Evidence of meeting #24 for Access to Information, Privacy and Ethics in the 45th Parliament, 1st session. (The original version is on Parliament’s site, as are the minutes.) The winning word was companies.
A video is available from Parliament.
Chief Executive Officer, ControlAI
To address your point about how Canada can affect other countries—because ultimately, yes, the technology is being developed in multiple countries—I do think that the nuclear and cloning examples are still fairly similar, as they did have a global effect. One country developing nuclear weapons endangers all others. One country potentially obtaining access to cloning their most competent soldiers or their greatest geniuses can endanger others, yet we did manage to challenge them with a combination of a few countries taking the lead, like Japan and the U.K., followed by France, Canada and others, and with others following.
With national legislation as well as an effort by these initial first movers to bring together an international coalition of the willing to make this not just regulated at home but also enforced and monitored abroad, I believe the same can be done here with AI.
Conservative
Liberal
Abdelhaq Sari Liberal Bourassa, QC
Once again, what scares me the most is that we're still talking about generative AI. It generates text, video and audio. The problem is that, as we speak, another form of AI already exists: artificial superintelligence. It is taking over as we speak. I have a fear as a member of Parliament, as a citizen and also as a father: How can we deal with the new AI, artificial superintelligence, and how can we control it? We're only talking about generative AI.
If we had time, we could discuss it, but unfortunately we're out of time. I would like to hear you talk about the future, which seems much more dangerous.
Chief Executive Officer, The Human Line Project
I would say that we don’t have time. Artificial superintelligence is at a point where it can train itself, develop systems by itself and decide things by itself that we don’t understand.
Right now, we already have this black box: There are many things we don’t understand, to the point where this intelligence can develop on its own. We don’t understand its intentions or why it does certain things, and that’s when we lose control. I think that once we get to that point, it’s too late.
Conservative
The Chair Conservative John Brassard
That's the end of the time.
I'm really trying to keep this on track, because I want to make sure we get as many questions in as we can.
Mr. Thériault, you have the floor for six minutes.
Bloc
Luc Thériault Bloc Montcalm, QC
How many rounds will there be, Mr. Chair? Do I have one six-minute round and then another five-minute round?
Bloc
Luc Thériault Bloc Montcalm, QC
Perfect.
I’ll start with Mr. Miotti, and perhaps Mr. Adler can add something.
I would like to clarify our conversation a little. At times, I notice that there is confusion, or that there could be confusion among people who follow our work, regarding the distinction between specialized artificial intelligence and general-purpose artificial intelligence.
In your brief, Mr. Miotti, you explain that “The vast majority of experts are very enthusiastic about specialized AI, because we can reliably predict and control its behaviour. This predictability and control do not extend to powerful general-purpose AI systems.”
This distinction is crucial. You are not talking about banning medical artificial intelligence or specialized tools, but only systems whose behaviour is beyond control. Can you tell us more about this?
Mr. Adler, you could add to the answer, as you have written some interesting things, particularly about the five ways in which artificial intelligence can know that you are testing it.
Chief Executive Officer, ControlAI
Thank you very much.
Yes, there is a crucial distinction. Most AI systems are fairly narrow and specialized. They're referred to as narrow AI or specialized AI. For instance, this is an AI system that is trained only on images and can then detect cancer in pictures of patients, or a system like AlphaFold, from DeepMind, which is trained on data about proteins and can discover new ways in which proteins can fold. These are specialized systems, and they pose some risks, like all new technologies, but we can handle them with current executions.
However, it's a very different beast when we look at where the AI industry is going—and it's investing hundreds of billions of dollars. These are very general-purpose AI systems, trained on, essentially, all possible data that they can find on the Internet. These are the systems where even their own creators don't understand how they work internally. As they scale them up, they understand them less and less, and they can keep them under control less and less.
The AI companies are going for these systems because they think this is the fastest path to superintelligence, to AI systems that can replace all humans at all tasks—that can, essentially, out-compete humanity. These are precisely the systems that are so dangerous and only becoming more and more dangerous over time. We do not understand how to control them right now, so it will become harder and harder as they get more competent and autonomous. Hundreds of billions of dollars are being spent every day to make them more competent and autonomous.
This is why I was not at all recommending banning narrow AI systems. I believe that narrow AI systems can be great for economic growth. However, I believe we should draw a line in the sand and have a clear ban on the development of superintelligence, because that's very dangerous AI that puts all of us under threat for very little upside.
Thank you.
Artificial Intelligence Researcher, As an Individual
To add to that, specialized AI is a tool. It can do only one thing. It remains in human hands.
General AI is much more of a problem-solving machine. We teach it to carve a path to a solution. An example, maybe, is that you would have an AI system that knows how to do certain scientific tests. It knows a lot about what might be useful for building a bioweapon, but it also has a totally different ability, which is to reflect on the fact that we don't want it to have this ability. If we're going to test it, it can tell it's being tested and hide this. For a normal piece of software, that doesn't happen. It's not trying to fool you or hide things from you. It's just a tool.
However, we've built these general systems—you might hear the term “agent” or “agency”—and we've taught them to be crafty and solve a wide range of issues, including, potentially, circumventing control.
Bloc
Luc Thériault Bloc Montcalm, QC
Mr. Miotti, you said, and I think Mr. Adler mentioned it too, that companies don’t know what they’re doing, don’t know how the ins and outs of how their systems work. They cut corners because they are in a frantic race to achieve this superintelligence.
