

Here Meta's past shadow falls again. Meta is a company long criticized over content moderation and fake news. Now it's putting a system that auto-adjudicates "what is fact" with AI into a mass-market app. If the AI is subtly biased or mis-resolves an event due to training-data limits, that misjudgment spreads to hundreds of millions as the "official result." A company once accused of swaying public opinion with its recommendation algorithm now seeks an even more sensitive power: a "truth-adjudication algorithm."

And the play-money design deserves a second look. Not wagering money sidesteps gambling rules, but the "prediction-stoking" mechanism itself remains. A structure that auto-generates new markets daily to keep people betting resembles, in essence, an addiction design that maximizes engagement. If sensational or polarizing topics drive livelier betting, the AI has incentive to optimize for generating more such topics. That's where the line between "fun" and "opinion distortion" blurs.

In the end, Arena is less a tech product than a vast social experiment. "Will people accept AI adjudicating truth?", "How will regulators view play money?", "Will Meta wield this power in a way that earns trust?" — not one of these three questions is answered yet. So this app's fate depends less on code polish than on two external variables: trust and regulation. Even if it launches, the real test begins from there.

🥄 Three Things You're Probably Wondering

— Isn't this gambling? It's designed to sidestep traditional gambling by using a daily play-money allotment rather than real money. But how regulators view the "prediction-stoking" mechanism itself is still unknown, so it's too early to call it fully free of the gambling debate.

— What does "Llama decides truth" actually mean? AI makes the final ruling on whether an event happened and resolves the market. Clear-cut events are fine, but how the AI rules on ambiguous or contested events is the crux of trust. Get it wrong there and "truth" can be distorted.

— Will it actually launch? It's still at the leaked-document stage, so nothing's confirmed. Meta has ambitiously built and shelved apps before, and regulatory variables are big. "In preparation" is true; "confirmed launch" is not. Watch and see.

Further reading

• Meta plans to release AI-powered prediction market app, documents show — NPR
• Meta is building a prediction markets app. These stocks fell in response — CNBC
• Meta Gambles With Its Trust In Prediction Markets — Forrester
• Meta Prediction Market App Puts Llama in Charge of Deciding What Is True — TechTimes
• Meta plans to build a prediction market app — Yahoo Finance

Numbers and criteria are as of announcement and may change.