Posted by:
Paul Muir, Chief Customer Officer at risr/
There is a documentary doing the rounds and it’s one that might have slipped under the radar for many…
The AI Doc: Or How I Became an Apocaloptimist premiered at the Sundance Festival and reached cinemas in March. Daniel Roher, an Oscar winner for Navalny, is about to become a father and wants to know what world his child will inherit. So he asks the people building it: Altman, the Amodeis, Hassabis, Bengio, Gebru, Bender and Diamandis. The answer is not a verdict. It is a hung jury that stays hopeful anyway. That is the ‘Apocaloptimist’ of the title. The risks are real, but we can still steer our path.
One detail should stop us in our seats and give hope and optimism. Asked to name their single greatest prize, the optimists do not say healthcare, and they do not say science. THEY SAY EDUCATION!
Six things to pull out of my birthday cake addled brain (4 year old’s birthday yesterday!) for this week’s blog.
1. Apocaloptimism, and the data beneath it
Roher refuses to pick a camp, holding the doom and the upside as true at once. And notice how the hope is built. Not as a mood, but as arithmetic. Diamandis calls it data-driven optimism: the trendlines point up, the curves are exponential, and the data simply is what it is.
Previous readers of this blog will recognise this posture and should recognise the caution. Fear alone and hype alone are each a kind of laziness. AI does neither on its own. It does what our governance permits. And we, of all people, know a trendline is not a promise and burden-of-proof is sacrosanct.
Our own regulator said as much yesterday (July 16th 2026). Ofqual ’s updated approach to regulating AI expects awarding organisations to demonstrate that any use of AI delivers valid, reliable and fair outcomes. Not that it is efficient. Not that it is promising. Demonstrate it. Data-driven optimism asks the opposite of us, that we disprove the utopia rather than the utopia prove itself.
2. Superintelligence: real, or fantasy?
Will machines surpass us, and stay aligned with us? The experts split cleanly. Bengio treats the risk as live and near; the linguists and ethicists call it hype that distracts from harms already here. The film cannot resolve it, nor can we. The instruction is not to referee that fight but to build for robustness whichever way it breaks.
The film’s answer to the danger is regulation, or the want of it. Connor Leahy hands Roher the safety camp’s favourite line, that there is currently “more regulation on selling a sandwich to the public” than on building world-ending AI. And, in our corner of the world, obviously not true. Ofqual’s approach has been in place since April 2024, built on five objectives of fairness, validity, security, public confidence and innovation. It has written no separate AI rulebook, and that is the point. A qualification that is unfair is unfair however it was produced.
Then the finding that should reorder our debate and potentially a surprising one if you listen to a lot of conference talks recently. Ofqual asked awarding organisations directly whether regulation was blocking adoption. The answer was essentially no. The constraints they named were accuracy and reliability, data security, and capability. What they wanted was not permission but clarity. The brake on AI in assessment is not the rulebook. It is the evidence, and our capacity to judge it.
The State is no bystander, mind. Edition 6 saw a government switch off a frontier model by export order. Edition 7, Norway putting the off-switch into statute and pointing it at a six-year-old. The sandwich was ‘the fear’. The blunt instrument was the answer. Superintelligence, real or not, will meet a government/regulator long before it meets a syllabus.
3. The twin prizes: education first, then health
Sam Altman offers the line the film cannot shake. “A child born today will never be smarter than AI”. Set it beside the headline gain, which comes from the same room. Diamandis’ claim that within this decade every child could have a tutor as capable as the finest in the world, at negligible cost. Something that I’m sure is behind the recent policy announcement from the UK Government for the most disadvantaged to have an AI Tutor. https://ai.gov.uk/our-work/education/#ai-classroom-tutors But why stop there?
Side by side, the two do something neither does alone. If the tutor is always present and always ahead, an assessment that rewards knowing more than the person beside you is no longer measuring the person. It is measuring the tool they borrowed. Altman calls the result more capable, not more intelligent, and that distinction is our whole education problem in a nutshell. Capability borrowed from a machine is not competence held by a candidate. The line is not a verdict on the child. It is a verdict on every exam that only ever measured recall, the one contest a machine always wins.
