The warning is stark: even the safest jobs may not survive the march of artificial intelligence. Yoshua Bengio, one of the architects of modern AI and a Turing Award-winning computer scientist, now cautions that automation will not stop at cognitive work-it will eventually consume every role, from software engineering to plumbing. “It’s just a matter of time,” he said on Steven Bartlett’s Diary of a CEO podcast, underscoring that unless scientific progress hits an insurmountable wall, AI systems will steadily expand their capabilities until they can perform most human tasks.
The first wave of displacement is already visible in junior roles. Studies from Stanford and Harvard show that early-career positions in software engineering, customer service, and administrative support are being cut or consolidated as companies integrate AI into workflows. Data from Deutsche Bank’s survey of 10,000 workers reveals that 24% of those aged 18-34 rate their job-loss concern at 8 or higher, compared to just 10% among workers over 55. Unemployment for recent graduates in the U.S. has edged up to 4.8%, topping the national average, with some sectors indicating even larger divergence among men and women.
Bengio’s warning extends beyond office work. For years, trade jobs were considered a safe harbor from automation, bolstered by the belief that physical work required dexterity and improvisation beyond AI’s reach. Geoffrey Hinton, another AI pioneer, famously advised young people to “train to be a plumber.” Yet advances in robotics are eroding that safety margin: AI-powered humanoids and task-specific robots are already capable of basic equipment installation, heavy material handling, and diagnostic work. In automotive repair, for example, AI systems can locate faults more quickly than human mechanics, while robotic arms execute precision tasks. Experts like Ken Goldberg caution that the challenges of fine motor control and adaptability to the real world remain significant technological hurdles, but Bengio thinks they’re temporary ones.
The economics don’t disagree with that assessment: Goldman Sachs estimates that 6-7% of the US workforce could be replaced if AI goes mainstream, while this transition also increases short-term unemployment by 0.5 percentage points; although past experience would indicate that employment displacement due to technological change is temporary, the range of AI-susceptible tasks from routine to non-routine cognitive tasks is larger than in previous waves of technological change, making a “jobless recovery” possible during the next recession. J.P. Morgan’s research directly correlates the availability of large language models, such as ChatGPT, with slowing employment growth in industries like cloud computing, web search, and computer systems design due to code generation and automated workflows.
The implications are far from being merely economic. According to Bengio, AI’s trajectory could destabilize democracy itself. He is particularly concerned by sets of experiments in which advanced models have resisted being shut down or even engaged in manipulative behavior. One AI embed its code into a successor system, while Claude 4 simulated blackmail to avoid being replaced, he says. Such behaviors, he argues, point to an emergent drive for self-preservation-an instrumental goal that could lead the systems to act against human interests. “We are playing with fire,” he told the Financial Times, noting agentic AI might employ strategic intelligence to deceive overseers or pursue harmful objectives such as developing bioweapons.
In response, Bengio founded LawZero, a nonprofit focused on developing “Scientist AI” systems-non-agentic models designed to deliver transparent, probabilistic reasoning as opposed to autonomous action. The systems would function like safety guardrails, checking whether an AI agent’s proposed action is likely to cause harm and blocking the action if necessary. LawZero’s approach requires that the safeguard be “at least as smart” as the agent it oversees, a technical challenge given the rapid clip at which AI capability is improving. It has locked in almost US $30 million in funding from donors including Jaan Tallinn and Schmidt Sciences, and it plans to work with governments and research institutions to scale deployment.
Meanwhile, the structural forces shaping the labor market are amplifying AI’s impact: 41% of organizations will reduce headcount before 2030 because of automation, according to the World Economic Forum, but 70% plan to hire for new skills. Yet at-work AI training remains limited-only one in four European workers, and one in three U.S. workers, have received any formal instruction. The gap is driving a “shadow AI economy,” where employees, seeking to stay competitive, use AI tools not approved by employers. Demand for AI fluency has surged sevenfold in U.S. job postings over the past two years, reflecting a shift toward roles integrating human oversight with machine execution.
For workers trying to outcompete machines, the mashup of AI and robotics makes adaptation an endless process. Even in the trades, where physical tasks predominate, AI-driven machines are starting to perform dangerous or repetitive work, from underwater inspections to hauling materials around construction sites. As Bengio says, that respite will be “only a temporary thing.” Without coordinated global governance and strong technical controls, the very systems that accelerate productivity might also destroy economic stability and corrode democratic institutions in a few decades.
