Definition
Human-AI teams producing more homogeneous output of higher average quality. In Ju and Aral's MindMeld experiment, 2,310 people built advertising campaigns in pairs, some with a human partner and some with an AI agent. The AI pairs were 60 percent more productive per worker. Their work also converged on similar styles.
What is at stake
The average goes up and the spread goes down, inside the same experiment. An organisation that measures quality alone will call this a win and never see what it lost. The loss shows up later, when every team's output starts to look like every other team's.
Where the word came from
- Creativity as a social dilemmaDoshi and Hauser, 2024, the narrowing measured across a pool of individual writers
- Diversity collapseJu and Aral, 2025, the same narrowing measured in teams
- Mechanised convergenceMicrosoft and Carnegie Mellon, 2025, the same effect reported by knowledge workers
How episcope came to it
Harang Ju and Sinan Aral, MIT, March 2025. A preprint on arXiv, revised since and not yet peer reviewed, reporting field experiments on a collaboration platform built for the study, with every message and every edit logged, and a subset of the ads run on X to link the work to real click rates.
What to watch for
The same study is also a jagged frontier result. The AI pairs wrote better copy and produced worse images, so the frontier ran through the middle of one task. The productivity figure travels with the convergence beside it, because the paper reports them together and the second is the price of the first.
Papers and articles behind it
Ju, H., & Aral, S. (2025). Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance. arXiv:2503.18238.