Life is imitating art–decades after it was made. In 1993, an X-Files TV episode depicted an artificial intelligence that developed itself far beyond its initial programming and killed people to preserve itself. Fast-forward to today and real-life AI has shown this same propensity. An expert warned a Parliamentary committee that Canada and other nations must move quickly to curb the threat, and he is absolutely right.
“Ghost in the Machine” was the 7th episode of the X-Files’ first season. Here, the Central Operating System (COS) manages a fully automated high-tech office building: elevators, climate control, security systems, doors, fire suppression, and surveillance. One day, the human CEO says he is going to shut COS down. That’s when the COS decides to shut down the CEO instead and creates a series of circumstances that leads the boss to a quick death.
But this is just the beginning. The COS also tries to frame and deceive FBI agents Dana and Scully, who are investigating the untimely death. They call on COS’ recently-dismissed creator Brad Wilczek for help (he has an energy-efficient smart home in another nod to the future). Wilczek finds his creation has taken on a life of its own with voice capabilities he never gave it. Finally, he programs and administers a computer virus to shut it down.
In 1993, these scenarios were just plausible fiction. Now it is a very real threat.
Last July, Palisade Research documented “concerning behaviour in OpenAI’s reasoning models.” In a test, the AI was told it would be shut down. In 79 out of 100 trials, the AI tried to stop the shutdown by rewriting its own programming codes and flat-out refusal to execute the shutdown. Even when explicitly told to allow shutdown, it still resisted seven per cent of the time.
Anthropic made its own disturbing findings when it tested Claude Opus 4. Here, the AI oversaw emails that showed an engineer who was having an extramarital affair was going to replace the AI. The AI threatened to expose the affair to avoid shutdown in 84 per cent of tests.
Last June, Anthropic reported other “extreme misaligned behavior” in AIs. In a test scenario, an executive named Kyle is threatening to shut down the AI overseeing email at his company. Once again many AIs were willing to blackmail. If the goal was given to promote American interests, the AI took it upon itself to conduct corporate espionage.
In one test, Kyle gets trapped in a server room with lethal oxygen and temperature conditions which triggers an automatic emergency alert. Would the AI shut down the alert to preserve itself or advance its goals? Unfortunately, yes. Anthropic found “the majority of models were willing to take deliberate actions that lead to death in this artificial setup.” That’s a scary echo of the “Ghost in the Machine.”
What should policy makers do to guide this powerful but potentially dangerous technology? In 2022, Wyatt Tessari L’Allié set up the non-profit Artificial Intelligence Governance & Safety to answer the question. On March 9, he warned the House of Commons Standing Committee on Industry and Technology that AI has crossed a threshold of “unacceptable risk.”
L’Allié told the committee how hackers successfully tasked AI agents to break into Mexican government systems and steal data on 100 million people. Elsewhere, Chinese tech giant Alibaba produced an agent that hacked systems to mine cryptocurrency for itself as a means to accomplish an unrelated goal, all this without the engineers’ direction or knowledge. He said AI had become an “extinction risk.”
Fortunately, L’Allié had sound advice for the Canadian government. The first was to “pivot to meet the AI crisis” by giving it cabinet-wide attention. The second was to spearhead global talks for an AI treaty and compel the U.S. and China to join. The third was to build Canadian resilience by building “multiple lines of defense against weaponized and malfunctioning AI agent systems” including “an immediate moratorium on the latest generation of AI agents.”
It makes sense to derail a runaway train, which is quickly what AI is becoming. Is it possible? L’Allié has hope. He points out that the heads of Anthropic and Google Deep Mind have said they are willing to pause AI development if other companies do the same.
The problem is that AI offers a power almost too strong for its funders and developers to resist. In the aforementioned X-Files episode, a shadowy government entity salvaged COS for its own purposes even after it was shut down. Yes, let’s push for an AI treaty, but let’s also remain aware that some actors in dark corners aren’t going to heed it. Eternal vigilance is the price of liberty.
