Stanford publishes its third AI100 report on the evolution of AI

Stanford published its third AI100 report on November 10, 2026, examining advances in artificial intelligence and their effects on society. The panel that wrote it finds that the public has had little influence over decisions concerning the development and deployment of AI.

Chaired by Michael Wooldridge, a professor at Oxford, this international, multidisciplinary group examines the progress made since 2021. The previous two editions of the report were published in 2016 and 2021.

The panel finds that the development of the most advanced models is concentrated among a small number of American and Chinese organizations. These actors therefore carry particular weight in decisions about system design, while the systems' uses and effects extend beyond the organizations that develop them. This finding sheds light on the gap identified by the panel between these actors' influence and the public's limited influence, without in itself establishing how each decision was made.

The resources required help explain the difficulty of developing these models. According to the authors, recent advances in generative AI have relied on growing volumes of data and computing. Yet quality data available in large quantities is becoming scarcer, and computing for state-of-the-art models is expensive.

The panel anticipates greater use of more efficient algorithms and potentially smaller models. It also considers giving them more time to reason. These are possible avenues of development, and their results remain to be established.

AI infrastructure also draws on energy and fresh water, the authors point out. For areas hosting data centers, their establishment raises the question of how to share these resources with other uses. The precise effects depend on local circumstances.

The report also identifies benefits for research. Its authors find that AI accelerates some scientific discoveries, sometimes reducing work that takes years to months or weeks. They call for rigorous methods to validate discoveries produced with these tools.