Meta builds AI-powered prediction market app using Llama

Per NPR, Meta is building a prediction-market app internally codenamed 'Antwerp' and 'FBForecast,' powered by its Llama large language model. The app's AI would automatically generate market questions from trending topics and surface personalized market recommendations to users — and, critically, the AI would also resolve markets, making the final 'yes' or 'no' call on whether an event occurred.

The AI-as-resolver design is the most consequential and contentious piece. Prediction markets live or die on trustworthy, neutral settlement; handing final adjudication to an LLM raises immediate questions about accuracy, manipulability, hallucination and bias — especially when the same company sets the questions, recommends bets and decides outcomes. It also wades into the regulatory minefield prediction markets already face (gambling and CFTC-style oversight in the US).

Strategically it fits Meta's broader 2026 pattern of shipping AI-native standalone apps and assistants: a creator companion app, an agentic assistant, and now a prediction product, all leaning on first-party Llama rather than third-party models like ChatGPT. Zuckerberg has repeatedly pitched 'agents that understand your personal context' as Meta's north star, and prediction markets are a engagement-and-data flywheel that fits social distribution.

This lands amid other Meta pressure points the same week — US pressure to agree to AI security reviews and litigation over Llama training — making the company's aggressive AI productization look both ambitious and exposed. Caveats: this is pre-launch from documents, feature set and timing may change, and regulatory approval is far from assured. Watch for launch confirmation, how settlement disputes are handled, and regulator reaction.