Trump calls AI catastrophe warnings a “hoax”. That does not settle the safety questions

Donald Trump has rejected the push for stronger AI guardrails, arguing that presidential leadership and existing government powers are sufficient. In a follow-up post, he called warnings about AI taking over the world and destroying humanity a “HOAX”.

The original Truth Social posts make the disagreement clear. They do not provide a technical rebuttal to the safety proposals, nor do they themselves announce a new law. For readers following the debate, the useful question is what safeguards each side is prepared to make verifiable.

What Trump actually said

In the first post, Trump presents a strong president as the necessary control, asserts that the government already has substantial powers over AI companies, and alleges a conspiracy against AI and data centres. He also targets Anthropic’s Dario Amodei.

There is a SICK conspiracy going on against AI and Data Centers

That is an allegation made by Trump, not a finding supported by the post. The message offers no evidence establishing such a conspiracy. We are not adopting its accusations against critics or treating criticism of a technology as proof of disloyalty.

The follow-up questions why industry leaders would seek regulation, predicts enormous economic benefits, and dismisses catastrophic scenarios. His claims about China and other companies also require separate verification; repeating them as facts would go beyond what these posts establish.

The safety proposal is more specific than a demand to stop AI

Amodei’s essay distinguishes pacing from halting training. It proposes embedded third-party evaluators, coordination among democracies and eventually wider international coordination. It argues for more time and scrutiny to keep safeguards in step with capability.

That proposal deserves examination on its details, including whether evaluators can inspect enough information and report uncomfortable findings. It should not be reduced to an assertion that all development must end. Equally, Amodei’s forecasts are assessments of risk, not a timetable of events known to happen.

Read our earlier report on the frontier-safety debate

Existing powers and preventive testing answer different questions

A government’s ability to respond to wrongdoing is different from the ability to discover dangerous behaviour before a system is deployed. That is the distinction a useful policy debate needs to address. This is an analytical point, not a legal assessment of the powers Trump refers to.

Imagine a company preparing an assistant that can take actions in other software. A statement that someone will be held responsible after a failure does not tell a customer what was tested beforehand, what permissions the assistant has, or how it can be stopped.

Preventive testing also needs scrutiny. Tests can be incomplete, expensive or poorly matched to real use. Asking for evidence should apply to safety claims and performance claims alike, without assuming a company’s preferred policy is automatically in the public interest.

The companies should not get a free pass either

Our view is that rules should be judged by what they require and who can meet them. A proposal may improve accountability, create unnecessary barriers for smaller competitors, or do both. The useful assessment examines those effects rather than guessing the private motives of everyone supporting it.

Microsoft’s newly published consultation provides one example of a document readers can examine. It sets out intended human-control principles and invites public feedback. Its publication is a starting point for scrutiny, not evidence that the company has solved the safety problem.

The same standard should apply to claims that faster development is essential. What benefit is expected, over what period, and with what protection for people affected by failures? A national competition argument does not answer every operational question.

What this means for Australian readers

A US political statement does not change the settings on an Australian customer’s AI account. It does help show how contested the oversight of major AI developers has become. Readers should look for subsequent decisions, published requirements and implementation evidence rather than treating a social post as a completed policy change.

For a business using AI today, keep the purchasing questions concrete: what can the tool access, what can it do without approval, and how will you detect and recover from a mistake? Those questions remain relevant regardless of which public figure sounds most certain.

Neither a prediction of disaster nor a declaration that danger is a hoax is enough. The most useful next development would be evidence outsiders can inspect and safeguards whose operation can be checked. That is where this debate should lead.