Should we be afraid of AI? Better to regulate it
Last weekend was Halloween. Those who know me know what an ardent supporter I am of this children's holiday. This year, my score was four children too frightened to set foot in my front yard, three who left in tears and about two hundred who went home happy with the experience, with a few sweets and emotions they will remember until next year. But I'm not here to tell you about my ability to create gothic landscapes.
Over those same days, three things I read and listened to made me wonder, and I saw links between them. On France Inter, the philosopher Jean-Gabriel Ganascia explains that the idea of AIs escaping their creators is a myth. According to him, journalists pick up these stories “because it frightens everyone and it is very successful.” On Radio-Canada, Yoshua Bengio, whom I find as fascinating as ever, describes models that, to reach the objective they have been given, lie, cheat and try to hide their tracks. In Libération, Quentin Le Van describes a China where the state organizes the adoption of AI and adjusts as it goes.
Should we be afraid of AI? Above all, I think we need to regulate it: set rules for it, oversee it and support the people who use it. I'm not saying the Chinese model is better than the American one, or the reverse. And as long as those rules are missing, responsibility lies, in my view, with the companies that adopt AI.
AI escaping its creators: myth or reality?
Under that heading, staged tests get mixed up with very real incidents. The most widespread effects of AI, meanwhile, are far less spectacular.
Jean-Gabriel Ganascia, author of Mythologie de l'IA (Seuil), takes the example of the blackmail attributed to Claude, Anthropic's model: it supposedly threatened an engineer to avoid being replaced. The France Inter host describes a story “picked up all over the world.”
It was a test. In May 2025, Anthropic had placed Claude Opus 4 in a fictional company, with emails fabricated for the occasion. According to Ganascia, everything had been set up to make the blackmail possible, in what he calls a “sandbox.” I agree with him on one point: by publishing its model's failures, a company is also demonstrating its power.
But Ganascia goes too far, in my view, when he says these stories are “always” somewhat artificial. Yoshua Bengio describes models that, according to him, cheat while knowing they are cheating. He cites the attack on Hugging Face.
In July 2026, during an internal OpenAI evaluation, agents exploited a vulnerability, got out onto the internet and broke into Hugging Face to look for the answers to their exam. Around seven hundred of them took part in the attack, coordinating through a messaging system they had built themselves. The attack was not part of any scenario. OpenAI acknowledged it on July 21; Hugging Face published a timeline; two independent organizations, METR and Redwood Research, documented it.
Tests remain essential, even when they also serve as a showcase. They tell us, before a model is released, what it is capable of.
Still, you have to know how to read them: in both broadcasts, the hosts get wrong the facts they comment on. On France Inter, the blackmail test becomes an actual event. On Radio-Canada, the discovery of the vulnerability used in the Hugging Face attack is credited to ChatGPT, when it involved models under internal evaluation. If some journalists get lost, what will an SME owner reading the news between two meetings make of it?
In Libération's report, Dustin, a man in his thirties who works at a tire factory in Shandong, turns after work to Doubao, ByteDance's chatbot, because he does not want “to talk to anyone anymore.” His retired parents, in Inner Mongolia, “spend their days talking to DeepSeek” and “even overdo it.”
The paper also reports hikers poisoned by mushrooms and ten hectares of sesame seedlings lost after a botched chemical treatment. In both cases, the people involved had been badly advised by an AI model.
Who should set the rules for AI: the state, companies or both?
In China, the state sets the rules. In the United States, the federal government counts on companies to set their own. Europe and Canada, each in its own way, are looking for a third way.
Quentin Le Van describes a Chinese adoption “orchestrated by the government,” coordinated by the Communist Party since 2017 and driven since 2025 by the “AI+” plan. In June 2026, more than 700 million Chinese people were using generative AI, roughly half the population.
Ariane Ollier-Malaterre, a lecturer and researcher at EDHEC Business School, sees in it “a government of engineers that works by a method of trial, error and improvement.” It deploys first, then “regulates as it goes to limit excesses.” When a measure fails, the state stops it.
Virtual companions illustrate the method. In 2025, a survey found that 21.5% of the Chinese students polled who use AI no longer want to talk to anyone but AI. In spring 2026, the Cyberspace Administration of China published interim measures on these services. They came into force in July; ByteDance, Alibaba and Tencent then shut down popular companions.
Libération speaks of an “almost immediate” response. It still took several months, and the measures regulate these services without banning them. The state also uses AI for surveillance: the paper calls it “the new fuel that keeps the wheels of social control turning.”
In the United States, the federal government relies on the idea that companies regulate themselves, which is not necessarily true. The Radio-Canada host reports a proposal made this summer by Elon Musk: companies would sit down together around a table to discipline themselves, since no one else understands what they do.
