Yoshua Bengio

And some people will become maybe political activists to make sure governments will move in the right direction.

Governments do, to some extent, not enough, listen to public opinion.

And if people don't pay attention or don't put this as a high priority, then there's much less chance that the government will do the right thing.

But under pressure, governments do change.

It is really important that we evaluate the risks that specific systems pose.

So here it's the one with OpenAI.

These are different risks that researchers have identified as growing as these AI systems become more powerful.

Regulators, for example, in Europe now are starting to force companies to go through each of these things and build their own evaluations of risk.

What is interesting is also to look at these kinds of evaluations through time.

So that was 01.

Last summer, GPT-5 had much higher risk evaluations for some of these categories.

And we've seen actually...

real world accidents on the cybersecurity front happening just in the last few weeks reported by Anthropic.

So we need those evaluations and we need to keep track of their evolution so that we see the trend and the public sees where we might be going.

All of these.

So companies are doing it themselves.

They're also hiring external independent organizations to do some of these evaluations.

One we didn't talk about is model autonomy.

This is one of those more scary scenarios that we want to track.

where the AI is able to do AI research.