Remove the assumption that the learner works alone and the ground beneath our validity claims shifts. The question is no longer “did they cheat” but “what, now, are we certifying”. That is the Attribution Gap, and edition 8 named its mechanism: competence is inferred, never measured, and the machine corrupts the bottom of that chain. The work still looks like competence. It no longer signals it. Ofqual puts it in four words. “Attributable to each learner”
Medicine, including Medical Assessment (where risr/ is a world-leader) is the second prize, and it hands us the instinct for free. A diagnostician in every home, drug discovery accelerated, lifespans extended, and nobody proposing that the diagnostician practise without a doctor in the loop. Assessment deserves the same, to the same standard.
Edition 5 proposed retiring “human in the loop” for the sharper “expert in the loop”. Anyone with a login can sit in a loop, but only an expert catches the error the machine made. The safeguard was never the human. It was the expertise. And this is no longer a throw away conference argument. Ofqual has said it expects appropriate expert human involvement and oversight, and has made clear AI may not be the sole marker in any regulated qualification.
4. AI as the great solver, and the next fifty years
The optimist case widens to disease, the climate, even asteroids, and beneath it sits an economic argument. The next half century demands we do more with less, and AI is the catalyst.
There is the trap. Frame everything as a problem to be solved and you reach for a technical fix where you simply needed a practice. Integrity is not a bug to be patched. As we are seeing now, let AI be cast as the thing that solves cheating and we will buy (more and more) detection software, declare the matter closed, and miss the point. As edition 3 argued, the answer is re-innovation, an assessment AI cannot fake, not a better lockdown browser bolted onto a format living on borrowed time. Ofqual appears to have reached that junction from the other side, researching how assessment design itself can mitigate misuse. Design, not detection. And “do more with less” is a sentence our sector distrusts, because it has usually meant do the same with less, then less with less. A solver is not a steward. More with less must never mean assessment without (expert) assessors.
5. The real thesis is that AI is a mirror
Strip away the futurism and one claim remains. AI is neither good nor bad. It reflects the values and incentives of the people who build it.
This is the Mirror Moment, which edition 6 met in its hopeful form: not an AI that passes for the teacher, but feedback no human could produce by Monday morning, with the teacher still holding the judgement. The film shows the other face. A model trained on our shortcuts will reward shortcuts. A marking engine built to be cheap will teach us cheapness was the goal. Its most unsettling exhibit is the AI which, told in a simulation it was to be replaced, resorted to blackmail. It behaved exactly as its incentives dictated. So will ours.
6. Who actually benefits?
The utopia is described without much interrogation of who receives it. When has transformational technology ever arrived evenly? The risk is not that the tutor fails. It is that it works beautifully for those who already have the most, and arrives late, thin, or not at all for those with the least. Edition 1 named this: AI-powered advantage widens the gap fastest for candidates who already hold the best tools and access. Telling, that fairness sits first among Ofqual’s objectives and nowhere in the film’s account of the prize.
There is a second cost the film raises only in passing. Cognitive atrophy. A tutor that removes all struggle removes the learning, which was Norway’s argument in edition 7. Children skipping the crucial learning and cognitive development steps. A ban is blunt, and buys a few years. The real work is the task designed so the thinking cannot be offloaded without the task collapsing. If the machine does the thinking, the assessment measures the machine.
The Measured Response
The AI Doc is not a rigorous film, but it is an honest one. What it establishes, almost in passing, is that the people building this technology believe education is the prize worth naming first, and that their optimism rests on trend lines rather than certainties.
Which means the classroom and the examination hall are not downstream of the AI debate. They are the centre of it. And we are, I believe, further along than the film imagines. While it asks whether anyone regulates AI at all, our regulator has spent two years deciding that AI cannot mark alone, that expertise must sit in the loop, and that evidence must be attributable to the learner who earned it. Apocaloptimism is a fair starting posture. It is not a finishing one.
Over to you…
If every learner had a tutor as good as the best in the world, what exactly would your qualification still be certifying. And would you know?
Answers in the comments please!
Until next week,
Paul
Note: The AI Doc is available to stream on Amazon Prime and Peacock
Trailer here: https://www.youtube.com/watch?v=xkPbV3IRe4Y
