Influence Benchmarks for AI Systems

Description: DARPA seeks proposals for a testbed using economic frameworks to benchmark AI biases and latent behavioral preferences in dynamic simulated markets, including adaptation, collaboration, and deception. The effort requires a sandbox that uses only subject-system queries and outputs, supports auction or market models and news-driven agent behavior, and develops classifiers and quantitative comparisons against human-like behavior. The official posting identifies the opportunity as a September 2, 2026 pre-release with proposals opening September 23 and closing October 23, 2026, and lists $300,000 for Phase I over 12 months and $1,800,000 for Direct to Phase II over 24 months.