Four days in June. Tuesday, , its most powerful publicÌýartificial intelligence model. Wednesday, its CEO, Dario Amodei, publishes a calling for government oversight of frontier AI. By Friday, the US government to that model for foreign nationals, citing national security concerns, forcing Anthropic to take it offline.

Amodei’s essay is arguably the most substantive AI policy document a frontier AI lab’s CEO has published. His central argument is exactly right: AI moves in weeks while policy moves in years, and the mismatch is the defining governance challenge of our moment.

But his solutions are offered at the level of nation-states and regulatory bodies: Federal Aviation Administration-style model testing, export controls, democratic coalitions. He has nothing to say about what happens inside universities, 94 per cent of whose staff use AI tools for work, according to EDUCAUSE’s of nearly 2,000 US higher education staff, while 46 per cent were unaware of any institutional policy governing that use – including 38 per cent of executive leaders. In many cases, the policies simply don’t exist.

Andrew Maynard, a professor in the School for the Future of Innovation in Society at Arizona State University, what falling behind looks like institutionally: governance working from a mental model of AI lagging where practitioners already are.

É«ºÐÖ±²¥

Three clocks run at permanently mismatched speeds. The AI development clock moves in weeks: capability shifts that make Tuesday’s assumptions obsolete by Friday and render an assessment task judged AI-resistant in semester one demonstrably not by semester two.

The institution clock moves in committee, semester and procurement cycles. A working group that spends 14 months on an assessment policy is doing exactly what institutional governance is built to do: make robust decisions. What has changed is the environment that trade-off assumed, in which the things being governed moved slowly enough to wait for.

É«ºÐÖ±²¥

The practitioner clock runs in the gap between the other two: in the daily decisions of course coordinators, educators and professional staff about whether and how to use the latest AI upgrade. The who are using AI tools their institutions haven’t provided, evaluated or sanctioned are not being reckless. The tools work, the work needs doing, and the approved alternatives either don’t exist or can’t do what the unapproved ones can.

These three clocks cannot be synchronised – call it the three-clock problem. Speeding up governance sacrifices the deliberation that makes its decisions legitimate. Suppressing the practitioner clock loses the most honest signal the institution has. And the development clock certainly won’t slow for a working group. Coherence across permanently mismatched speeds calls for a different kind of movement, not a faster one.

Specifically, committees need better, more up-to-date intelligence on what is happening – at both the developmental and practitioner coalfaces. Between meetings, a committee has no eyes – and no reform of committee protocols closes that gap because the only one that would do so – meeting continuously – would stop it being deliberative. But they can at least make sure that they are deliberating over this year’s issues, rather than last year’s, if they can build a dedicated sensing and translation function to monitor AI developments and synthesise what practitioners are actually doing internally.

That’s the ask, and it’s small: name someone – a role, or a small team – and give them a standing channel into governance, rather than relying on an annual survey. That person or team can also translate committee decisions back down so practitioners aren’t working in a vacuum. The loop runs both ways.

É«ºÐÖ±²¥

None of this bypasses or shortens committee deliberation. The committee itself would set the conditions under which it must be told about developments, such that the channel holds no authority the institution didn’t grant it; and the committee retains every decision-making power it had before. What changes is the input, not the clock speed – and it is a principle that other committees could adopt too: intelligence that is advisory in content but mandatory to consider.

Universities run this logic internally all the time. A paper reaches a committee without legal sign-off and comes back procedurally incomplete: the advice doesn’t bind the decision, but the decision can’t be taken without it.

It’s worth knowing what it looks like when the channel exists. It isn’t a chief AI officer in the boardroom, a training programme, or a policy refresh. At RMIT University, it has involved the deputy vice-chancellor education, Sherman Young, presenting institutional AI strategy directly to a 900+ member community of practice spanning three continents and taking questions from the floor. Nothing was decided faster that way, but it meant that the people making the decisions about the strategy’s evolution knew more about the institution they were deciding for.

None of which stops Washington DC from imposing overnight bans on specific technologies, of course. No institutional sensing capacity can affect whether the shock lands. But it can minimise the blast radius if the committee knows about it straight away – and also knows which courses were built on that model, which assessments assumed its capability, what the fallback is and how long a migration takes.

É«ºÐÖ±²¥

Most universities couldn’t have answered those questions on that June Friday. And while , the fact that they still can’t should worry all vice-chancellors.

The answer is architectural and internal, but for most institutions, it still hasn’t been built.

É«ºÐÖ±²¥

Nick McIntoshÌýis a Google higher education faculty AI fellow and a learning futurist at RMIT University Vietnam.Ìý