What Happened
Two sharply different visions for the future of AI collided this week. Anthropic, the San Francisco-based AI safety company behind the Claude family of models, raised the possibility of a coordinated, voluntary pause on frontier AI development. At almost the same moment, US President Donald Trump signed a national security memorandum ordering the opposite: faster, broader deployment of AI across the US military and intelligence apparatus.
Trump's memorandum stated: "Under my Administration, the United States can and will responsibly accelerate the use of AI across intelligence and war fighting domains in line with American values." The directive gave US Defence Secretary Pete Hegseth 90 days to update an existing directive on the autonomy of weapons systems "to ensure the deliberate adoption of AI systems that respect the chain of command." The administration also asked leading AI developers to voluntarily submit their most capable models for government cybersecurity testing before public release, and stressed the technology must not be used "to censor free speech ... or conduct unauthorised or unlawful surveillance activities."
Anthropic's position sits in direct tension with that acceleration. The company floated "the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology." It acknowledged the practical difficulty, noting that any such pause "would also require that each can verify that the others have actually stopped."
For Australian organisations working through their own AI strategy, this week's events are a useful stress test: the gap between safety-first thinking and speed-first policy is now very public, and very wide.
Why It Matters
The tension here is not abstract. Anthropic disclosed that as of May 2026, more than 80 per cent of the code merged into its own coding system was authored by Claude. That is a striking figure - an AI company's core product is now writing most of its own infrastructure. The company was candid that "the Claude-written code was still worse in quality than human-written code at Anthropic in late 2025," but the trajectory is clear.
Anthropic described the logical endpoint of that trajectory as "an AI system capable of fully autonomously designing and developing its own successor." Its assessment of that prospect was measured rather than dismissive: "These trends have huge implications. AI that can build itself would be a major development in the history of technology - one that could bring enormous good for the world in science, healthcare, and beyond." The implied concern is that the same capability could also bring enormous harm if human oversight erodes before the necessary safeguards are in place.
For Australian businesses, the practical stakes are immediate. Organisations that adopt AI systems without structured governance frameworks risk being caught between a US policy environment pushing for speed and a safety research community warning that speed without oversight is the core problem. Our AI automations and AI training work consistently shows that the organisations that move well - not just fast - are the ones that build governance in from the start.
Key Details
Trump's national security memorandum covers several distinct areas. On the military side, it directs acceleration of AI use across intelligence collection and warfighting, with the chain-of-command constraint built in as a guardrail. On the civilian side, it asks AI developers to submit frontier models for government cybersecurity testing before public release - a voluntary measure, not a legal requirement.
Anthropic's freeze proposal is also voluntary in character. The company acknowledged the verification problem directly: a coordinated pause only works if every major developer actually stops, and confirming that is technically and politically difficult. The proposal is framed as an option worth having available, not as an immediate demand.
The self-coding data point deserves attention on its own terms. More than 80 per cent of code merged into Anthropic's coding system coming from Claude as of May 2026 is a concrete, measurable sign of how quickly AI is taking over technical work - even at the organisations most focused on AI safety. The quality gap noted for late 2025 may close faster than most organisations are planning for.
Australian government and public sector bodies watching the US military AI push should note that the chain-of-command requirement is the key governance mechanism the Trump administration is relying on. Whether that is sufficient is precisely what Anthropic is questioning.
Background and Context
Anthropic was founded by former OpenAI researchers with an explicit focus on AI safety. Its public communications have consistently argued that the pace of AI development is outrunning the social and technical infrastructure needed to manage it safely. The freeze proposal is consistent with that long-standing position, though it represents a more concrete policy ask than the company has made previously.
The Trump administration's approach reflects a different set of priorities. The US is in active competition with China on AI capability, and the national security memorandum is partly a response to that competitive pressure. Accelerating military AI adoption is framed as a matter of strategic necessity, with safety addressed through the chain-of-command requirement rather than through development slowdowns.
The two positions are not entirely incompatible - it is possible to accelerate deployment in some domains while pausing capability development in others - but the practical and political distance between them is significant. The verification problem Anthropic raised is real: a voluntary pause that one major developer ignores is not a pause at all.
For context on how Australian organisations can build structured approaches to AI adoption that account for both the opportunity and the risk, see our AI strategy services and editorial standards.
What Comes Next
Hegseth's 90-day deadline to update the weapons autonomy directive is the most concrete near-term milestone. That update will set the formal US military standard for how much autonomous decision-making AI systems are permitted in warfighting contexts - a standard that will influence procurement, development, and deployment decisions globally, including by Australian defence contractors and allied governments.
Anthropic's freeze proposal will likely generate formal responses from other major AI developers. Whether any coordinated mechanism emerges - and whether it could be verified - remains to be seen. The company's own disclosure that Claude is writing more than 80 per cent of its code suggests the window for a meaningful pause may be narrowing faster than the policy debate is moving.
Australian organisations should treat this week as a prompt to review their own AI governance frameworks. The question is not whether to adopt AI - that ship has sailed - but whether the structures around adoption are robust enough to hold up as the technology continues to accelerate. Our AI training programmes are designed to help teams build exactly that kind of structured capability.
For more on how Mindiam approaches these questions, see our about page and authors/ben-rogers.
Frequently Asked Questions
What is Anthropic proposing with its AI freeze call?
Anthropic raised the option of a coordinated, voluntary slowdown or temporary pause on frontier AI development. The stated purpose is to give societal structures and AI alignment research time to catch up with the pace of technological change. The company acknowledged that any such pause would require participating organisations to verify that the others had actually stopped - a significant practical and political challenge that has not yet been resolved.
What did Trump's national security memorandum actually direct?
The memorandum directed the US military and intelligence community to accelerate the adoption of AI across warfighting and intelligence domains, in line with what the administration described as American values. It gave US Defence Secretary Pete Hegseth 90 days to update the existing directive on weapons autonomy to ensure AI systems respect the chain of command. The memorandum also prohibited use of AI for unlawful surveillance or censorship of free speech, and asked AI developers to voluntarily submit frontier models for government cybersecurity testing before public release.
Why is the 80 per cent code-authorship figure significant?
As of May 2026, more than 80 per cent of the code merged into Anthropic's own coding system was written by Claude, Anthropic's AI chatbot. This is significant because it shows that even the organisation most focused on AI safety is already operating in an environment where AI is doing the majority of its own technical development work. Anthropic noted that the Claude-written code was still worse in quality than human-written code as of late 2025, but the direction of travel is clear and the quality gap is likely to close.
What does self-improving AI mean in practical terms?
Anthropic described the concern as the prospect of "an AI system capable of fully autonomously designing and developing its own successor." If an AI system can design and build a more capable version of itself without meaningful human input, the pace of capability growth could accelerate beyond what human oversight structures can manage. Anthropic described this as potentially bringing "enormous good for the world in science, healthcare, and beyond," while also raising serious questions about whether humans would retain meaningful control over the process.
What should Australian organisations take from this?
The core lesson is that the gap between safety-first and speed-first approaches to AI is now a live policy debate at the highest levels of government and industry. Australian organisations building AI capability need governance frameworks that can hold up under both pressures - the commercial imperative to move quickly and the safety imperative to move carefully. Structured AI strategy, clear accountability, and ongoing staff training are the practical tools for navigating that tension.
