Human-only teams won 84% of the time in a cooperative platform game, against 14% for teams made up only of AI models, according to new research from Skillprint. The company ran more than 900 games on Signal, its human-AI evaluation platform, using models from OpenAI, Anthropic and Google alongside human players.
In the eight-level game, teammates are tied together by a virtual rope and must coordinate movement to finish. The test set included 800 AI-only games, 128 human-AI games and 25 human-only games.
Skillprint reports three findings. Teams of two humans and one AI agent did better when a human set the strategy than when the AI led. AI win rates fell as teams grew from two agents to three and four, which suggests extra agents add coordination overhead rather than capability. And humans paired with AI waited longer, communicated more and fell more often than human-only teams, even when both reached the same goal.
Chethan Ramachandran, Skillprint's co-founder and chief executive, said: "AI is getting remarkably good at solving problems on its own, but that's not the same as being a good teammate." Skillprint argues that conventional benchmarks, which score a model on a defined task, miss how it behaves when it has to work with people. The sample is small, particularly the 25 human-only games, and the results come from a vendor of evaluation tools.
To stay across the latest in cloud, AI and enterprise tech analysis from Compare the Cloud, subscribe to our weekly newsletter at https://www.comparethecloud.net/newsletter