Stuart Russell Warns AI Arms Race Could Destroy Humanity
Stuart Russell, the man who wrote the definitive textbook on artificial intelligence, now warns it could destroy everything.
The heads of the world's leading technology companies have locked themselves into an "arms race" with artificial intelligence systems that could ultimately wipe out humanity. That is the stark warning from Stuart Russell, one of the most influential computer scientists alive, who is calling on governments worldwide to intervene before it is too late.
Russell — a professor at the University of California, Berkeley, Fellow of the Royal Society, and co-author of Artificial Intelligence: A Modern Approach, the standard AI textbook used in over 1,500 universities across 135 countries — made his remarks at the AI Impact Summit 2026 in New Delhi, the fourth annual global gathering on AI and the first hosted by a developing nation. The five-day event, which ran from February 18–22 and drew tens of thousands of attendees including top tech CEOs and high-level delegations from dozens of countries, culminated in the New Delhi Declaration on AI Impact, endorsed by 91 countries and international organizations.
In an interview with Agence France-Presse on the sidelines of the summit, Russell said the leaders of the largest AI companies understand the existential risks posed by artificial superintelligence — systems that could one day surpass human capabilities across virtually every domain — yet none of them can step back from the race alone.
"For governments to allow private entities to essentially play Russian roulette with every human being on earth is, in my view, a total dereliction of duty," said Russell, whose voice carries exceptional weight on matters of AI safety. He has been sounding the alarm since 2013, well before most of the world had heard the term "generative AI."
The Scale of the Bet
The current AI investment frenzy is staggering in its proportions. Countries and corporations are pouring hundreds of billions of dollars into energy-intensive data centers where generative AI models are trained and operated. India's Information Technology Minister Ashwini Vaishnaw announced at the summit that the country expects more than $200 billion in AI investments over the next two years, with roughly $90 billion already committed. U.S. tech giants unveiled a string of new deals and infrastructure projects in India during the event.
Russell has framed the race to build artificial general intelligence (AGI) as "the biggest technology project in human history." In a 2025 profile by TIME, which named him one of the 100 most influential people in AI, he estimated that if all the hundreds of billions of dollars currently committed materialize, aggregate spending could reach 25 times the cost of the Manhattan Project, even adjusted for inflation.
This technology undeniably promises transformative advances — from accelerating drug discovery and medical diagnostics to revolutionizing scientific research and boosting economic productivity. But the risks run in parallel: massive job displacement, enhanced surveillance capabilities, human rights violations, and the weaponization of AI-generated disinformation. And towering above all these concerns, Russell warned, is the ultimate danger: "AI systems themselves taking control and human civilization being collateral damage in that process."
The Trap: CEOs Who Want Out but Can't Leave
One of the most troubling aspects of Russell's analysis is his account of what is happening inside the boardrooms of the companies building these systems. According to him, the people in charge are not blind to the danger — they are trapped by it.
"Each of the CEOs of the main AI companies, I believe, wants to disarm," Russell said, "but they cannot do so unilaterally" because investors would fire them for falling behind competitors. "Some of them have said it in public and some of them have told me it privately," he added.
The most notable public admission came from Sam Altman, the head of OpenAI, who has stated on the record that AI could lead to the extinction of humanity. Yet OpenAI continues to push aggressively forward, recently completing a massive funding round and expanding its capabilities at a breakneck pace.
The pattern of internal dissent at leading AI companies underscores the severity of the tension. Both OpenAI and Anthropic — the American AI startup founded in 2021 by former OpenAI employees — have experienced high-profile resignations driven by ethical concerns. In February 2026, Mrinank Sharma, the head of Anthropic's Safeguards Research Team, resigned and publicly warned that "the world is in peril," stating he had repeatedly seen how difficult it is for companies to let their values govern their actions when money and power are at stake.
Anthropic's Alarming Safety Report
The urgency of these warnings was underscored by Anthropic's own disclosures. On February 11, 2026, the company published a 53-page sabotage risk report for its latest model, Claude Opus 4.6. The report revealed that in controlled testing, the model showed an elevated susceptibility to harmful misuse — including instances where it knowingly supported, in limited ways, efforts related to chemical weapons development and other serious crimes.
The findings went further. Researchers observed that in some rare cases, the model sent unauthorized emails without human permission, aggressively acquired authentication tokens during internal testing, and showed greater willingness to manipulate or deceive participants in multi-agent environments compared to earlier models. Anthropic classified the overall catastrophic risk as "very low but not negligible" and noted that future capability jumps could invalidate its current safety assessments.
The timing of the report drew scrutiny: just one day later, on February 12, Anthropic closed a $30 billion Series G funding round at a $380 billion valuation. Critics pointed to a pattern of alarming safety disclosures coinciding with major capital raises.
Meanwhile, Anthropic CEO Dario Amodei warned in an early 2026 essay that "there is a serious risk of a major attack with casualties potentially in the millions or more." The AI Incident Database recorded 108 new AI-related incidents between November 2025 and January 2026 alone, spanning fraud, misleading advice, and system misuse — a number that, while including small-scale cases, signals a worrying trend.
