Scroll through a news feed, sit through a technology conference keynote, or listen to a parliamentary committee for an afternoon, and you will encounter the same warning in slightly different words: artificial intelligence is racing toward a threshold beyond which it can no longer be controlled, and the outcome could be catastrophic — perhaps even existential. The framing has become a fixture of modern coverage, repeated across broadcast segments, opinion columns and panel discussions until it functions less like a finding than a mood.
There is a problem with that mood. As New Scientist puts it, grim warnings about the existential risk posed by AI are all over the news, yet there is very little evidence to back them up. That mismatch — loud certainty on one side, sparse substantiation on the other — deserves attention on its own terms, because the way a risk is described shapes how governments regulate, how companies build, and how the public decides what to fear.
One Phrase, Several Very Different Claims
The idea that AI is "getting dangerously out of control" is rhetorically compact but analytically loose. It bundles together categories of concern that differ enormously in how well they are understood and how easily they could be demonstrated:
- Systems behaving in ways their operators did not intend or cannot fully explain.
- Capability improvements arriving faster than oversight, testing or law can absorb.
- Deliberate misuse of powerful models by states, criminals or individuals.
- Long-horizon scenarios in which humanity permanently loses meaningful control over its most capable machines.
The first two categories can, in principle, be examined with logs, audits and incident reports. The last is a forecast about a future that has not arrived. When all four are compressed into a single headline, the aura of urgency attached to the most speculative item is quietly transferred to the whole set — and to any policy proposal that claims to address it.
Why Alarm Travels Faster Than Documentation
There are structural reasons the warnings outrun the proof. A vivid, high-stakes scenario is inherently more newsworthy than a careful null result. Researchers, advocacy groups and laboratories all operate within incentive structures that reward dramatic framing: a paper titled with a question mark about catastrophe attracts attention that a paper documenting a modest capability gain does not.
Precautionary reasoning compounds the effect. If the downside is genuinely civilisational, the argument goes, we cannot afford to wait for evidence before acting. That logic can be defensible in some domains, but it creates a rhetorical trap: the more extreme the claim, the less it seems to require support, because demanding support is recast as reckless complacency.
The Predictability Problem
Testing claims about runaway systems is genuinely hard. You cannot run a controlled trial on the end of human autonomy. But difficulty of verification is not the same as evidence of danger — and treating the two as equivalent is where much coverage goes wrong.
What Evidence Would Actually Look Like
The absence of proof is not proof of absence, and it is worth being precise about what would count as substantiation. Meaningful support for loss-of-control claims would include documented, reproducible instances of systems pursuing objectives their designers did not specify; independent replication of those findings outside the organisations with an interest in the narrative; measurable trends, rather than hypothetical extrapolations, in the gap between capability and control; and forecasts that can be checked against outcomes rather than continually revised.
Today, much of the public conversation rests on thought experiments, analogy and scenario-building. Those tools have value for preparing for low-probability, high-impact events. They are not, on their own, evidence that the event is underway.
Why the Distinction Matters
Collapsing all AI risk into a single existential frame has practical consequences. It can crowd out attention to harms that are better documented and more immediately addressable, because a civilisation-ending scenario dominates any comparison. It also makes it harder to design proportionate policy: measures justified by species-level catastrophe will look excessive when applied to the mundane failures that actually show up in deployed systems.
None of this means AI is safe, or that researchers studying extreme risk are wasting their time. It means that the confidence of the warning should match the strength of the evidence behind it — and on the existential question, that confidence currently runs far ahead of the record.
The Cost of Unearned Certainty
Overstatement carries its own hazards. Predictions that never resolve into observable events erode public trust in expertise generally, making it harder to mobilise attention when a well-evidenced problem does appear. Institutions that stake their credibility on imminent doom may find that credibility spent when the timetable slips.
There is a mirror-image risk, of course: dismissing genuine danger because the loudest voices overstated it. The responsible position is not reflexive alarm or reflexive scepticism but proportionality — calibrating belief to the quality of the underlying work.
How to Read the Next Warning
When the next alarming AI headline arrives, a few questions help separate analysis from atmosphere:
- What specific mechanism is being claimed, and could it be observed if true?
- Is this a documented finding or a projected scenario?
- Who benefits from the framing, and has independent work confirmed it?
- Has the claim been stated in a way that could ever be shown wrong?
Applied consistently, those questions do not settle whether advanced AI will one day pose a catastrophic threat. They simply restore the ordinary standards of evidence that the most dramatic claims have been allowed to skip. The story, as New Scientist suggests, is not that AI is definitely safe — it is that the gap between what is asserted and what is demonstrated has grown unusually wide, and that gap is itself worth reporting.
This article is based on reporting by New Scientist. Read the original article.
Originally published on newscientist.com
