What came of the moment Marcus described?
On 20 May, Gary Marcus asked whether generative AI could become “the tech industry’s Vietnam”, and whether public backlash might “get AI to a better place”. He pointed to a specific prize: Donald Trump was, he wrote, “actively considering the kind of AI preflight check” that he and Michelle Rempel Garner had proposed three years earlier, modelled on the FDA. Thirteen days later Trump signed an order. According to the Guardian’s report, it invites developers to hand over models for review on a voluntary basis, as much as 30 days before launch, and its text says it shall not be construed to authorise “a mandatory governmental licensing, preclearance, or permitting requirement”. Britain’s testing also rests on voluntary practice, reached by a different route. Polling from the Ada Lovelace Institute found 82% of UK adults it polled backed mandatory safety testing of AI, while the UK’s AI Security Institute says it is “not seeking to mandate requirements” for frontier developers.
Strategic Insight: Marcus read the moment half right. Trump did sign an order on checking models before release, as Marcus reported he was considering. What it built is close to what Britain already has: voluntary access, agreed with the companies.
Where does the Vietnam analogy hold?
Marcus’s starting point was a post by All In’s Jason Calacanis, who wrote that “AI is Gen Z’s Vietnam: they want nothing to do with it and they are vocally opposing it”, whilst giving the students “small to no chance” of stopping it. The trigger was a run of commencement ceremonies at which at least three speakers, including former Google chief executive Eric Schmidt, were booed for mentioning AI. Marcus suggested Calacanis may be underrating the students’ chances, and that the analogy runs deeper.
His version rests on two parallels. The first is cost: he puts the Vietnam war at “about a trillion in today’s dollars” and sets it against what he calls the “multitrillion dollar investment in AI”. The second is persistence in the face of evidence. The war, he argues, continued for years despite signs it was not working, and AI is “still struggling with hallucinations, unreliability and misalignment” after massive investment. Thinking about the “lack of RoI for most pilot studies” and the reliability problems, he writes, makes him answer yes.
That is a question, not a finding, and Marcus frames it as one. The post puts no figure on the pilot-study claim, and it invites readers to argue the point in the comments. As a piece of rhetoric the analogy is strong. As evidence it is thin: the cost comparison shows scale, not failure, and whether the spending is a mistake depends on returns, and the post offers no figures for them.
| Measure | Figure | Why it matters |
|---|---|---|
| Earliest release review in the signed US order | Up to 30 days, voluntary | No developer is required to submit a model |
| Earlier draft of the order | Reportedly 90 days | The signed text is lighter than the reported draft |
| UK public supporting mandatory safety testing | 82% (33% comfortable with voluntary) | Most respondents back the kind of mandate the US order rules out |
| Models AISI says it has evaluated before release | More than 55 | Britain’s voluntary testing is already established |
Sources: the Guardian, 2 June 2026; Ada Lovelace Institute, Great Expectations (fieldwork September 2025, 1,928 UK adults); AI Security Institute, October 2026.
How did a preflight check become a voluntary review?
The second half of Marcus’s essay is the more testable one. In January he had predicted in Politico that “By the end of 2026, President Trump will have begun to distance himself from the aggressively pro-AI industry policies” of 2025. By May, he wrote, the idea looked like it had legs “far sooner than I expected”, with Trump’s turnaround “perhaps prompted by fears over Mythos”. The Guardian describes Mythos as an Anthropic model with “advanced cybersecurity capabilities”.
The Guardian’s account of what followed partly bears him out. It describes the order as “tacking against” Trump’s earlier deregulatory stance, and it gives federal agencies a role in vetting powerful models before release. But the route there matters. The signing was postponed in late May after, according to the reports the Guardian cites, Mark Zuckerberg, Elon Musk and David Sacks, previously Trump’s AI adviser, pressed him in private calls to change course. On 21 May, the day after Marcus published, Trump said, speaking in the Oval Office: “I didn’t like certain aspects of it, I postponed it.” An earlier version reportedly asked for models 90 days before release. The signed version made the whole review voluntary.
Reality Check: The order also says the US “continues to lead the world in AI” and refuses “to stifle this innovation with overly burdensome regulation”. The Guardian ties the order to fears about models such as Mythos and to hardline Maga supporters pushing for a stricter process. Industry pressure, on the reported accounts, pulled it back part of the way.
Nor was the voluntary review entirely new. The Guardian notes that the Commerce Department’s Center for AI Standards and Innovation already had similar arrangements with OpenAI and Anthropic, and that in May the administration agreed early-model reviews with xAI, Microsoft and Google DeepMind. The order built on an existing practice. It did not create the FDA-style gatekeeper Marcus and Michelle Rempel Garner had argued for.
Has Britain already been through this?
For UK readers, the US outcome looks familiar. The AI Security Institute set out its priorities for frontier developers in October, and it was plain about its method. “Much of how frontier developers manage risk today is voluntary,” it wrote, adding that it has “independently evaluated more than 55 models before release”. Its aim is best practice that “could later inform formal technical standards”, not a licensing regime.
