Governing the Partnership
This paper applies to the classroom an argument made in the California Management Review in October 2026 (Jarrahi, Ritala and Akbarin, 2026). Like No. 3, it names no vendor as an adversary.
النسخة العربية من هذه الورقة ←
Read the full paper
Tell us who you are and the rest of this paper opens straight away. It is free.
Most AI governance asks what the AI must not do. The harder question is what the teacher must keep doing.
This paper applies to the classroom an argument made in the California Management Review: that AI governance "should extend beyond" guardrails "and encompass the orchestration of a symbiotic relationship between human capital and AI implementations" (Jarrahi, Ritala and Akbarin, 2026). The article is written for companies. Its central idea fits schools closely, because teaching is exactly the kind of specialist work in which the professional's judgement is the point.
The first wave of classroom AI policy was defensive, and rightly so. Protect student data. Check for bias. Keep a record. Know which provider sees what. Perspectives No. 3 set out three questions a school must be able to answer about the AI its teachers use: can you see it, can you shape it, can you stand behind it?
Those questions remain necessary. They are not sufficient. The authors observe that most governance frameworks "focus on defensive mechanisms and 'AI guardrails' aimed at risk mitigation and harm reduction". Their alternative is to govern the partnership itself, so that "human strengths complement AI capabilities by preserving human agency" (Jarrahi, Ritala and Akbarin, 2026).
They describe four kinds of human-AI work. One of them, augmented expertise, is collaboration that deepens professional mastery in a specialised field: their examples are lawyers, developers and radiologists, whose tools "support expertise only when professionals can evaluate their outputs in context". Teaching belongs in that company. A teacher reviewing an AI lesson draft is a professional evaluating an output in the context of a particular class, and the governance question is whether that evaluation is real.
A teacher's approval is only worth something if it is a judgement. A click is not a judgement.
The article makes a pointed observation about "human-in-the-loop" policies: governance "must determine where human involvement adds substantive judgment and where it risks becoming merely ceremonial" (Jarrahi, Ritala and Akbarin, 2026).
A partnership between teacher and AI can fail in two ways.
- Under-use is easy to see. Teachers find the governed tool slow or weak and quietly return to whatever is in their pocket. No. 3 described this as shadow use.
- Ceremonial review is harder to see, because from the outside it looks like success. Every draft is approved. Usage climbs. The dashboard is green. Yet the approval may have become a reflex: the AI's version, accepted as it came, week after week.
The article warns against making a professional "a nominal accountability backstop without the information, competence, and authority needed to exercise meaningful control", and against "ritualistic review that adds paperwork but little substantive oversight" (Jarrahi, Ritala and Akbarin, 2026).
If every teacher draws on the same few models, lessons begin to look alike.
Generative models, the authors note, "privilege statistically prevalent patterns" unless their design and use deliberately encourage alternatives (Jarrahi, Ritala and Akbarin, 2026). They cite evidence that generative AI can raise the creativity of individual work while reducing the collective diversity of what a group produces (Doshi and Hauser, 2024), and they warn that "similar AI architectures trained on similar data" can homogenise an organisation's thinking.
Schools should take this seriously. A system where thousands of teachers ask the same few models for the same lesson will tend towards the same lesson. No guardrail catches this, because nothing has gone wrong in any single draft. The loss is in the variety across classrooms: the local example, the unexpected way in, the teacher's own voice.
Two things counter it, and neither is a guardrail
- Ground the AI in the school, not in the average. In EdSmartly, drafts start from the school's own curriculum and grade, follow the school's own content guidance in its own words, and cite their objectives as Verified or Indicative so the teacher can see what the draft rests on.
- Treat the teacher's departure from the draft as the asset it is. The changes a teacher makes are where the class in front of them enters the lesson. Governance should protect that, and learn from it.
The AI a school approved in September may not be the AI drafting its lessons in December.
The article distinguishes AI governance from ordinary IT governance on exactly this point: AI systems can be updated, and "the performance and organizational consequences of these systems also change through deployment, feedback, and human use", so "they require a more adaptive form of governance" (Jarrahi, Ritala and Akbarin, 2026).
For a school the consequence is practical. A policy that names a tool, but not the models behind it, governs something that may already have changed.
What adaptive governance looks like
- Every draft carries its provenance. In EdSmartly, each generation records the provider, the model and the version of the instructions that produced it.
- The engines are checked daily. Every configured provider is tested each morning through the same code that serves teachers, and failures are reported the day they begin.
- Changes are told to the school. School leaders see a dated record of each change to the models behind the school's drafts, in English and Arabic, each with one sentence on why. Providers that receive teaching content only, never student data, are listed on their own line, with the setting a school uses to leave them out. A build check stops a model change reaching production without an entry.
