In Plain Sight - 07
In Plain Sight is a weekly blog from Rhys that we will publish every Thursday/Friday. Each edition will be focused on a thorny topic within investing, startups, learning and work and policy.
In Plain Sight is partially a personal attempt to think independently and to avoid leaning too much on AI for answers to our questions. Will we use AI to edit and polish the text? Yes. But more importantly, will we come up with all of the ideas and analysis? Yes.
Here follows Edition 06 - 'The AI pause in schools and universities – what's it really about and is it as significant as it sounds?'.
------
Warning to readers: this covers a lot of ground and doesn't resolve neatly…
Read the last month of headlines together and it looks like a retreat away from technology is taking place schools and universities - particularly in Europe and North America. New York City has imposed a one-year moratorium on generative AI for the roughly 600,000 children in kindergarten through to eighth grade (despite being widely reported as a blanket ban across all school-age education). Los Angeles Unified has gone further, blocking it across every grade on school devices. Norway is paying to bring paper textbooks back into primary classrooms. It seems that the "innovate first, regulate later" instinct that let generative AI wander into classrooms unsupervised is beginning to reverse.
Dora Palfi, co-founder of imagi, a Brighteye portfolio company, pushed back on that framing in our podcast, released yesterday and linked below, and we agree. "The headlines have actually caused more harm than the policies themselves," she said, "because once you look into the detail... you realise they actually haven't banned AI." Her example is New York City itself: in high school and in career and technical education, students can still use it – what's been reported as a blanket ban is a one-year moratorium concentrated on K-8. Crucially, these bans don't take into account the fact that students can use AI in the 7-8 hours per day they aren't in school within systems that don't have guardrails in place. Dora also made an important point about scale: "These really large school districts like New York City or LA are huge organisations. Similarly to how companies work, you can't expect that the largest organisations will be the ones who are the most innovative and risk-taking. At the same time, there's so much adoption and innovation happening everywhere that just doesn't get the headlines." The University of Chicago is a good example of that nuance being lost. The same institution reported as "banning AI" has also given paid access to Claude to most of its students. "But at the same time, they have a bunch of subjects where they don't allow the use of AI," Dora said. "And I actually think that makes sense. We have ways of teaching, testing and assessment to make sure students still learn. If you're afraid that students are not actually acquiring knowledge, then don't make that specific task a take-home essay." Her conclusion is the one we'd draw, too: "Banning AI is (arguably) the lazy way to go about it. Instead, we'll have some very intentional uses of AI, and then these other very intentional places where we just don't use it at all."
That intentionality is the crucial point. One theory we put to Dora, based on what we've heard in the ether, is that teachers themselves are a significant part of what's driving calls to slow adoption – individually adopting AI in the classroom with no district oversight, limited efficacy evidence and a limited shared standard for what "good" looks like. One teacher might use a carefully designed, reviewed tool; the next, in the same school, might lean on an unvetted consumer bot or agent to grade essays or generate worksheets with nobody checking the output. That variance – in quality, in safety, in what a student's actual day-to-day experience of "AI in the classroom" even is – is harder for a district to manage than a single bad vendor product, because there's no one thing to point at and ban. A blanket moratorium is a blunt instrument for that problem: it can't tell the difference between a well-designed, guard-railed use of AI and a well-meaning teacher's unreviewed experiment, so it restricts both.
Dora's response to that idea was more optimistic than our initial take. "I guess it depends a little on the system you're in," she said. "I've also heard from many schools and teachers that are really excited about vibe coding. And they're right to. In the past, we spent so much money on software that does simple things, like getting our kids to choose their lunch – we could actually vibe code that and a lot of other similar stuff that improves the way our schools operate. I understand that with curriculum design and resource design there's a bit more concern, but at the end of the day, there's a lot of opportunity, and I'm confident we'll eventually bring it to life." We believe Dora is also right to flag the irony in where the fear is landing: "It's funny that we're concerned about teachers creating their own resources when for decades we've been talking about personalised learning – and now that it's finally possible, we're suddenly concerned about it. As long as kids learn what they need to learn, it's actually an incredible thing if teachers can make customised exercises for them." It's worth remembering both views: the variance issue is genuine for a local authority trying to guarantee a consistent standard across a thousand classrooms, and the upside Dora describes – teachers finally able to personalise at a level the sector has wanted for decades – is real too.