The UK also has detailed evidence of what the public wants. The Ada Lovelace Institute’s Great Expectations briefing, based on a nationally representative poll of 1,928 adults in September 2025, found:
- 89% agree AI products should be held back until proven safe, even at the cost of slower progress.
- 82% back mandatory safety testing, against 33% who are comfortable with voluntary measures.
- 84% worry that, when it comes to AI rules, ministers will favour the technology sector’s needs over the public’s.
- 48% disagree that keeping up with other countries justifies lighter rules, with 38% agreeing.
Critical Context: The 84% figure is the one to keep in mind when reading the US story. The Guardian’s account of private calls from tech leaders that delayed and softened the order describes the kind of influence that UK respondents told Ada they were worried about.
We covered the briefing’s gap between government priorities and public expectations in January. That gap remains: like the US order, the UK’s AI Security Institute relies on voluntary testing.
Where is the British backlash showing up?
Calacanis described a generation that wants “nothing to do with” AI. A recent British protest we reported looks different. In September, more than 100 people protested against Xlinks’ proposed £14bn data campus with battery storage near Great Torrington in north Devon, according to the BBC. Emily Clark, who chairs the local action group, told the BBC: “I’m not actually against AI - I’m against where the location is.” She also complained that the area lay outside the AI growth zones.
Campaigners had asked ministers to pause data centre planning approvals. The BBC reports the government rejected it, and a government spokesperson said fresh data centre proposals have to go through planning, where their effect on communities and the environment is assessed.
Strategic Reality: In Britain, the planning system gives local objection to AI infrastructure a formal process to work through, with local and environmental effects assessed. A booing commencement crowd has no equivalent.
We looked at whether the UK can buy local consent for AI datacentres last week.
Who is affected?
Frontier developers and the firms that build on them. On both sides of the Atlantic, testing before release depends on developers agreeing to it. That keeps product timetables in the developers’ hands, and neither regime requires a government check before a model reaches a UK buyer.
UK organisations deploying AI. The Ada polling shows a public that expects proof of safety before rollout. That expectation applies to whoever puts the system in front of people, whatever the law requires.
Communities and councils near proposed data centres. The Devon case shows where national AI ambitions meet planning committees. The government has declined to pause permissions, so each proposal is argued on its local merits.
The UK government. Its AI Security Institute is pursuing voluntary best practice while 82% of respondents to Ada’s 2025 poll supported mandatory testing.
What remains unresolved?
Whether voluntary testing reaches the models that matter. The US order and the AISI approach both rely on developers choosing to take part. Neither the Guardian’s report nor the AISI blog says what happens to a developer that declines.
Whether Marcus’s cost argument can be settled. The Vietnam comparison turns on AI spending being a mistake. Marcus cites no figure for the returns on pilot studies in this post, and the post leaves that question open.
Whether public opinion turns into political pressure. The Ada figures show strong support for mandatory testing. Support in a poll is not the same as pressure at the ballot box, and the same briefing found 60% think people like them get little say over government decisions.
What to watch
Marcus’s own deadline. He forecast that Trump would distance himself from the 2025 pro-industry stance before 2026 is out. The voluntary order is evidence both ways. A mandatory element added before 31 December would make the forecast look right. Further loosening would make it look wrong.
Uptake of the US review. If major developers publicly submit models under the 30-day framework, voluntary testing is working on its own terms. If the arrangements stay as opaque as the agreement whose details the Guardian says were removed from a federal website, it is not.
AISI’s standards track. The institute hopes its best practice will feed into formal standards in time. A standard with a named body and timetable would move Britain beyond voluntary practice. If the priorities remain open questions, it has not moved.
The Xlinks proposal in north Devon. Xlinks told the BBC the project was “still at a relatively early stage”. How the planning process handles it will show whether local objection can stop AI infrastructure on a site campaigners say lies outside the AI growth zones.
Marcus ended his essay on the hope that “if the people unite, we might actually be able to get AI to a better place”. The US order is what followed the moment Marcus described: a voluntary review, not a gatekeeper, signed after industry leaders lobbied the president directly. In Britain, polling shows the public already backs mandatory testing before release. The AI Security Institute, which runs Britain’s pre-release testing, has said it is “not seeking to mandate requirements” on the labs.
Source and attribution
This analysis responds to Gary Marcus’s essay “Could generative AI turn out to be the tech industry’s Vietnam? And could public backlash lead AI to a better place?”, published on Marcus on AI on 20 May 2026. Details of the US executive order are from the Guardian’s report of 2 June 2026. UK polling is from the Ada Lovelace Institute’s Great Expectations briefing. The AISI material is from its blog post “Towards technical best practice for frontier AI development”, and the Devon protest is from BBC News.
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching whether Britain’s voluntary testing regime turns into a standard before the public’s patience with it runs out. Read more analysis at Insights, or get in touch.