What each component asks of a school, and where EdSmartly stands
The article closes with four components of governing the partnership (Jarrahi, Ritala and Akbarin, 2026). The table below says what each asks of a school, what EdSmartly does today, and what comes next.
| Component | What it asks of a school | In EdSmartly today | Next |
|---|---|---|---|
| Co-learning architecture: mutual learning between people and AI, without extra workload | Do teachers' judgements about drafts flow back into how the AI serves the school? | Teachers rate drafts up or down in one click; school leaders see the totals, not individual ratings | How often teachers edit assessments before approval In development; turning repeated changes into the school's own guidance, an open question we are working on |
| Dynamic role design: preventing skill atrophy as AI takes on more | Do teachers keep practising the craft the AI now assists? | Every draft is a starting point for the teacher, never a finished product | A matter for school policy more than software: see move 4 |
| Decision and enforcement rights: who decides, with accountability not deflected onto "the AI" or a nominal reviewer | Is it clear who decides, and does that person have what they need to decide? | The teacher approves; the approval is recorded with who, when and a fingerprint of the exact text, never the text; the school sets thirteen controls; a school can require approval before printing, download or export | |
| Partnership safeguards: beyond guardrails, protecting against over-reliance on either partner | Can the school tell whether approval is still a judgement? | Every draft begins "AI-generated, not yet reviewed"; an edited assessment no longer matches its approval and reads "not yet reviewed" again | A school-wide view of how often assessments are edited before approval In development; the same view for written drafts once teachers can edit them inside EdSmartly Planned |
The table is honest in both directions. The decision rights are built. The co-learning loop is only partly built. Role design belongs to schools as much as to any product.
The right measures describe the school. None of them grades the teacher.
Most AI dashboards count use: logins, drafts, hours saved. Those numbers show adoption. They say nothing about the partnership. The article recommends "tracking metrics beyond" volume, including "instances where human intuition succeeded despite AI's recommendations", and examining "successful departures from AI suggestions" to see whether they should inform future work (Jarrahi, Ritala and Akbarin, 2026).
For a school, the partnership question is: how often does a teacher's judgement change what the AI proposed? A school where nearly every draft is approved unchanged may be a school where the AI is excellent. It may also be a school where approval has become ceremony. The figure alone cannot say which. A head of school who knows the figure can ask.
There is a line here that we hold as a principle. These figures describe the school, never the individual. There will be no ranking of teachers by how much they edit, no per-teacher score, and no prompt telling a teacher they did not change enough. A measure that grades teachers teaches them to perform for the measure, which is the opposite of protecting their judgement. It would also turn a teacher into the "nominal accountability backstop" the article warns against.
Five moves for heads and ministries
- Ask whether review is judgement or ceremony. Add one question to every AI review: when a teacher approves an AI draft, what does that approval mean, and how would we know if it stopped meaning it?
- Keep the record of review, not the content. Record who approved what, and when, as a fingerprint, never the text. Proportionate records make accountability real without turning review into paperwork.
- Govern the model, not only the tool. Require your provider to tell you when the models behind your teachers' work change, and to keep a dated record of it.
- Keep the craft practised. Make room for teachers to plan some lessons without AI, so the judgement they bring to AI drafts stays sharp. The article cites evidence that AI used without guardrails can harm learning (Bastani et al., 2025); the same care applies to professional skill.
- Measure the school, never grade the teacher. Use school-wide figures to start conversations, not to rank staff, and treat what teachers repeatedly change as knowledge about what your context needs.
The AI proposes; the teacher decides
EdSmartly's first principle has not changed since No. 1: the AI proposes; the teacher decides. Every draft begins as "AI-generated, not yet reviewed". The approval is the teacher's, recorded and checkable. A school that requires approval can block printing, downloading and export until it is given. That is a safeguard, not a lock: copying stays possible, because the teacher is a professional, not a suspect.
This paper adds one commitment to that principle. Governing the partnership means making the teacher's judgement visible to the school without making the teacher watched by it. We think that line can be held, and we are building to hold it.
An invitation
If the question in this paper is one your school or authority is asking, we would welcome the conversation.
References
- Bastani, H. et al. (2025) 'Generative AI without guardrails can harm learning: evidence from high school mathematics', Proceedings of the National Academy of Sciences, 122(26), e2422633122.
- Doshi, A.R. and Hauser, O.P. (2024) 'Generative AI enhances individual creativity but reduces the collective diversity of novel content', Science Advances, 10(28), eadn5290.
- Jarrahi, M.H., Ritala, P. and Akbarin, M. (2026) 'Governing AI through human-AI symbiosis', California Management Review, Insights, 12 October. Available at: cmr.berkeley.edu (Accessed: 17 October 2026).
Contact
hello@edsmartly.com · EdSmartly.com · app.edsmartly.com
EdSmartly is a service of Axiom Intelligence Partners LLC.
This document may be shared freely in unmodified form.