Suivre
- Site webbelfercenter.org
Stephen Casper
Professeur adjoint de politiques publiques à la Harvard Kennedy School et coauteur d'un pre-mortem du Belfer Center sur le risque catastrophique de l'IA, qui avertit que toute capacité néfaste d'une IA qui peut exister finira par se diffuser largement, et que les risques crédibles se comptent en mois, ou en un an ou deux, si ce n'est dès maintenant.
Fonctions
- oct. 2026 – aujourd’huiAssistant Professor of Public Policy, Harvard Kennedy SchoolAlso a Faculty Affiliate of the Harvard School of Engineering and Applied Sciences and a Faculty Affiliate with the Belfer Center's Program on Emerging Technology, Scientific Advancement, and Global Policy. No start date for the professorship is sourced anywhere; the date given is the earliest a source attests the role. The same page records that he completed his PhD at MIT and did a research residency with the UK AI Security Institute before joining Harvard, that he is a writer for the International AI Safety Report and a lead writer for the Singapore Consensus, and — in its own words — that he "never has and never will work for big tech, and has never signed a non-disparagement agreement".Source →
Au registre
19 oct. 2026 · Entretien · Harvard Gazette
So the first thing is to recognize and account for the fact that minimizing risk is incompatible with maximizing legitimate usefulness. There’s a trade-off.
Assistant Professor of Public Policy, Harvard Kennedy School à l’époque
The next sentences of the same answer, and the first item on his list of safeguards. He goes on to scoping capabilities through training — filtering biohazard-related research out of the training set — and then to training a model to refuse harmful requests even when it is prompted for them.
19 oct. 2026 · Entretien · Harvard Gazette
You can’t maximize useful capabilities without also enabling some harmful capabilities.
Assistant Professor of Public Policy, Harvard Kennedy School à l’époque
Asked which AI safeguards help reduce the risk of a major incident. The first he names is scoping a model's capabilities, and this is his reason for it: a model that can help with benign and legitimate virology experiments can probably also help with experiments creating harmful pathogens.
19 oct. 2026 · Entretien · Harvard Gazette
A second thing that can be predicted is that a very large, AI-enabled terrorism incident is probably going to be the result of a closed-weight frontier AI model that is poorly safeguarded or an open-weight model.
Assistant Professor of Public Policy, Harvard Kennedy School à l’époque
The first thing he says can be predicted is the shape of the harm — bio, chemical and cyber risk. He holds that closed-weight models can be made very secure against consistent misuse with a well-developed toolkit, and that the incident-level threats therefore come from closed-weight models that are poorly safeguarded and from open-weight models, which are hard to safeguard at all.
19 oct. 2026 · Entretien · Harvard Gazette
The trends here would suggest there are credible risks within months, or a year or two, if not now.
Assistant Professor of Public Policy, Harvard Kennedy School à l’époque
The next sentence of the same answer. He adds that it is hard to point at events and say "this is here" or "this is a few months away", and that risk management at a government level is about decreasing the rate of catastrophic incidents and making them more endurable.
19 oct. 2026 · Entretien · Harvard Gazette
Something’s eventually going to happen. The lessons of the past few years have taught us that if there is a bad AI capability that can exist, it will eventually widely proliferate.
Assistant Professor of Public Policy, Harvard Kennedy School à l’époque
Answering whether it is inevitable that a major AI-assisted event will occur. The Gazette prints the passage twice, once in the body of the edited conversation and once as a pull-quote.