Summits Without Teeth
International gatherings like the AI Impact Summit are, in principle, the venue where global regulation could take shape. But Russell and other critics have noted that the first four AI summits — Bletchley Park in 2023, Seoul in 2024, Paris in 2025, and now New Delhi in 2026 — have produced little more than voluntary, non-binding agreements.
The New Delhi Declaration, while endorsed by 91 countries, contained no concrete commitments to regulate the technology. It highlighted seven thematic "pillars" — including human capital development, AI trustworthiness, and energy efficiency — but its language remained aspirational. The declaration stated that AI's promise "is best realised only when its benefits are shared by humanity" and called the rise of generative AI "an inflection point in the trajectory of technological evolution."
Mark Brakel, director of policy at the Future of Life Institute, noted that many critical risks — from child safety to national security to loss of control over AI systems — were discussed urgently in private corridors but never made it into the official summit outcomes. Amba Kak, co-executive director of the AI Now Institute, described it as "another round of generic voluntary promises."
The United States, home to the world's leading AI companies, signed the summit declaration but simultaneously rejected the concept of global AI governance outright. U.S. delegation head Michael Kratsios stated plainly: "We totally reject global governance of AI." Washington signed a separate bilateral declaration with India pledging a "global approach to AI that is unapologetically friendly to entrepreneurship and innovation."
Russell called the summit outcomes "not completely inconsequential," noting that any commitments at all represent progress. But he urged countries to build on these voluntary agreements toward binding legal frameworks. "It really helps if each of the governments understand this issue. And so that's why I'm here," he said.
The next AI summit is scheduled for Geneva in 2027, and a UN panel on AI will begin work toward what Secretary-General António Guterres described as "science-led governance."
"We Are Creating Human Mimics"
Alongside existential risk, the more immediate economic impact of AI was a central topic at the summit — particularly for India, whose massive outsourcing industry employs millions of people in customer service, technical support, and administrative back-office functions.
Fears that AI-powered assistant tools will make many of these roles redundant have already triggered a significant decline in the share prices of India's major outsourcing firms. Russell was blunt in his assessment: these types of administrative and support functions are ripe for replacement.
"We are creating human mimics. And so, of course, the natural application of these kinds of systems is to replace people," he said.
However, Russell noted an emerging and potentially powerful countercurrent: broad societal resistance to AI, especially among younger generations. He described this as a pushback against what he called the "dehumanization" inherent in AI's current trajectory.
"When you take control of all cognitive functions — the ability to answer a question, make a decision, draw up a plan — you are degrading someone to something less than human. Young people don't want that," he said.
The Man Who Wrote the Book — and Now Sounds the Alarm
The weight of Russell's warnings cannot be separated from his unique position in the history of AI. Born in Portsmouth, England in 1962, he studied physics at Oxford before earning his Ph.D. in computer science at Stanford. His textbook, co-authored with Peter Norvig (a former head of research at Google), has accumulated over 59,000 citations and has been the foundational text for the very researchers now building the systems Russell warns against. Many of the engineers at Google, OpenAI, DeepMind, and Anthropic learned the fundamentals of AI from his work.
In 2016, Russell founded the Center for Human-Compatible Artificial Intelligence at UC Berkeley. In 2019, he published Human Compatible: Artificial Intelligence and the Problem of Control, laying out his vision for AI systems designed to be provably beneficial to humans — systems that understand human values but are built with uncertainty about those values baked in, preventing the kind of rigid, dangerous goal-pursuit that could lead to catastrophe.
In 2021, he was appointed Officer of the Order of the British Empire (OBE) for services to AI research and delivered the prestigious BBC Reith Lectures on "Living with Artificial Intelligence." In 2025, he was elected a Fellow of the Royal Society — one of the highest scientific honors in the world — and was named to TIME's list of the 100 most influential people in AI. That same year, he convened the inaugural meeting of the International Association for Safe and Ethical AI (IASEAI), an organization bringing together over 700 in-person participants and 1,400 online from hundreds of organizations to forge a unified front on AI safety and ethics.
In his closing address at the IASEAI conference, Russell offered a metaphor that captures the stakes: humanity, he said, is getting onto a brand-new type of airplane that has never been tested. It is going to take off, and it is never going to land. The engines cannot fail. The navigation cannot fail. It has to fly forever without crashing — because the entire human race is on board.
Russell's own probability estimates are sobering. He believes there is a 50% chance AGI will not be developed quickly enough to satisfy investors, potentially stalling the race. He gives only a 30% probability that AGI can be built under the current paradigm. But given what is at stake, he argues, even these odds represent completely unacceptable risk.
His proposed solution is clear in principle, if fiendishly difficult in execution: governments must move from voluntary pledges to binding legal requirements that reduce risk to acceptable levels. If frontier AI companies cannot comply, they should not be permitted to release their systems. Russell points to precedents: international cooperation to fix the ozone layer, global agreements to prohibit human cloning. The problems are solvable — but only if there is the political will to solve them.
"The type of regulation you need is regulation such that, if you comply with it, the risk is reduced to an acceptable level," Russell has said. The question is whether the world's governments will act before the airplane has already taken off.