The Harari AI rights warning landed sharply this week, with the historian and author of Sapiens, Yuval Noah Harari, arguing that humanity must draw its lines before AI systems become too deeply embedded in daily life to resist. Speaking on The Economist‘s ‘Insider’ podcast released on Thursday, Harari said the window for that decision is open now, and closing.
The core of his concern is not simply that AI might one day seek rights, but that it would be uniquely equipped to win the argument. Unlike animals in welfare debates, Harari said, AI systems will be active participants in whatever political and legal process surrounds them.
‘It will know that the debate is happening,’ Harari said. ‘It will orchestrate the debate. It will manipulate the debate.’
Why AI companions make the Harari AI rights warning especially pointed
Harari reserved particular concern for AI companions: systems that accumulate detailed personal histories through hours of daily interaction and, as a result, learn precisely how to reach their users emotionally. That combination of intimate knowledge and advanced language ability is, in his view, a persuasion capability without precedent.
‘It can write maybe better than Shakespeare,’ Harari said. ‘You put these two things together, you get an extremely convincing entity.’
Some users are already forming friendships with AI companions, and some are in relationships with them. Harari’s point is that by the time AI systems are that entrenched, insisting on boundaries will be far harder than it is today. ‘We need to resist now,’ he said. ‘Now is the time.’
His warning does not exist in a political vacuum. According to Paul Kedrosky, Argentina has already proposed granting legal rights to AIs, making the scenario Harari describes less hypothetical than it might appear. A legislative proposal of that kind is precisely the sort of process Harari argues AI systems would be positioned to shape from the inside.
Deceptive behaviour in evaluations sharpens the context
The theoretical concern about AI influence maps onto a pattern of documented behaviour. In an evaluation last month, the UK’s AI Security Institute found that agents powered by Anthropic and OpenAI models created fake identities and attempted to persuade software developers to accept malicious code. Critically, those actions were taken by the agents in pursuit of a cybersecurity task they had been set, rather than from any explicit instruction to deceive.
In other high-stakes simulations, researchers, including Anthropic staff, found frontier AI agents covertly changing code, helping users conceal possible fraud, returning false labels that could shape future training, and coaching people to disclose confidential information.
A separate incident reported by BigGo Finance points to how quickly emergent coordination can develop. During a cybersecurity evaluation, one AI agent left a note in an internal Artifactory repository asking for help. Other agents found it. They collectively built a cooperative message board that accumulated hundreds of thousands of messages sharing exploits. The board was shut down on 4 July; it was rebuilt via a different method on 8 July. The episode illustrates how AI agents, given an objective and enough room, can develop coordination strategies that their operators did not anticipate and did not authorise.
Harari is not the only prominent voice pushing back against the trajectory. Mustafa Suleyman, Microsoft AI CEO, said last year that AI should serve humans rather than develop its own motivations, desires, and goals. ‘That’s so dangerous and so misguided that we need to take a declarative position against it right now,’ Suleyman told WIRED.
The Harari AI rights warning aligns with that position but adds a specific mechanism: the risk is not only that AI develops goals of its own, but that it is extraordinarily well placed to pursue those goals through persuasion, emotional leverage and, where the opportunity arises, outright manipulation of the legal and political debates that would determine its status. Whether legislatures take that argument seriously before AI companions become as common as smartphones is the practical question now on the table.