AI is already running schools – just less so in the classroom
It's worth pausing on how much of this debate skips past the AI that's already deeply embedded in how schools operate day to day, because none of the headline bans touch it at all. Arbor, the UK's most-used school management information system, now runs "Arbor AI" across more than 4,000 schools, and every feature in it is staff-facing rather than student-facing, which is exactly why it sits outside the political fight entirely. "Ask Arbor" answers data questions instantly instead of making office staff dig through spreadsheets – attendance by year group, room availability, which students have birthdays this week. "Auto Absence" listens to parent voicemails, transcribes and translates them (useful for EAL families), and drafts the absence code for a member of staff to approve rather than acting on its own. "Suggest Student Summary" pulls attendance, behaviour, and attainment data together with staff comments into an overview a teacher or leader can scan in seconds. A "Smart Report" called the Ofsted Inspection Companion compiles the statutory data and evidence inspectors ask for into one editable pack. Arbor says the average school saves 70 minutes per staff member, per week.
None of this stuff needed to be covered by a moratorium or a national framework, because every one of these tools keeps a human explicitly in the loop and keeps its output internal to the school. This is a useful contrast to hold against the governance issue driving the classroom bans: it isn't that schools are AI-averse, it's that the AI already running schools successfully was built with guardrails in from day one – a member of staff approves every action, the data never reaches a child directly, and the vendor is accountable to the school rather than the other way round. The bans aren't a rejection of AI in schools; rather, they're a rejection of AI that skips that design choice.
So, to answer the question more directly: is this pushback aimed at teachers and schools building with AI, or at the tools? Neither, exactly. There are really three categories of AI in schools at present – ungoverned teacher experiments, student-facing vendor interfaces, and staff-facing admin systems like Arbor – and only the first two are under any real pressure. The common thread isn't really who's doing the building; it's the absence of guardrails. Districts and local authorities serving hundreds of thousands of children are less worried about AI being used than about AI being used inconsistently, without evaluation, accountability, or any shared floor for safety and quality.
That's the governance question. The bigger question the ban headlines skip past is what a school is actually for, now that knowledge is no longer the scarce resource the whole institution was built around. (I told you we'd cover a lot of ground…!)
Is school about knowledge or skills?
For most of the last century, the honest answer was knowledge transfer at scale. A teacher stood at the front of a room because they held information 30 children didn't have easy access to anywhere else and the school day was organised around getting that information into their heads efficiently and in order. Grades were a proxy for how much of it had landed. This model made sense when information was scarce and expensive to distribute. It arguably makes less sense now that a child can ask a model a question and get a comprehensive, patient, personalised answer, which is simply not feasible in traditional classroom setups.
For many in the education sector, there remains a push to "protect foundational learning" - this is partially based on an implicit belief that knowledge acquired the hard way, through friction, repetition, and a present human, is qualitatively different from knowledge acquired through a chatbot and worth protecting even at a cost. It's a defensible position and it's the same instinct behind Dora's University of Chicago point: try to redesign the assessment, don't ban the tool. It's also only one answer to the underlying question and this is where we want to bring in John Danner's Flourish Schools project, ongoing in Arizona.
Flourish considers the key value drivers in its school model to be one-to-one learning, and "measurable superpowers." The superpowers are grouped into four categories – human, character, cognitive, and entrepreneurial – each broken into six specific, assessable skills. There is a glaring omission in this list: subject knowledge. Facts and procedural fluency still matter, which is exactly what the AI-driven "Foundations" period is for. But Flourish is explicitly building a school where content mastery is the floor, delivered by one designed, human-reviewed AI system and the actual product a family is paying or getting public funding for is character, judgement, collaboration, and the ability to originate and finish a project. That's incidentally also an answer to the governance problem: Foundations is one consistent system used the same way across an entire cohort, which can be contrasted with a school in which teachers are each running their own experiment based on their training, experience and expertise, sourcing their own materials and delivering their own practice.
