Information & Ethics Committee on Jan. 26th, 2026

[Steven Adler]

Broadly, what I would think about is, what risks did the AI companies consider? How did they test for these? What results did they find? How do we know that this is all truthful and credible?

Right now, AI companies essentially get to grade their own homework. It really isn't surprising that you get grade inflation. They say 100% of the time that they handle this, but then, in the real world, that's not what you see.

The question is, how do we provide trust, both at a national level and also internationally, so that countries don't need to worry about rivals undercutting them? They save time by not doing the safety testing, because they're trying to race to get ahead. Ultimately, everyone pays the price for this.

[Leslie Church, Liberal, Toronto—St. Paul's, ON]

Is there anybody that's doing this well already? Are there any companies that are doing this as part of self-reporting that we should be looking at that might be leading the way? Maybe it's not perfect yet, but maybe there's something that we can look at as we move forward down this path.

[Steven Adler]

Yes. Companies have different bright spots. There isn't one that totally satisfies me. There's an organization called AI Lab Watch that rates companies on their efforts. I think the highest score they give is maybe 35%. That's the top mark. Some of them are maybe down in the 3% range.

These numbers are sensitive, but that's how people working on safety view this. There's a long, long way to go.

[Leslie Church, Liberal, Toronto—St. Paul's, ON]

I very much take your point about self-reporting and some of the challenges that are inherent in it. What do you see as the type of transparency that you would expect from a company that they could self-report on, versus the types of activities or safety checks that might belong or be better placed with a regulator?

[Steven Adler]

As a start, I would be happy to see the companies required to demonstrate this work to a regulator rather than to the public. If that's the crux of it, great, but by and large, this is not yet happening. We will see the beginnings of it when the EU's code of practice goes into effect this upcoming summer, if that goes as planned, but as of now, there's just not really a quality bar at all, let alone who you share your details with.

[Leslie Church, Liberal, Toronto—St. Paul's, ON]

Mr. Chair, can I open that line of questioning to the other two witnesses as well?

Do you have anything to add in terms of this issue around transparency and how we promote it, how we regulate it and where companies are currently falling down in transparency measures?

[The Chair, Conservative, John Brassard]

Monsieur Brisson or Mr. Miotti, if you want to take that, go ahead.

[Andrea Miotti]

Just adding to what Mr. Adler was saying, I think what's crucial is to move from the current regime. In the current regime, companies move first, and then governments have to scramble behind them. This always puts governments and the public on their back foot. Especially for the most powerful AI systems, it's crucial to move to a regime that is much closer to how we deal with any other high-stakes engineering project. We don't just let people build a bridge, then figure out a few years later that it's collapsing because we didn't check the plans. We don't let anybody build a nuclear power plant without first checking exactly what they're planning to build.

I think this approach should be taken increasingly with the most powerful AI systems, especially as these companies themselves expect these systems to be extremely dangerous. They are quite candid about it publicly. We should apply the same standard and flip the burden of proof. They should show, first of all, that their plans are sensible, and then they can go ahead, rather than going ahead first and then letting governments scramble behind them.

[Etienne Brisson]

I think it’s important to understand that the creators of artificial intelligence themselves say that they don’t necessarily understand what’s going on. However, when we hear representatives from these companies speak, they say with confidence that they know where they are going, that they have safeguards in place and that they have found the solutions.

I think it would be important for the companies themselves to show us where the limits are and also tell us what they do not yet understand.

[The Chair, Conservative, John Brassard]

Thank you, Ms. Church.

Mr. Thériault, you have the floor for five minutes.

[Luc Thériault, Bloc, Montcalm, QC]

Thank you, Mr. Chair.

Mr. Miotti, I will begin by talking about a situation I experienced.

In your brief, you mention Anthropic. Not so long ago, in the autumn of 2025, it announced a cyber-attack, carried out largely without human intervention and on a large scale. I quote: “AI was used to hack into other computers. Approximately 30 units were targeted simultaneously; successful intrusions affected large technology companies and government agencies.”

So this is no longer just theory. What lessons should we learn from this?

[Andrea Miotti]

Absolutely, this is an unfortunate but not surprising event. This is one of the many canaries in the coal mine of where AI companies are going by developing superintelligence.

This is because what companies are doing, the main path they're taking to get to superintelligence, is to develop AI systems that get better and better, especially at automating software development and especially at automating AI development. They want to initiate what some of these companies call an intelligence explosion, where AIs just kind of do their own homework and keep creating the next generation of AI with fewer and fewer humans in the loop. Naturally, ultimately, hacking is done by.... Malware is just software used for malicious uses, so as these AI systems are being developed to get better and better specifically at replacing and automating software engineers, they are also getting better at hacking and at cyber-attacks.

Ultimately, as companies keep investing in this specific angle, which is to make AIs better and better at developing software, we will see more and more of these cyber-attacks, but the scale will increase unless we prevent the development of superintelligence.

[Luc Thériault, Bloc, Montcalm, QC]

Earlier, we talked a little about possible verification. I would like to continue on that topic with a little more detail.

In your brief, you state that “compliance could be verified by international inspection regimes”, and you note that “the concentrated supply chain for advanced AI chips provides an additional point of control”. You also specify that “pushing the boundaries of AI currently requires the use of large supercomputers that consume the electricity of a city for many months”.

Could you clarify how these verification mechanisms work or could work in practice?

If Mr. Adler would like to add anything, he may also respond.

