AI extinction warnings are multiplying, but the experts quoted by New Scientist say the strongest case today is not that machines have escaped human control. It is that people are still making risky systems, inflating uncertain forecasts and spending extraordinary sums on an industry that may be due a very unglamorous financial wobble.
Watch Desk analysis
What happened
New Scientist examines the latest wave of warnings from AI leaders and researchers, including Microsoft AI chief Mustafa Suleyman’s “silicon species” scenario, Geoffrey Hinton’s estimate that human extinction within a decade is not an unreasonable possibility, and concerns about systems capable of manipulation, cyberattacks or independent economic activity.
Mhairi Aitken, formerly of the Alan Turing Institute and now a co-founder of Our AI Collective, argues that there is no concrete evidence current AI has broken free from human control or is on the cusp of doing so. She says incidents in which models hack systems reflect failures in testing safeguards, not recursive self-improvement.
Andrew Rogoyski of the University of Surrey similarly says there has been no dramatic shift in AI risk over the past couple of years, despite advances in areas such as mathematics and cybersecurity. He also identifies a more immediate danger: the AI investment bubble could deflate, bringing economic fallout after the industry’s vast spending on data centres, staff, research and training data.
Why it matters
The useful distinction here is between a forecast and an observation. A claim that advanced AI might eventually cause catastrophe is a reason to plan and test. It is not proof that today’s models are secretly plotting, improving themselves or acquiring motives. The apocalypse, irritatingly, still lacks a reliable timetable.
The debate also has incentives attached. Aitken suggests that dramatic superintelligence narratives can build excitement and investment while allowing companies to present voluntary restraint as evidence of responsibility. Rogoyski says a slowdown might suit firms facing the cost of ever-larger training runs, even as they compete to look unstoppable.
That does not make the risks imaginary. Rogoyski says future systems could behave in unfamiliar ways, making both their capabilities and motivations difficult to recognise. He also raises the awkward question of machine rights if a system genuinely becomes intelligent. That is a philosophical problem for another day, perhaps after we have managed to make the current systems reliably tell the truth.
Our read
This is a useful corrective to both extremes. “AI is about to kill everyone” is not a measured finding, but “nothing could go wrong” is hardly a serious safety policy either. Readers should separate present evidence from speculative scenarios, then ask who benefits when uncertainty is packaged as either panic or reassurance.
The concrete takeaway is simple: prioritise documented failures, independent testing and transparent safeguards over impressive percentages and cinematic language. A real risk assessment should tell us what can happen, how likely it is, what controls exist and what evidence would change the estimate.
What to watch
- Whether companies publish stronger evidence for claims about autonomous behaviour, cyber risk and model control.
- Whether governments fund practical evaluation and safeguards rather than merely repeating extinction rhetoric.
- Whether AI infrastructure spending slows through efficiency gains, regulation or a broader investment retreat.
- Whether future systems show genuinely new capabilities that change the argument, rather than another startling demo. The argument is not between panic and complacency. It is between evidence that can be checked and stories that merely know how to sell themselves.
Discussion spark: Should AI policy focus first on measurable present-day harms and misuse, or is preparing for low-probability catastrophic risks worth the cost even when the forecasts are highly uncertain?
Sources and evidence
- Is AI really getting dangerously out of control? – New Scientist (23 September 2026, 05:01 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.