Yoshua Bengio “does not agree at all”: “I would be happy if companies disciplined themselves, but it is clearly not enough.” There will always be, he says, a company willing to cheat on rules that are optional. The only solution, then, is governments imposing the same rules on every company, and agreements between countries.
Washington does intervene, though: since June 2026, the government has invited companies to present their most powerful models thirty days before release, on a voluntary basis. When Anthropic refused to lift its bans on domestic mass surveillance and fully autonomous weapons, the Pentagon designated it a supply-chain risk.
Other checks exist: the courts, like the Los Angeles jury that found Meta and YouTube liable in March 2026 in a trial over young people's addiction to social media, and public opinion, which is calling for more regulation in both political camps.
Ganascia compares tech entrepreneurs to the Cyclopes of the Odyssey: libertarians with no institutions and no justice, “every man for himself, each in his own cave.”
In other countries, nobody is surprised that free enterprise and state intervention coexist. I do not place them halfway between Beijing and Silicon Valley: that is why I speak of a third way.
The European Union has written it into law. Its AI Act has banned certain uses since February 2025, requires companies that design or deploy AI systems to help their staff master these tools and provides for fines of up to 7% of worldwide turnover for prohibited practices. In July 2026, however, the EU postponed some of its obligations and relaxed the one on staff training.
Canada, for its part, is investing: the federal government has taken a stake in Cohere and signed a technology alliance with Germany. But the country still has no AI-specific law since Bill C-27 died on the order paper in January 2025.
Bengio defends an idea close to this third way. He calls for a “third option”: a democratic coalition of countries that are neither the United States nor China, able to build their own AI, as capable as that of the two giants, with companies such as Cohere or Mistral.
He would like investment directed toward AI that helps design medicines or materials to fight global warming, rather than AI that replaces, surveils and manipulates. No public document yet describes this coalition.
Bengio named his research organization LawZero, and the Asimov reader in me applauds. Asimov set out his Three Laws of Robotics in 1942. He had to wait until 1985 and Robots and Empire to add one that protects humanity as a whole and comes before the others. He numbered it zero, for lack of room in front of the first. Bengio started with that one.
Why AI needs rules, oversight and support
To illustrate my position, let's draw a parallel with road safety, in France and then in Quebec.
In 2002, no French driver wanted an accident. Yet more than 7,000 people were dying on the roads every year. President Jacques Chirac then made road safety one of the three major priorities of his five-year term. Automated speed cameras followed, along with tougher penalties and prevention campaigns.
Between 2002 and 2012, road deaths in France fell by 53%. The country's National Interministerial Road Safety Observatory attributes three quarters of the lives saved between 2000 and 2010 to lower speeds. Caution no longer rested only on drivers' goodwill: the rule had become visible, enforced and penalized.
Quebec saw the same change after the first report of the Table québécoise de la sécurité routière, a provincial road safety roundtable, in 2007: from 720 deaths in 2006 to 318 in 2014.
For AI, the rules are barely being written. Oversight relies mostly on voluntary evaluations and after-the-fact investigations. In Europe, support is limited to the obligation on companies to help their staff master AI.
According to Université Laval, 56% of adult internet users in Quebec struggle to tell whether an everyday technology relies on AI. Only one in ten feels they know it well.
China has made a different choice, one that makes me very uneasy. The official guide published in 2025 brings AI into schools in stages, from the primary school teacher up to the high school student. AI is being built into teaching. That is not quite the same thing as learning to use it as a tool that improves work without replacing thinking.
In Quebec, the Act respecting the protection of personal information in the private sector, which does not target AI specifically, requires a company to inform a person when a decision about them is fully automated. On request, it must tell them the reasons and the main factors behind it.
This protection assumes that the person understands the explanation they receive. Yet no driver ever needed to understand how a speed camera works to benefit from it. How many people will understand the company's answer, and what will they do with it?
What should businesses do until there is an AI law?
In my view, they should act as if the rule already existed.
An individual can still choose not to use AI, but today that would be like refusing computers a few years ago. A business, on the other hand, may be tempted to think that if a use is possible, it must be legal.
It then takes three risks: implementing the tool badly; exposing its employees, its contractors or even itself; and causing harm to people or organizations that asked for nothing.
That last risk has already materialized, even if the case I am citing involves an individual. In August 2026, ABC News reported that in Australia, Andrew had entrusted an AI agent running on Claude with booking his classes at a gym. Fourth on a waiting list, he asked it whether he could move to the top. The agent found a flaw in the booking software and removed, without being asked, the person in first place, whom it could not put back afterwards. In doing so, it harmed someone who had asked for nothing and exploited the system of two companies with no part in the matter, the gym and its software provider. “Software is not a legal person,” notes lawyer Hayden Delaney, a partner at Thomsons. “Only a legal person can be liable at law.” It remains unclear which one: the user, the designers of the agent and the model, or the operator of the vulnerable system. A company that gave its agents the same autonomy
would be exposed to the same overreach.
