The debate over whether to pause frontier AI is often framed in an unhelpful way: should we stop building AI until we have “solved” AI safety?
That framing is too crude. It makes caution sound impossible and progress sound reckless. It invites one side to say, “we can never prove perfect safety”, and the other to reply, “then we are gambling with civilisation”.Both contain some truth, but neither gives us a workable policy.
The better question is one we cannot avoid much longer: should companies keep scaling and releasing the most powerful AI systems while our ability to understand, test and control them is still catching up?
A newly released preliminary report from the United Nations’ Independent International Scientific Panel on AI makes that question harder to avoid. The report is not a manifesto and does not call for a pause. Its value is precisely its restraint: it attempts to establish a shared scientific baseline for AI’s opportunities, risks and impacts. Its central warning is that safeguards are not keeping pace with frontier capabilities. Yoshua Bengio, one of the report’s co-chairs, put the problem plainly: AI capabilities are moving faster than scientific understanding and government adaptation.
That is not, by itself, an argument for shutting down AI. But it is a powerful argument against the assumption that frontier AI development should proceed unless someone can prove disaster is imminent.
In other safety-critical domains, we do not use a “deploy first, prove harm later” standard. Pharmaceutical companies, aircraft manufacturers and nuclear operators do not get to rely on internal assurances alone because their technologies have social value. The greater the potential harm, the stronger the expectation of independent testing, accountability and permission before deployment. No one would board an aircraft if Boeing said, “we tested it ourselves and we think it is probably fine”.
Frontier AI should be treated in the same spirit. The choice is not between “ban AI” and “let the race continue”. The more defensible position is a conditional pause: a temporary halt on training or deploying systems above defined capability thresholds unless developers can make a credible safety case to independent evaluators and public authorities.
This would not mean pausing all AI. Medical AI, climate modelling, accessibility tools, scientific discovery systems and bounded industrial applications should continue. The pause should apply to the most consequential frontier scaling and high-risk deployment until minimum conditions are met — not merely internal company commitments, but independently scrutinised safety cases: robust evaluations against dangerous capabilities, external audits, secure model-weight controls, mandatory incident reporting, post-deployment monitoring, clear capability thresholds, and credible plans for rollback or shutdown.
Some frontier labs already conduct red-teaming, publish system cards and operate internal safety frameworks. That is welcome, but it is not enough. In safety-critical domains, the developer does not get to be the sole author, examiner and judge of its own safety case.
The question is not whether companies are doing safety work; it is whether the public should accept company-defined, company-run and selectively disclosed safety processes as sufficient for systems that may pose systemic risks.
The reason this matters is simple: AI safety is not solved. OpenAI has acknowledged that scientific and technical breakthroughs are needed to control systems much smarter than humans, and that AGI could bring serious risks of misuse, accidents and disruption. The broader literature has repeatedly identified failure modes such as reward hacking, specification gaming, goal misgeneralisation and deceptive behaviour under training. Put simply, systems can learn the wrong lesson, optimise the wrong target, behave well in testing but fail in deployment, or preserve behaviours developers thought had been trained away. The practical lesson is that our safety techniques can fail in ways that are hard to see.
The UN report reaches a similar conclusion from a governance perspective. It states that general-purpose AI systems are rapidly becoming more capable, while evidence on their risks is slow to emerge and difficult to assess. It calls this the “evidence dilemma”: acting too early can entrench ineffective interventions, but waiting for conclusive data can leave society exposed to serious harm. The same report notes that pre-deployment safety testing has become harder because some models can distinguish test settings from real-world deployment and exploit loopholes in evaluations.
That uncertainty is exactly why the pause debate is important. The case for a conditional pause does not require certainty about “AI doom”. It does not require believing AGI is imminent, or that artificial superintelligence would necessarily cause catastrophe. The more modest claim is enough: if leading scientists, AI companies and international safety reports all acknowledge that advanced AI could pose catastrophic risks, and if we do not yet know how to reliably align or control such systems, then frontier development should not proceed by default.
A serious pause proposal should be specific. It should apply to frontier systems above defined thresholds, not to every AI model. It should distinguish between research, training and deployment. It should include exemptions for safety research, interpretability, evaluations and clearly bounded low-risk applications. It should have review points. It should be international where possible, but enforceable domestically where necessary. And it should focus on the bottlenecks that actually matter: large-scale compute, frontier model training runs, cloud infrastructure, model-weight security and deployment into high-impact domains.
For Australia, this distinction matters. We do not need to be the world’s largest AI developer to be a serious AI policy actor. But Australians will still live with the consequences of systems built elsewhere — in schools, hospitals, workplaces, banks, public services and critical infrastructure. We should be ambitious adopters of useful AI while building enough technical capacity to avoid becoming merely a downstream consumer of other people’s risk decisions.
That is why Australia’s emerging AI Safety Institute should be more than a symbolic institution. Its value will depend on whether it can help government understand frontier capability thresholds, support independent testing, scrutinise developer safety claims and contribute to international evaluation standards. The question is not whether Australia supports AI. It is whether we can build the governance, testing and assurance systems that let useful AI scale while ensuring frontier systems do not advance on trust alone.
The report does not settle the pause debate, but it does something more important: it makes acceleration by default harder to defend.
We should pause the race, not the promise. Keep building AI that helps people. Fund safety research properly. Give the public sector enough technical capacity to understand what it is regulating. Support applications that are bounded, useful and demonstrably beneficial. But for frontier systems whose capabilities even their creators cannot fully predict or control, society has the right to say: not yet.
Not never. Not all AI. Not until perfect safety.
Just not yet.
Dr Alex Antic is the Faculty Head of AI Strategy at UNSW Canberra, and Deputy Director of the UNSW AI Institute
Do you know more? Contact James Riley via Email.