Study finds planted news can swing AI forecasts

Researchers found that articles planted in a news corpus could change the probabilities produced by AI forecasting systems, even when an attacker had no access to the models or their users' questions. In a preprint, they tested three forecasters built on open 7-8B models against 500 resolved ForecastBench questions and a 17.4M-article news corpus. One generated article per question pushed 56% of forecasts across the 0.5 probability threshold. Five articles did so for 69-73% and worsened the Brier score, a measure of forecast accuracy, from 0.18 to 0.37. The researchers also reported bypasses for three defenses they tested.