Exhibit A – Measured Response #6

Welcome back. Last time I left you with a verdict on the e-Assessment Association conference and a question about whose trust we are really protecting. This week I was in Vilnius, Lithuania for the ninth FLIP+ Annual Conference, and I came home with a different kind of verdict. Two of them, in fact, handed down by two very different ‘courts’.

Just over a week ago, on a panel at the eAA, I made a claim I half expected to be challenged: that the thing most likely to stall AI in assessment would not be the technology. It would be governments and regulators. Nobody pushed back hard. Days later, Washington proved the point for me, in the bluntest terms available.

1. The letter from Washington

Over the past week the US government, citing national security, issued an export-control directive ordering Anthropic, the owners of Claude, to suspend access to its two most capable models, Fable 5 and Mythos 5. For every foreign national on earth, including its own non-American staff! The practical effect was that Anthropic switched both models off for everyone, worldwide, overnight, to stay compliant. Its other models kept running. The stated concern was an alleged method for jailbreaking the model’s cyber capabilities; Anthropic disputes the severity, noting the same trick reportedly works on other freely available models.

Set aside who is right. Notice the shape of it. The capability did not fail. A frontier system did exactly what it was built to do, and then a government, one previously so anti-regulation, reached over and took it off the board. That is the future I described from the panel, arriving early. The constraint on AI in assessment was never going to be whether the model could mark the essay accurately. It was always going to be who is allowed to switch it on, and who can switch it off.

Now, back to the scheduled programming and the key points from the FLIP-plus e-assessment community conference in Vilnius, Lithuania…

2. The Netherlands is digitising the real thing

While we in the UK see our highest-stakes school exams remain firmly on paper, the Dutch are mid-way through changing theirs. Niek van den Berk and Emilia Montanus-Roelofs of Cito walked us through a five-year national programme to digitise the central exams across eighteen subjects, run jointly with CvTE and DUO. The aim is exams that are more flexible, more authentic and more accessible, built on a single construction platform rather than the patchwork of processes they are replacing.

The clever part is where they are pointing the AI. Not at the headline act of writing or marking the exam, but at the unglamorous, load-bearing work: checking items, catching the answer that stopped being correct after a last-minute edit, spotting the distractor that quietly became the right one. Their guiding principle, borrowed from Max Planck researchers: “humans think critically first, then let AI play the comparison.” That is ‘re-innovation’ in its quietest form, a genuine leap applied to the dull machinery that actually decides whether a high-stakes exam is safe.

3. The UK is doing the research, not the deployment

To be fair to ourselves, we are not idle. AQA’s Jeanne Marie Ryan presented a careful pilot, built with Imperial College London, putting LLM-generated formative feedback in front of Year 10 English pupils. Three-quarters found it helpful. Teachers were warmer still, but every one of them wanted it supervised; one described watching the output “like a hawk.” The honest findings were the valuable ones: the feedback ran too long, sometimes addressed the wrong thing, and lacked a recognisable “teacher voice.”

That last point is the ‘Mirror Moment’ in miniature. The win is not an AI that passes for a teacher. It is feedback at a level of detail no human could produce by Monday morning, with the teacher still holding the judgement. The same logic ran through Friday afternoon, where AQA’s Cesare Aloisi reframed marking as “assisted judgement” and Cambridge Assessment’s Sanjay Mistry and Collin Andrew Sherriff described moving automarking into digital mocks. All real, all serious. And all, tellingly, aimed at formative work and mock papers rather than the live summative exam that still decides a pupil’s future on paper.

4. Shared and auditable beats proprietary and revocable

The FLIP+ Library made the contrast sharper still. Release 2 of the ‘library’ has grown the shared item bank to more than two thousand items across fifteen countries and twenty organisations, and turned a repository into something closer to a decision-support tool, with semantic search, AI-generated item descriptions and recommendations.

Take translation. The Library pools items in many languages and renders them into one shared tongue, six languages into English so far, eight by the summer, with SMEs now able to hand-edit the output. The engine, shown in the session “Co-Building Auditable AI,” is locally hosted and auditable and importantly built in the open. That is the ‘moat argument’ from an earlier blog turned on its head: shared, auditable infrastructure is harder to take away than any proprietary model.

5. The off switch nobody owns

Carmen Strigel pushed that idea to its conclusion with her session on Tangerine (https://www.tangerinecentral.org/) , the open-source, offline-first platform now used in more than sixty-five countries and over six hundred languages, much of it on ministries’ own servers. It already does AI-assisted scoring of oral reading fluency, in classrooms with barely any connectivity at all.

Set it beside the Claude Fable story and the lesson is hard to miss. One capability vanished worldwide on a single government’s say-so; the other keeps running in the hardest-to-reach classroom precisely because nobody owns the off switch. Data sovereignty used to sound like idealism. This fortnight it started to look like resilience.

The Measured Response

I closed the FLIP+ conference with a session called “The Future is Human: Considerations for Responsible Use in Assessment ,” drawn from the Association of Test Publishers (ATP) white paper on human oversight. The argument is simple. Meaningful human oversight is the mechanism, transparency about it is what turns the mechanism into trust, and innovation without that integrity is just exposure.

So here is the verdict, and Washington and Donald Trump entered it into ‘evidence’ before I could. Two pictures from one eventful fortnight in assessment: a government removing the most capable models from everyone outside its borders, and a sector quietly digitising real exams, building auditable shared libraries, and running open tools where the signal barely reaches. The technology was never the thing in the ‘dock’. The UK is behind not for want of capability or talent, because AQA and Cambridge put paid to that, but for want of the will and regulatory nerve to put AI and digital assessment where the stakes are real. Vilnius further convinced me the capability is ready. Washington reminded me the constraint was never technical at all.

I said it from the eAA panel last week, and I will say it again here with no pleasure. Integrity is the specification, but it is the gatekeeper, whether law makers or regulator, who decides whether the spec ever sees the light of day.

That’s the Measured Response.

Over to you…

Here’s the homework (for the comments section):

If a single directive can switch off the most capable model on earth overnight, is the open, auditable, sovereign route now the safer bet for high-stakes assessment rather than the idealistic one? And for the UK in particular, are we protecting candidates by keeping our highest-stakes exams on paper, or simply mistaking caution for integrity while countries such as the Netherlands get on with it?

Be honest. The sector that leads the next decade will not be the one with the most capable model, because that can be revoked on a Friday afternoon. It could be the one that built something no one else could switch off.

Until next week…

Paul

Note: Anthropic’s own statement on the ‘switch-off’ directive is here: https://www.anthropic.com/news/fable-mythos-access