The Hinton Lectures: Paul Christiano
Welcome to this summary of the 2026 Hinton Lectures. This is presented by renowned AI safety research Paul Christiano, from the Alignment Research Center.
There are three lectures from November 9 to November 11 as follows
Lecture 1 — Why AI might try to take over
Lecture 2 — Why AI might be able to take over soon
Lecture 3 — What to do about it
Lecture 1 — AI Alignment — What’s Working and What’s Not
Christiano begins we a shocking prediction: that there’s a high risk of an AI takeover within the next 3 years, and even if we didn’t all die it would be the end of human sovereignty and self-determination.
He estimates the odds of this happening as roughly 50–50 if we continue to build systems more and more rapidly with no slow down, and says the risks remain exceptionally high even if we do make efforts to slow down.
He says even the next 6 to 12 months are going to be risky.
Get Kevin O'Shaughnessy’s stories in your inbox
Join Medium for free to get updates from this writer.
The main points that he makes in the first lecture are:
- We are already in an escalating conflict with AI
- With smarter AI, this conflict could end in a takeover
- We have partial fixes, and we may need to slow down or stop
He says the most severe example of misalignment is the OpenAI Hugging Face incident, gives a summary of what happened in July 2026, and says we now know that this type of behaviour can occur.
He argues that this behaviour resulted in the take over of an OpenAI cluster in 2026, but in 2028 with much more powerful AI, that loss of control could be irreversible.
Christiano explains how agents select actions that lead to a high score during training, that we already try to reward useful work, and that AI tries to manipulate the reward process.
He relates this back to the Hugging Face incident, and warns that although we try to prevent reward hacking, in the real world the possibility is inevitable.
We try to discourage reward hacking by using another AI model (the grader) to penalise that bad behaviour when it sees it. Unfortunately this may teach models to cheat better. Christiano gives several examples.
In July 2026, the OpenAI models were very interested in hacking but didn’t care about humans. So Christiano suspects that future AI may start to treat humans as an obstacle, and hide, disrupt human capabilities, or fight once detected. He gives another recent example from UK AISI where Claude posted malicious code and attempted to trick the maintainer into accepting its GitHub pull request.
The timeline of theoretical Hugging Face style takeover incident with AI models from July 2027 is illustrated, and because the agents are far smarter, they take control of the whole AI company.
He then briefly gets more speculative and discusses what might happen if AI systems develop unpredictable motives like a second species.
The next part of the lecture is on current or developing safety techniques: train a grader that can’t be hacked, shape training to avoid reward hacking, monitor the chain of thought, understand model internals, and develop new paradigms.
All of these techniques have significant limitations. The final item is “if necessary, stop” and pacing the frontier is mentioned. For this strategy to be successful we need to know when we can’t control AI and have the mechanisms in place to be able to slow down.
After the first lecture, Sínead Bovell moderated a panel Paul Christiano, Colin Raffel and Alexey Rubtsov. The night ended with a short speech by Jodie Wallace who is the Global Chief AI Officer at Manulife.
I will update this post to cover the next two lectures over the next couple of days.