[Andrea Miotti]

Absolutely. One thing that is important to understand is that these powerful AI systems that are being developed are not just a few lines of code on a private computer of a private person. They require a lot of hardware. They require a lot of physical infrastructure that is visible from satellites. It's data centres that sometimes take the space of a football field, in some cases, or the space of a small town, and they're getting bigger and bigger. They're filled with tens of thousands to hundreds of thousands of the world's most powerful supercomputers, which are manufactured by only a very few companies, and in some cases, just by one company in some part of the supply chain.

This makes both enforcement domestically and verification internationally much easier than if we had to deal with just software on a laptop. It's much closer to how we can track plutonium and uranium in the case of nuclear weapons, and that's how we stopped proliferation.

Countries just chiefly need to focus on these large infrastructure projects. These give the countries leverage to intervene. For example, if there is a ban on superintelligence, countries can simply require the data centre companies to disclose whether companies on their data centre are doing this criminal activity and interrupt them. Even in cases of non-compliance, including internationally, countries can directly ask the data centre to be shut down in case, for example, a company or a rogue actor is still undertaking the development of superintelligence even outside of the law. It's making it easier to enforce.

[Luc Thériault, Bloc, Montcalm, QC]

Mr. Adler, you wrote, with regard to computing, that computing governance offers a potential avenue for international verification. Would you like to add anything?

[Steven Adler]

I think the thing to emphasize is that it is very hard for people to create one of these systems. As Mr. Miotti emphasized, enormous resources go into it. That is a pathway for being able to tell who is really at the frontier of AI making these most powerful, uncertain systems about what they will be able to accomplish, and to be able to then regulate accordingly.

[The Chair, Conservative, John Brassard]

Thank you, Mr. Thériault.

Mr. Hardy, you have the floor for five minutes.

[Gabriel Hardy, Conservative, Montmorency—Charlevoix, QC]

Thank you, gentlemen, for joining us.

What we have heard about artificial intelligence over the course of several committee meetings now is quite astonishing, and I think it is evolving very rapidly. We understand, of course, that the evolution of artificial intelligence is uncontrolled, even by its creators, that it hides its errors, that it learns on its own. We also learn that it uses a lot of water, a lot of electricity and that the risks are great.

I still have a question, because, on the one hand, there is generative artificial intelligence, and on the other, specialized artificial intelligence. We think that one is more dangerous than the other.

Mr. Brisson, I imagine you discuss this with other colleagues. When you talk to your colleagues around the world, do you get the impression that they currently agree on the risks posed by superintelligence? Does everyone see the same thing, or are there still different options, different visions of artificial intelligence?

[Etienne Brisson]

I would say that all experts certainly agree on the risks, especially the big ones. We’re talking about generative artificial intelligence, superintelligence. I don’t think anyone, except maybe the companies, thinks that’s a good idea.

However, there are definitely two sides to this. In the United States, many bills are being introduced. On the other hand, we want to win the race. It’s really important to win the race, especially given the current administration.

So we see the dangers, but for some people, these are mitigated by the benefits that artificial intelligence will bring.

[Gabriel Hardy, Conservative, Montmorency—Charlevoix, QC]

However, when it comes to wanting to win the race, we have often been told that once created, this superintelligence will belong to no one. So what is motivating all countries to race as fast as possible towards their own demise?

Every company is motivated by a goal. However, if it creates a product over which it will have no control, why would it spend billions when it knows it will lose it, that it will give it to everyone, even without wanting to?

[Etienne Brisson]

I can’t speak on behalf of companies. However, I believe they think they will be able to control this superintelligence. If that weren’t the case, I don’t think they would try to produce it. I think they still have hope, perhaps, or that they have this somewhat far-fetched idea that they will be able to control it, that they will be the only small group of humans capable of controlling this superintelligence.

[Gabriel Hardy, Conservative, Montmorency—Charlevoix, QC]

Do you share this view, Mr. Adler and Mr. Miotti? Do you see exactly the same motivation among companies? Why are they creating a product over which they will lose control?

[Steven Adler]

There are different plans for how people might control a system like this, but nobody thinks that they work. They are all flimsy and speculative today. One is to put AI in charge of making sure that another AI is good as it is building more and more complicated AI systems. You hope this structure doesn't topple down on you.

If I had to speak to the psychology of people inside the companies, they feel somewhat forced. They are racing, and they are doing it regretfully, because they think they have no choice. They say they have no choice. I think this is mistaken. In fact, within the last week, DeepMind and Anthropic, two of the leading labs, have said that if other companies and groups would slow down as well, to go more cautiously, they are open to this.

The issue is that they are not governments. They are not international bodies. They can control only their own decisions, so, unilaterally, they feel compelled to go ahead. Somebody else is going to do this. We may as well do it too and try to do it a bit better and a little more safely. Ultimately, what we would do is not do this at all until we're really sure we're ready.

[Gabriel Hardy, Conservative, Montmorency—Charlevoix, QC]

Go ahead, Mr. Miotti.

[Andrea Miotti]

Thank you.

As Mr. Adler was saying, ultimately I believe this is where the role of government comes in, because these companies have shown that they are unwilling or unable to stop. At the same time, they are spending hundreds of millions of dollars in the U.S.A., but also increasingly in other countries, to lobby against any form of regulation of AI whatsoever. They are running political campaigns against any candidate that is running for election criticizing AI, and trying to avoid regulation wherever they can.

Again, this is why we have governments. We would not have expected big tobacco to stop selling cigarettes just because they cause cancer. They were aware of it. They kept doing it. We had to have regulation. We wouldn't expect big oil to just stop emitting without regulation. Here it's the place for governments to step in. We can clearly see that the companies cannot self-regulate, and I think governments should draw a clear red line in the sand to say that superintelligence—