Dartmouth, 70 years on: AI speeds up security and education, but responsibility can't be delegated

By Dartmouth · November 3, 2026.

Dartmouth, the university where the term “artificial intelligence” was coined 70 years ago, held a three-day conference titled “AI at 70: What Must Remain Human?”. It brought together more than two dozen panels, talks and presentations featuring pioneering researchers, industry executives and scholars of AI's social impact. The university itself has published a summary of several sessions; what follows is based on that summary.

National security: faster and at greater scale. According to Sasha Baker, head of national security policy at OpenAI and former acting Under Secretary of Defense, foundation models change two things: they speed up the pace of change (in science, cybersecurity and other fields) and they make scaling possible. Baker said most militaries and governments are still absorbing what that means. For his part, James Patton Rogers of King's College London argued that the proliferation of drones and increasingly autonomous weapons points to a global shift in power. He cited Iran's response to the US operation “Epic Fury”, with drones and precision missiles at longer ranges.

There were two useful points on governance. Patton Rogers noted that the UN has spent a decade discussing lethal autonomous weapons with little regulatory progress. Baker cited a US Department of Defense directive requiring “appropriate human judgment” in the use of autonomous weapons, adding: “there is always an accountable commander.” She also rejected the analogy with nuclear non-proliferation: nuclear weapons depend on physically scarce, controllable materials and know-how, whereas with AI, in her words, “we are probably already past that point.”

Education: AI literacy as a subset of competencies. James Dobson, a professor of English and creative writing, described real costs: AI has “drained some of the color” from learning, makes assessment very difficult and takes an emotional toll on those navigating shifting norms. He also saw opportunities to rethink pedagogy. Roopika Risam, chair of the Department of Film and Media Studies, explained that a faculty group defined four core competencies: knowledge in one's own field, critical inquiry, ethical and sociopolitical awareness, and “responsible agency.” AI literacy is a subset of all of these. Her example is clear: if you haven't mastered a field, you can't tell whether what a language model gives you is any good.

The other sessions. Molly Crockett (Princeton) warned against the “illusion of generalizability”: a model matching human performance on cognitive tasks says more about the limits of those tasks than about the model's capability. Hany Farid, a pioneer of digital forensics, argued that there will be technical solutions to manipulated images, but that “trust is very human” and is currently eroded when it comes to institutions. On the arts panel, Nona Hendryx presented a “digital twin” trained on her music, and the speakers argued that the creative process matters as much as the result. Finally, Lisa Marsch presented Evergreen, a generative platform supporting student well-being that is being evaluated in a randomized controlled trial with 200 Dartmouth undergraduates; the idea is to intervene before a crisis, not after. It is an ongoing initiative, and the summary includes no published results yet.

Our take. The common thread across the panels is that what scales with AI is capability, not accountability. That fits the thesis we hold: the short term will be uncomfortable, with educational assessment in crisis, eroded trust in images and a race in drones and military autonomy without a solid international framework, but none of that is an argument for slowing down; it is an argument for governing with evidence.

Two caveats. First, Baker's warning about the nuclear analogy is honest and uncomfortable: if AI can no longer be controlled like a scarce material, governance will have to rely on accountability, standards and human oversight, not on turning off the tap. Second, Crockett's critique serves as an antidote to hype: mistaking a good test score for human understanding inflates expectations and misdirects investment. It is worth distinguishing demonstrated capability from aspiration.

The question in the title, what must remain human, has a practical answer: judgment, accountability and the prior knowledge needed to evaluate what a machine produces. That is also the key to the education ahead, and the reason why Evergreen, or any AI tutor, must be tested in controlled trials before being accepted. If we get this right, long-term abundance will free people to devote themselves to what they are passionate about, but only if, during the transition, we preserve the human capabilities that allow us to steer it.