Does AI Quietly Push Teaching Towards the Bell Curve?

Adam Sturdee
Feb 17
5 min read

Recently, I had the privilege of a searching and generous conversation with colleagues from a leading UK university about the future of AI in teacher development. They were enthusiastic about the potential of transcript-based lesson analysis to deepen professional reflection. They were also clear-eyed about the risks. Their central question was subtle but profound:

If AI works by identifying patterns, does it quietly push us towards the middle of the bell curve?

And if so, what might that mean for inclusive practice, culturally responsive pedagogy, dialogic classrooms, or work with pupils with SEND?

These are exactly the questions we should be asking.

The Bell Curve Concern

Large language models are statistical systems. They learn patterns from vast amounts of text. It is reasonable to worry that what is statistically common could begin to masquerade as what is professionally desirable.

In education, that would be dangerous. Teaching is relational and contextual. What works in a Year 11 revision lesson may not work in a Year 7 philosophical enquiry. What works for a largely neurotypical cohort may not work for a class with significant additional needs. What works in one community may not translate to another. If AI were to flatten this diversity into a single model of “good teaching”, it would fail the profession. But here is the crucial distinction. AI does not automatically enforce an average. It responds to the frame it is given. Normativity does not emerge simply from mathematics. It emerges from design. The risk is not that AI is statistical. The risk is that we design it carelessly.

Reflection Engine, Not Grading Engine

One of the strongest themes in our discussion was the role of judgement. In my own work building Starlight, we have been clear about this from the beginning. The platform is designed as a reflection engine, not a grading engine. That distinction matters far beyond product design. When teachers upload a lesson transcript, they are not submitting themselves to inspection. They are generating a hypothesis about their practice. Outputs are framed as contestable. Feedback points to specific moments in the transcript and invites professional judgement: agree, disagree, nuance. The aim is not to replace coaching conversations, but to make them more specific, timely and scalable. If AI becomes prescriptive, it becomes reductive. If it becomes interrogable, it becomes developmental. The difference is philosophical as much as technical.

Design Choices Matter More Than Algorithms

The university colleagues rightly raised the risk of standardisation. My response was that the safeguard is not primarily technical. It is architectural and cultural.

Three principles are critical.

1. The Lens Must Be Controllable

Schools should be able to define what matters in their context. One school may emphasise adaptive teaching and SEND provision. Another may prioritise dialogic talk or retrieval practice. A multi-academy trust may want coherence across its professional framework. Effective practice cannot be treated as universal. It must be contextualised. AI systems that do not allow for contextual weighting will inevitably drift towards generic norms.

2. Outputs Should Be Treated as Hypotheses

Feedback grounded in transcript evidence is different from feedback grounded in generalities. When an AI tool anchors its observations in specific dialogue, it invites interrogation. “What led you to that conclusion?” is a healthy question. Professional judgement must remain central. The moment AI outputs are treated as verdicts rather than prompts, development narrows.

3. Data Should Support Inquiry, Not Compliance

Quantitative indicators can illuminate patterns. They can also distort behaviour. Reduce teaching to a score and you narrow practice. Introduce leaderboards and you incentivise performance over reflection. This is where governance matters. AI in schools must sit inside thoughtful leadership frameworks. It must support inquiry, not surveillance. It must inform coaching, not replace it. The gravitational pull towards normativity lies less in language modelling and more in metric design.

Ethics, Consent and Trust

Another strand of our discussion focused on audio recording and ethics. In UK schools, transcript analysis typically sits within the lawful basis of public task, with schools acting as data controllers and AI providers operating as processors. That is the legal frame. But legality is not the same as trust. Clear communication with staff. Transparent governance policies. Explicit boundaries around how data is used. These are non-negotiable. AI in education must be both compliant and credible. Leaders who treat governance as an afterthought will quickly lose professional confidence.

The Real Question Beneath the Bell Curve

The deeper question is not simply whether AI normalises teaching. It is whether leaders use AI to narrow or to widen professional thinking. Used carelessly, AI could entrench conformity. Used thoughtfully, it could surface blind spots, broaden perspectives, and make high-quality reflection accessible to every teacher, not just those with regular coaching access. In many schools, sustained instructional coaching is aspirational rather than universal. Transcript-based analysis offers a way to scale reflective practice without multiplying lesson observations. But scale without humility would be dangerous. That is why these conversations with university colleagues matter. They sharpen thinking. They challenge assumptions. They force clarity about purpose.

Shaping AI With the Profession

One of the most encouraging aspects of the discussion was the spirit in which the critique was offered. It was not defensive. It was collaborative. Starlight itself has been shaped through partnership with teachers, senior leaders, trusts and researchers. Our recent work on transcript-based lesson analysis has been accepted for presentation at the BERA Teacher Education and Development Conference 2026, where we will explore how AI-supported reflection can scale coaching without sliding into surveillance or evaluation.

That framing matters.

AI should strengthen professional judgement, not standardise it. It should widen access to reflection, not flatten diversity. It should sit inside leadership, not above it.

The Question We Must Keep Asking

Does AI risk making teaching more “normal”? Yes, it could. If designed carelessly. If governed weakly. If reduced to metrics. But it could also do the opposite. It could help teachers see patterns in their own discourse. It could illuminate unnoticed habits. It could support inclusion by making invisible interactions visible. The outcome is not predetermined by the mathematics. It is determined by intent, architecture and leadership.

AI will shape education. The only question is whether we shape it thoughtfully in return.

Adam Sturdee is a senior leader and co-founder of Starlight, the UK’s teacher-first AI-powered transcript-based coaching platform for educators. His work sits at the intersection of dialogic practice, instructional leadership and responsible AI strategy for schools and trusts.

He will be presenting his research on AI-supported coaching at the BERA Teacher Education and Development Conference 2026: https://www.bera.ac.uk/conference/bera-tean-conference-2026

He is also speaking at the annual gathering of the SOPHIA Network – European Foundation for the Advancement of Philosophy with Children: https://www.sophianetwork.eu