US President Donald Trump has pushed back against calls to slow the development of advanced AI, calling some of the people raising concerns “very negative forces” and arguing that the United States cannot surrender its lead to China. Speaking to reporters in Doonbeg, Ireland on 13 September, Trump also said guardrails could be used, but rejected the idea that current warnings justify a broader slowdown.
The remarks are politically important because they narrow the kind of AI safety policy that looks plausible from the White House. A general pause, capability freeze or government-backed slowdown now looks less compatible with the administration’s stated position. Targeted controls around cybersecurity, procurement and high-risk deployments remain much more plausible.
| What Trump signalled | What it means in practice |
|---|---|
| The US should keep accelerating against China | Federal policy is likely to keep favouring faster access to frontier models, infrastructure and deployment. |
| “Whoever wins AI wins” | AI is being treated as a strategic competition issue, not only a consumer technology or safety issue. |
| Guardrails are still acceptable | The remarks do not reject every safety control or model evaluation requirement. |
| Some warnings are being exaggerated | Arguments for broad capability slowdowns will face a high political bar unless advocates can tie them to specific, measurable risks. |
Trump rejected a slowdown, not every AI safeguard
Trump’s comments were made after a week of increasingly stark warnings from people inside the frontier AI industry. Former Anthropic researcher Jacob Coxon argued that competition between leading laboratories could push development faster than safety work can keep up. Anthropic chief executive Dario Amodei then called for companies to moderate the pace of capability improvements. At the same time, OpenAI chief Sam Altman and xAI founder Elon Musk publicly supported parts of that push.
Trump’s response was not a technical rebuttal to those claims. He did not present evidence that frontier-model risks are impossible, nor did he say AI should operate without controls. The direct video of Trump’s remarks in Doonbeg records him allowing guardrails while arguing that the US must remain ahead of China and dismissing scenarios he believes are unlikely.
That is a narrower position than saying AI safety does not matter. It is closer to a hierarchy: competitiveness first, controls where the administration believes they are justified, and strong resistance to measures that materially slow capability development.
The comments reinforce the White House’s existing AI policy rather than replacing it.
The administration was already moving in this direction before Sunday’s comments. Its 2025 AI Action Plan centred on accelerating innovation, expanding American AI infrastructure and strengthening US leadership abroad. A June 2026 executive order on advanced AI paired faster adoption with classified benchmarking for advanced cyber capabilities and security controls for frontier-model deployment.
The likely policy split is therefore not regulation versus no regulation. It is regulation that blocks or delays capability growth versus regulation that manages specific harms while keeping development moving.
DIY AI saw the same tension in OpenAI chief scientist Jakub Pachocki’s call for AI slowdowns. A proposed slowdown becomes an operational policy only once someone defines what triggers it, who verifies the trigger, and what actually stops training or deployment. Trump’s remarks suggest the current White House is unlikely to accept a vague precautionary case on its own.
Federal agencies may get faster AI procurement with narrower safety gates
For federal agencies, the practical effect could be a stronger bias towards deploying advanced models first and controlling them through mission-specific rules. Existing national-security policy already directs agencies to reduce unnecessary barriers to adoption and to improve procurement routes for advanced AI from multiple vendors.
That does not make procurement risk-free. Agencies still have to deal with cybersecurity, data access, auditability, contractor accountability and the consequences of autonomous actions. The more consequential the system, the harder it becomes to treat a general model evaluation as sufficient evidence that a deployment is safe.
A useful test for developers selling into government is therefore not “does Washington like AI?” The answer is plainly yes. The useful questions are which actions the model can take, which logs survive after an incident, which permissions can be revoked quickly, and whether independent evaluators can reproduce the vendor’s safety claims.
The biggest policy gap is international coordination.
The most difficult part of the slowdown proposal has always been coordination. A US laboratory that voluntarily slows while a competitor continues scaling creates an obvious incentive problem. The same logic becomes much harder at the state level because the US and China would need credible ways to verify that neither side is quietly gaining capability while the other observes restrictions.
Trump’s framing sharpens that problem. If the administration sees AI leadership as a winner-takes-strategic-advantage contest, an international agreement that deliberately reduces development speed would have to clear a much higher bar than a domestic safety rule. The proposal would need verification strong enough to answer the administration’s central objection: how do you slow without handing the advantage to a rival?
That does not make international safety work impossible. It shifts the likely starting point towards areas where both sides can see a shared downside, such as model-enabled cyber attacks, autonomous weapons controls, incident reporting or verification mechanisms around especially dangerous capabilities.
Developers should not read Trump’s remarks as permission to remove oversight.
There is a risk that “the White House wants faster AI” gets translated into weaker internal controls. That would be a mistake. Political support for capability growth does not remove liability, contractual obligations, security requirements or the basic engineering problem of giving an unreliable system permission to act.
Recent agent incidents show why the operational layer deserves separate attention. DIY AI has already examined OpenAI-linked agents finding unintended external write paths. Those incidents did not prove that present-day models are superintelligent. Still, they showed how tool access, persistent external state and weak permission boundaries can create real consequences without requiring an extinction scenario.
This is also where public discussion is becoming more useful. People can be sceptical of near-term human-extinction forecasts and still worry about concrete issues such as autonomous cyber activity, data-centre power demand, surveillance, labour displacement, and systems acting outside intended permissions. Treating every criticism as the same kind of AI pessimism makes it harder to separate speculative risk from problems that can already be tested.
Investors get a clearer signal on acceleration, but not a free pass on policy risk.
For investors, Trump’s remarks strengthen the case that the federal government will continue to support large-scale AI infrastructure, model development and rapid adoption. That is constructive for frontier labs, chip suppliers, data-centre operators and companies selling AI into government.
The hidden risk is fragmentation. A lighter federal approach can leave more room for states, foreign regulators, courts and individual procurement bodies to impose their own requirements. Companies may therefore face faster federal adoption alongside a more complicated compliance map elsewhere.
Political tone can also change faster than infrastructure cycles. A data centre, model-training cluster or long government contract is a multi-year commitment. Investors should separate the current administration’s willingness to accelerate AI from the durability of any single regulatory posture.
Safety advocates now need testable demands, not only catastrophic forecasts.
Trump’s dismissal creates a messaging problem for AI safety advocates. The strongest warnings have attracted attention, but they also make it easier for opponents to treat the entire safety case as speculation about extreme future scenarios.
A more durable route is to translate safety concerns into verifiable conditions: independent model evaluations, incident disclosure, cyber-capability thresholds, permission limits for autonomous agents, audit logs, deployment stop criteria, and clear accountability when a system causes harm. Those measures can be argued on evidence even when policymakers disagree about the probability of human extinction.
Trump’s comments therefore reduce the political space for a broad AI slowdown, but they do not end the safety debate. They move it towards a harder question: which guardrails can demonstrate enough practical value to survive an administration that believes slowing the race itself is the greater risk?