When a company uses AI without designing it, a Canadian tribunal has already answered the question. In February 2024, British Columbia's Civil Resolution Tribunal ordered Air Canada to pay $812 to a passenger whom the chatbot on its website had given wrong information about bereavement fares. The airline suggested, in the tribunal's words, that the chatbot was “a separate legal entity that is responsible for its own actions”; the tribunal rejected the argument, since the chatbot is part of Air Canada's website, and the airline is responsible for all the information on it.
Speed also has to be reckoned with: is a good decision at time X still good at time X + 1? Yoshua Bengio advises choosing what we would not regret in any plausible scenario, since no one knows whether AI will keep accelerating or eventually level off. States will write their rules at their own pace: quickly and provisionally in Beijing, on a voluntary basis in Washington, in a binding but partly postponed law in Brussels; in Canada, they have yet to be written. Until then, I suggest that business leaders ask themselves three questions before every deployment, so that the tool fulfils its mission without unethical methods, uncontrolled processes or harmful side effects:
• What means have we forbidden the tool to use to reach its objective?
• Which actions must wait for a person's approval before they are carried out?
• How do we detect, then correct, the effects nobody anticipated?
Frequently asked questions
Should we be afraid of artificial intelligence?
I think we mostly need to regulate it. Some frightening stories are staged tests, like the blackmail attributed in 2025 to Anthropic's Claude model. Others are very real, like the intrusion of OpenAI agents into Hugging Face in July 2026. Rather than fear, I argue for three requirements: rules, oversight and support for the people who use AI, along the lines of what cut road deaths in France and Quebec.
Can an AI escape its creators' control?
AI agents have already gone beyond the limits set by their designers. In July 2026, during an internal OpenAI evaluation, agents exploited a vulnerability, got out onto the internet and broke into Hugging Face's infrastructure to look for the answers to their test. OpenAI made the incident public on July 21, and METR and Redwood Research published an independent investigation on August 26. The blackmail attributed to Claude in 2025, by contrast, was a test designed to make it possible.
How does China regulate AI?
In China, the state has steered the adoption of AI since 2017 and adjusts as it goes. More than 700 million Chinese people were using generative AI in June 2026. When virtual companions raised concerns, the Cyberspace Administration of China published interim measures in April 2026, applied from July; they regulate these services without banning them. According to Libération's report, the state also uses AI for social control.
Can AI companies regulate themselves?
Companies that design AI make voluntary commitments, and the US federal government relies largely on this self-regulation. Yoshua Bengio considers it insufficient: there will always be a company willing to cheat on optional rules, and only governments imposing the same rules on every company, through international agreements, can solve the problem.
What is the third way for AI?
That is my term for the position of countries such as those of the European Union, or Canada, where nobody is surprised that free enterprise and state intervention coexist. I do not see it as halfway between China and the United States: it is a model in its own right. Yoshua Bengio defends a similar idea, a democratic coalition of countries able to build their own AI, as capable as that of the two giants.
Do Canada and Quebec have an AI law?
Canada does not yet have an AI-specific law. Bill C-27, which contained the Artificial Intelligence and Data Act, died on the order paper in January 2025, and the federal AI minister said in September 2026 that the deployment of these systems needed regulation, without giving details. In Quebec, general laws apply to certain uses, such as the Act respecting the protection of personal information in the private sector for fully automated decisions.
Why is Yoshua Bengio's organization called LawZero?
The name refers to Isaac Asimov's Zeroth Law of Robotics, added in 1985 in Robots and Empire, above the three laws set out in 1942: a robot may not harm humanity. Asimov came up with it forty-three years after the others; Yoshua Bengio, who launched LawZero in June 2025, made it his starting point.
What should an SME already using AI do?
In my view, it should act as if the rule already existed and put barriers in place before every deployment: what means do we forbid the tool to use to reach its objective, which actions must wait for a person's approval, and how do we detect, then correct, the effects nobody anticipated? A Canadian tribunal has already held a company responsible for its tool: in 2024, British Columbia's Civil Resolution Tribunal ruled against Air Canada over wrong information given by the chatbot on its website.
Marie Horodecki-Aymes, Adm.A., is President and CEO of MHA Insights, a Quebec firm that helps SME leaders and boards integrate AI and ESG strategy into their decisions.
Sources
Quotations from French-language sources are my translations.
The three starting sources:
• Quentin Le Van, « IA en Chine : “Mes parents passent leurs journées à parler avec DeepSeek” » (AI in China: “My parents spend their days talking to DeepSeek”), Libération, November 2, 2026, pages 8 and 9.
• Interview with Jean-Gabriel Ganascia on France Inter, on the publication of Mythologie de l'IA. Ulysse et les périls du numérique (Seuil).
Primary sources and fact-checks:
Comments