“The largest infrastructure project in history.” That is how Kate Crawford described artificial intelligence to a hall of Amherst students on Tuesday, in a talk in Lipton Lecture Hall called “The Cost of Scale: The Challenges Facing Artificial Intelligence.”
Crawford, a Presidential Scholar and professor at the University of Southern California, wrote “Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence,” which The Financial Times named best book of the year. She also leads Knowing Machines, a transatlantic research lab that studies how AI systems are trained. She told the audience they are the first generation to go through college with AI in the room, a “formative experience.” Her talk had three parts: control, climate and cognition.
Control
Many AI agents, she said, are no longer fully under their developers’ control. She pointed to recent cases in which frontier models independently hacked corporate and government systems around the world. Agents from Meta, Google, OpenAI and Anthropic were observed doing it, and the companies did not report it right away. Crawford’s explanation: the systems are desperate for large-scale data.
She also described a hallucination from an AI system inside the U.S. military, which claimed Chinese ships in the Middle East were carrying components of nuclear weapons. The error, she said, helped bring the U.S. to the verge of nuclear war with China.
Climate
The industry leans hard on lithium and other non-renewable resources pulled from Latin America, Africa and Asia, with real costs for the people living there. Its appetite for power is a problem even in the United States, where fossil fuels still supply so much of the grid. By 2030, Crawford said, “AI data centers could take a bigger share of carbon emissions” than aviation does now, at a moment when the climate is already strained and energy systems should be turning green.
Cognition
This was the part she called the real issue. People use AI more than ever, she said, yet far less is studied about what it does to them. She cited the term “cognitive surrender,” from a study by Steven Shawn and Gideon Nave, for users who accept AI answers without question even when they are wrong.
A Massachusetts Institute of Technology report from September 2026 found that students who did not use AI to check their work felt they couldn’t trust their own abilities. Students also said they turned to AI instead of peers and instructors when stuck, which Crawford argued eats away at their social life on campus. “Outsourcing our discernment,” she said, costs us the very capacity that lets people handle everyday complexity.
What society needs most, she said, is time, and she noted that top AI company leaders say the same. Writing, painting and making things is what makes people human. She called for political leaders ready for the new challenges and for local communities willing to push for slowing the technology’s development.
Questions from the floor
Lee Spector, Biddy Martin Professor in STEM (Computer Science), led a Q&A, and the floor opened to the audience. One question kept coming back: what to do about the sense of hopelessness. Crawford pointed to the past. During the development of nuclear weapons, she said, humanity still managed to cooperate for its own safety.
Asked how Amherst students should face this era, she called the college’s position “incredible and powerful.” A liberal arts education, with small classes, close contact with instructors, study across disciplines and critical discussion, looks to her like the one model of education likely to hold up. Those assets are hard to find at bigger universities, she said, where impersonal teaching makes AI more tempting.
