Generative models succeed on two levels at once: observable behaviour, and the internal representations of the world they build for their own use. The talk asks how those internal representations compare to the ones humans build, and what happens at the seam.

The practical stake is a failure mode rather than a capability gap. Where the system's model of the world and the person's model diverge, the system can "set us up to fail" through ordinary interaction — not by being wrong on its own terms, but by being right on terms the person does not share.

Worked examples named in the abstract:

• Chess. A model trained to win, then paired with a weaker partner who makes some of the moves.
• Navigation. A model trained to find shortest routes, then handed an unexpected detour.

And a theoretical result that sharpens the whole thing: successful generation is achievable by agents provably incapable of identifying the model they are generating from.

Joint work with Ashton Anderson, Karim Hamade, Reid McIlroy-Young, Siddhartha Sen, Justin Chen, Sendhil Mullainathan, Ashesh Rambachan, Keyon Vafa, Fan Wei, Chris Qiu, and Prabhakar Raghavan.

The generation-without-identification result is the interesting half, and it lands near an argument already in the corpus. The shipping-the-model note works the same seam from the engineering side: a model that cannot disagree with the implementation it describes is not an instrument, because agreement is guaranteed and therefore uninformative. Kleinberg's version is stronger and stranger — an agent can generate competently from a model it could not identify if asked.

The chess-with-a-weaker-partner setup is also the human-in-the-loop case the attribution work kept running into from the other direction: two parties holding different states, and a system that resolves the difference silently rather than surfacing it.

Kleinberg works on the interaction of algorithms and networks, their role in large-scale social and information systems, and the societal consequences. Member of the National Academy of Sciences, the National Academy of Engineering, the American Academy of Arts and Sciences, and the American Philosophical Society. Served on the National AI Advisory Committee and the National Research Council's CSTB and CSTL. MacArthur, Packard, Simons, Sloan and Vannevar Bush fellowships; Nevanlinna Prize, World Laureates Association Prize, ACM/AAAI Allen Newell Award, ACM Prize in Computing.

To be written after the talk.

• Event page, Harvard SEAS
• SEAS events calendar