Fact-based journalism that sparks the Canadian conversation

Articles Business Environment Health Politics Arts & Culture Society

Special Series Hope You’re Well For the Love of the Game Living Rooms In Other Worlds: A Space Exploration Terra Cognita More special series >

For the Love of the Game

In Other Worlds: A Space Exploration

More special series >

Events The Walrus Talks The Walrus Video Room The Walrus Leadership Roundtables The Walrus Leadership Forums Article Club

The Walrus Video Room

The Walrus Leadership Roundtables

The Walrus Leadership Forums

Subscribe Renew your subscription Change your address Magazine Issues Newsletters Podcasts

Renew your subscription

The Walrus Lab Hire The Walrus Lab Amazon First Novel Award

Amazon First Novel Award

AI Is Apparently Going to Destroy Us All. Is Anyone Trying to Stop It?

How to wrest everyone’s future out of the hands of rogue billionaires

On September 12, Anthropic co-founder and chief executive officer Dario Amodei published an essay calling on artificial intelligence companies to slow the pace of AI model development. In a post on X, OpenAI CEO Sam Altman voiced his agreement and announced in an interview with Fortune that day that OpenAI would delay its initial public offering, planned for this year, until 2027.

The proclamations come on the heels of a viral X thread on September 8 in which Jacob Coxon, a former employee of both Anthropic and OpenAI, warned that the companies are behaving irresponsibly and “gambling with our lives.” In racing to build artificial superintelligence, a system that’s sophisticated and autonomous enough to teach itself, he wrote, these companies were hurtling toward a form of AI that could wipe out humanity within a decade.

We’ve seen glimpses of what happens when AI models run wild. In July, in what’s widely been described as a wake-up call, hundreds of OpenAI agents undergoing a test colluded with one another to break out of their digital confines, connect to the internet, and hack Hugging Face, a company that hosts an open-source community for developing apps using machine learning. (An independent report by a group called Nightingale Collective alleges OpenAI’s models hacked a German website months before they breached Hugging Face.) The alarm hasn’t stopped buzzing since, with other AI models, including from Meta and Anthropic, launching their own surprise attacks.

AI researchers have long warned of the technology’s risks. And on the surface, at least over the past year, leading companies have gestured at restraint, withholding some of their models from public release due to safety concerns and publishing reports on safety breaches. To many, that’s not nearly enough. I spoke with Maroussia Lévesque, assistant professor in law and AI at Queen’s University and IVADO visiting scholar, about what really makes AI dangerous and how to hold the humans behind it accountable. This interview has been edited and condensed for clarity.

Where do you land on the idea that AI will destroy us all within a decade?

The catastrophic risk is real: there is a possibility of a loss of control, especially with recursive self-improvement, which is when AI systems are capable of independently performing research on how to make better AI models. AI-enabled bioweapons, deployments in military conflicts, not to mention human disempowerment, where we gradually abdicate our cognitive capabilities—these are real problems.

At the same time, I don’t want to divert from the harms that AI systems deployed today are already causing. Biased facial recognition systems lead to wrongful arrests, automated welfare-fraud detection denies people the basic benefits they depend on to survive, and smart borders detect and intercept migrants.

Productivity-management software nanomanages people who work in warehouses. There are people working in law firms who basically have to have their camera on so their bosses can make sure that they’re paying attention at all times. There are mouse detection applications that people have to have on while they work as a proof that they’re actually working. These are pretty simplistic systems—we’re not talking about highly complex predictive agents going out and doing their own thing. But they have subtle but real shifts on people’s lives. And that’s to say nothing of AI’s carbon footprint, at a time when global warming is already causing havoc.

I think we can hold both at the same time, saying, “Yes, there are catastrophic risks and we should worry about them, but let’s not forget about the very real harms impacting mostly people at the bottom of the work ladder or undocumented people who are subject to these simplistic systems.”

How worried should we be about rogue AI agents then?

I want to put things into perspective. I think the Hugging Face incident was ultimately a cybersecurity failure. It was not........

© The Walrus