Mr. Chair and members of the committee, my name is Anthony Aguirre. I'm a professor of physics and the CEO of the Future of Life Institute, which is an international non-profit that seeks to steer transformative technology, especially advanced AI, in beneficial directions.
When we founded the Future of Life Institute 10 years ago, the field of AI policy didn't exist and the idea that AI could convincingly pass for human, prove new mathematical theorems, write complex computer code or ace nearly any human exam was a total fantasy. This was thought to be decades or even centuries away.
Today, AI can do all of these things and progress is showing no signs of slowing down. If this trend continues just as it is—in other words, if the same sort of progress we've seen from 2003 to now continues for the next couple of years—within one to five years, AI could well be able to do any cognitive task that a human can do. AI companies are aiming to create this artificial general intelligence specifically to replace rather than empower or aid human workers. That is the prize the companies are pursuing. That is what underlies the economics and the huge investment.
This is not labour displacement like previous technologies, but a wholesale drop-in replacement of tens of per cent of the labour force. I believe there's no viable plan in any country to address such a rapid and unprecedented upending of the human labour and income system, but this is not the biggest risk. AI could soon be able to do AI research and development by itself, rapidly improve itself and progress far beyond human capabilities into superintelligence. That is AI that competes not just with the best humans, but with humanity as a whole. This is, in fact, the stated goal of several of the AI companies. Multiple companies have explicitly said that they are building recursively self-improving AI.
I cannot emphasize enough how dangerous this is. Its developers would not be able to predict, understand or control for very long what such a system does. It would operate with inhuman speed, scale and sophistication toward goals that are fundamentally unknown to us. This is the reason that all of those eminent researchers and company heads signed the 2023 statement that AI presents an extinction risk to humanity. This is not a small risk. Many of the builders of the systems believe there are 10% to 20% Russian roulette odds. Some experts who are not building it put the numbers higher.
I understand that it is hard to take this in, given how big it feels and how unbelievably irresponsible it sounds, but this is what is actually happening.
The good news, however, is that this is not inevitable. The path laid out by a few giant U.S. AI companies is not the only possible path. There is still an opportunity to switch to a better path—one that develops powerful and trustworthy AI tools that empower and complement people rather than replace them, keeping humans in control. That is, we can develop systems designed from the ground up to be reliable and controllable and that have a particular scope and purpose, such as detecting cancer, predicting how proteins fold or doing advanced mathematics. Tool AI of this type would still enable medical and scientific breakthroughs, huge productivity boosts and products that allow people to do things they never could. We would just need to give up the poorly chosen goal of replicating and then exceeding all human capability in autonomous AI replacements.
At the Future of Life Institute, for these reasons we launched a statement calling for a prohibition of the development of superintelligence that is not to be lifted before there is, first, broad scientific consensus that it will be done safely and controllably, and second, strong public buy-in. This was signed by hundreds of experts from academia, employees of all the world's leading AI companies, politicians, religious leaders and many more.
I believe that Canada is well positioned to help shift tracks to this more responsible direction. Among other things, I recommend that Canada do the following.
First, support initiatives to create pro-human tool AI that aims to empower rather than replace humans.
Second, affirm that all AI systems must remain under meaningful human control. Strengthen and clarify liability laws so that responsibility for actions taken by AI systems rests where it should, which is with users, for systems they meaningfully control, and jointly with developers and deployers, if control is inadequate or absent. It's never with AI systems themselves.
Third, prohibit the development of superintelligence until and unless it can be shown to be safe and controllable and there is wide public buy-in.
Thank you.