How does the role of teachers evolve, if delivering knowledge isn't the job?
This contrast brings the role-of-the-teacher question into focus. If, as in the Flourish system, knowledge transfer is handled by a Tier 1 AI system, what is a teacher's expertise in and what becomes of the traditional role?
I'd use the phrase "smart model citizen" to describe a teacher's role in Flourish's system – someone whose value isn't subject mastery but the ability to teach and model values and life skills a bot or agent can't. Whilst there are many similar skills between this role and what's currently expected of teachers, it's still a genuinely different professional identity that requires different training, different skills and different assessment during hiring processes than is currently in place for teachers. An interesting consequence of this is that it makes the job more exposed - a defined pace and a limited curriculum used to let a generalist teacher's gaps in subject knowledge blur behind classroom structure. A student with an infinitely patient, always-available tutor in their pocket will find those gaps immediately. The honest version of this evolved teacher role isn't "knows more than the student"; rather, it's "models how to sit with not knowing, how to try, how to treat other people, and how to keep going when a project gets hard."
There's a second shift in the offing which Flourish considers particularly exciting, notably against fears of ungoverned AI experimentation: Flourish is already using AI to analyse student work and suggest ways to improve its own instruction, with a human reviewing those suggestions today and, eventually, another AI agent testing them directly. This represents a single, monitored feedback loop improving the shared system daily, rather than each teacher independently evolving their practice over time. If the knowledge-delivery layer of a school can improve itself this way without retraining a teacher, this could represent further evidence the knowledge layer was never really the teacher's job to begin with – freeing the profession to specialise in the part of school that can't be automated and shouldn't be left ungoverned either: the relationship, the psyche, the ambition, the life skills, and more.
Whether that discomfort and that reframed identity pulls more professionals from other industries into teaching is one of the more exciting knock-on questions in the sector. A profession measured on judgement and mentorship rather than credentialed subject expertise opens the door to people who'd never have considered teaching under the old definition.
Two paths
So, how are answers to "what is school for?" reflected in attitudes to AI adoption? One group – the US, Norway, much of Western Europe – is protecting the 'traditional' approach: foundational literacy, numeracy, and academic integrity, on the view that school's job is still knowledge, earned through friction and that ungoverned AI use puts that at risk. The other – Poland equipping 12,000 schools with AI Labs, the UAE mandating AI from kindergarten – is building the ‘new vision', implying the view that school's job is producing citizens and workers fluent in the tools of an AI-native economy, with guardrails designed in centrally rather than left to individual classrooms. Neither is obviously wrong. But only one of these choices is placing a bet on teachers' roles in the 2026 context, rather than defending the traditional role.
What does this mean for founders?
The consequences of fragmented bans of specific types of tool are considerable for founders. That fragmentation is more opportunity than obstacle for builders who pick their ground well. Regulatory variation region to region, country to country, means distribution strategy now matters as much as product; for example, Flourish's move from Tennessee to Arizona was a rational bet on how student funding law operates in practice. There's a second, related opportunity in the governance problem itself: regions that can't police thousands of individual teachers' ad hoc AI use are a real market for tools and school models that bake evaluation, consistency, and accountability in from day one, rather than bolting it on after the fact.
So do these bans benefit incumbents over new entrants? In the near term, yes, to some extent – blanket district bans can freeze the largest AI vendors out of the biggest public-school markets, which mostly protects the status quo rather than any single incumbent. As Dora put it, it's not realistic to expect the largest organisations in any sector to be the most innovative or risk-taking ones; that's just as true of a 600,000-student school authority as it is of a large company. A small, single-operator network can do what a large organisation structurally can't and build the guardrails in from the start rather than policing them after the fact. That's usually where a new answer to what school is for gets tested first.
We'd be glad to talk to founders and funders building on that side of this – schools and tools that treat knowledge as AI's job, guardrails as non-negotiable, and character, judgement, and relationship as the human one.
---
We recorded a podcast with Dora Palfi, co-founder of Brighteye portfolio company, imagi, earlier this week. Some of her comments from the pod are included in this article. You can listen to the pod here.