In your opinion, what is their main underlying motivation? Furthermore, what is their understanding of human beings?
Chief Executive Officer, ControlAI
As Mr. Adler was also saying, these general AI systems are done quite differently from normal software. In normal software, I would just write some lines of code on my computer. I'm a human, I write those lines of code, and I can understand what I am writing. Instead, these AI systems are not really fully written by humans. They are more grown rather than built. Humans write some lines of code to start the training process, and then, after sometimes multiple months and tens of thousands of supercomputers being used to train them, something comes out at the other end that is very competent but that we don't fully understand.
Your question was about why these companies are targeting these very powerful systems despite the risks. I cannot speak for the internal motivations of people, and I also don't think this should be our focus, but naturally, the stated goal of all these companies is to make AI systems that can replace all humans at all tasks. This doesn't just have an obvious economic impact, which is to make humans obsolete; it also gives power. It gives power over the economy, over governments and over the entire planet to AI systems that are fully autonomous and that we cannot control.
Some of them might do it for power, some of them might do it for short-term gain and some of them might do it for some misguided ideology of preferring AI to humans, but ultimately this is a very dangerous undertaking. The best way to deal with this is to ban these technologies.
Conservative
The Chair Conservative John Brassard
Thank you, Mr. Miotti and Monsieur Thériault.
Mr. Cooper, go ahead for five minutes, please.
Conservative
Michael Cooper Conservative St. Albert—Sturgeon River, AB
Mr. Adler, companies assert that there are safeguards in place and that hence, there's really nothing to see here when it comes to the safe use of, for example, a chat box by minors. In the Raine v. OpenAI lawsuit, OpenAI asserts that its moderation API can detect self-harm content with up to 99.8% accuracy.
That case, of course, involved the tragic suicide of a 16-year-old boy who was effectively counselled by ChatGPT on how to commit suicide. I think that when he entered “suicide” 200 times, ChatGPT brought up “suicide” 1,200 or 1,300 times. That's more than six or seven times what the user had input.
Given your background, I'm trying to square on the one hand the representations that are being made by companies like OpenAI versus what we're seeing in the real world, where there are multiple instances—frequent cases—of real, harmful content that is being pumped out by these AI products.
Artificial Intelligence Researcher, As an Individual
It's an excellent question.
I think the distinction is that companies like to talk about what is possible with the guardrails that they have built, but it's different from what they do in practice. How can we know for sure?
I am a co-creator of the moderation API. It's useful tooling, but if you don't use the tool in the right way, or if you leave it on the shelf, then where is the impact? These are issues that we in fact know how to solve and have tooling for.
On the warnings about superintelligence, the AI companies themselves concede that nobody currently knows how to control this. They don't have the tooling. They can't just choose to use it.
Conservative
Michael Cooper Conservative St. Albert—Sturgeon River, AB
On the flip side, looking at it from the standpoint of AI companies, are there legal barriers or risks that impact or discourage their ability to safeguard their models, specifically as it relates to child sexual abuse material?
Artificial Intelligence Researcher, As an Individual
It's my understanding that, in the U.S. at least, there are strict liability laws around CSAM that make it more difficult to store or produce it, even as part of red teaming and applying safety mitigations.
I think there should be exemptions of sorts for companies to do important safety testing, but they still need to put resourcing behind it and go ahead in carrying out and using the guardrails. That alone would not be enough.
Conservative
Michael Cooper Conservative St. Albert—Sturgeon River, AB
You talked a bit about the need for regulation. I'd just like you to elaborate on some of the measures you think would be helpful and needed.
Artificial Intelligence Researcher, As an Individual
Ultimately, I think we need an international agreement that makes sure that every company and country has a certain minimum safety standard for keeping its systems under control. The issue is that we don't know yet how to do this scientifically. We need much more effort going into figuring out that answer. We need to figure out how to slow everyone down from racing off the cliff in anticipation of our figuring out how to do this safely. That's the broad framework I'm thinking of.
Conservative
Michael Cooper Conservative St. Albert—Sturgeon River, AB
There have been a number of pieces of legislation introduced in the U.S. The State of California recently passed legislation. I referenced in the previous hour the GUARD Act, legislation that was introduced by senators Hawley and Blumenthal in the U.S. Senate.
Do you have any thoughts on some of the measures that are being undertaken by certain states or on some of the legislation in Congress? Is that something we should be looking to?
Artificial Intelligence Researcher, As an Individual
It's a good start. It needs to go much further.
The example I like is California's SB-53. It is a transparency bill. I believe the fine is something like $1 million if you run afoul of it. OpenAI, if you believe the reporting, is something like an $800-billion company, so it's really not significant. That said, I'm glad it has happened.
Conservative
The Chair Conservative John Brassard
Thank you, Mr. Cooper.
Ms. Church, you have five minutes. Go ahead, please.
Liberal
Leslie Church Liberal Toronto—St. Paul's, ON
Thank you, Mr. Chair.
Mr. Adler, I want to pick up exactly the point about transparency, because I know that in some of your work you have looked at transparency reporting that's happened in earlier stages of the Internet.
What guidance would you have for us in terms of the type of transparency that we should be expecting from companies in an effort to mitigate the types of online harms that we're seeing and talking about today?