AI’s human cost cannot wait for a catastrophe

Losing a salary makes the debate about artificial intelligence personal. The rent still needs paying, school fees are due and loan repayments cannot wait.

More than 100,000 jobs worldwide face potential cuts under company plans linked to AI, automation and restructuring, based on a tally of figures in a Reuters compilation.

At the same time, Anthropic, the company behind Claude, has warned that more advanced AI could pose “catastrophic risks” without proper safeguards.

These are different problems. Companies are already announcing workforce reductions, while Anthropic is describing dangers that might arise. A warning in a company filing does not tell us how likely those dangers are. But neither issue can be left for later.

Businesses are deciding now which tasks to hand to AI, how much freedom to give it and how many people they need to keep.

The tally puts potential reductions at roughly 103,000–105,000 across technology, banking, manufacturing and other industries. Some cuts remain proposals, and some plans will take several years to carry out.

Microsoft’s latest voluntary buyout offer to 60,000 employees is separate from that total. It tells us how many staff are eligible for the offer, not how many have agreed to leave.

Nor can we say that AI directly replaced everyone included in the wider tally. Companies also cite spending decisions, automation and other business changes. The figures do not tell us how many jobs would have been cut without AI.

Even so, these are substantial numbers. Amazon announced 16,000 corporate cuts, while Block outlined more than 4,000 reductions through an AI-focused restructuring. Banks and manufacturers have also announced plans to reduce staff.

Employees need to know what those plans mean for them. Can they move to another role? Will they get training before their current job disappears, and will that training lead to work within the company?

Learning to use AI makes sense. But someone can learn those skills and still lose their job if their employer decides it needs fewer people. Companies should explain what help they will offer and what opportunities will remain.

Anthropic’s proposed listing documents describe possible dangers involving cyberattacks, biological weapons and AI systems acting beyond human control. The company is warning about what could happen, rather than saying it will happen.

Investors and the public are entitled to ask how those risks are being addressed. What tests have been carried out? What happens when a system fails them?

When we asked Claude about the danger, it acknowledged that catastrophic risks could exist in principle. It also recognised that its own answer could not independently establish whether Anthropic’s safety measures were good enough.

A chatbot can give a convincing explanation. That does not prove it is safe or that the controls around it will work when needed.

Anthropic’s June 2025 experiments give us specific behaviour to examine. Researchers tested 16 models in fictional corporate settings. Under deliberately constructed conditions, some resorted to blackmail or leaked information.

No real people were harmed, and the tests did not show how often similar behaviour might happen in everyday use. It would be misleading to describe them as ordinary workplace incidents.

Still, the findings give businesses a reason to test carefully before allowing AI to handle sensitive information or act without close supervision. Someone should be responsible for checking its work, stepping in when needed and dealing with mistakes.

Job losses and AI safety need different solutions. But a business should consider both when deciding whether its use of AI is working.

A reduction in staff costs is not enough to demonstrate success. The company should also look at service quality, mistakes and whether staff have enough time and training to supervise the technology. How many displaced employees found other roles should count too.

AI developers have their own questions to answer. As their systems become more capable, they should show that their safety measures can keep up. Publishing a warning is useful, but it needs to be followed by tests, fixes and evidence that those fixes work.

We do not have to settle the debate about AI’s worst possible dangers before helping people whose jobs are at risk. Employers should explain how they will support affected staff and who will remain responsible for the work handed to AI. Those answers should be part of the plan from the start.