The strongest current evidence is for reducing teacher administrative load — planning, marking, reporting and communication. Claims about direct improvements to learning outcomes from AI tutoring remain contested and depend heavily on implementation quality. Schools generally get more value from applying AI to teacher workload first and to student-facing instruction second.
Where AI is genuinely useful in schools today
The uses with the clearest return are the ones that give teachers time back without touching the instructional decision.
- Lesson preparation. Drafting sequences, differentiating for a known class profile, generating retrieval questions. Teacher edits and approves.
- Marking and feedback. First-pass marking against a rubric, with the teacher moderating. Fastest measurable time recovery in most schools.
- Reporting. Assembling attainment, attendance and behaviour data into report comments and compliance returns.
- Communication. Drafting parent messages, translating them, and routing them across app, SMS and WhatsApp.
- Early warning. Flagging attendance and attainment patterns that predict disengagement early enough to act on.
Where the evidence is weaker than the marketing
Being straight about this matters more than it costs.
- AI tutoring as a substitute for instruction. Results vary enormously by subject, age and how much teacher oversight is retained. Treat headline effect sizes with caution.
- Automated essay scoring at high stakes. Adequate for formative feedback; contested for terminal assessment, particularly across dialects and second-language writers.
- Predictive risk models. Useful as a prompt to look, dangerous as a verdict. A model trained on historical outcomes will reproduce historical inequities unless deliberately checked.
- AI detection tools. False positive rates are high enough that they should never be the sole basis of an academic misconduct finding.
The five risks a school must actually manage
1. Data protection
Student data entering a general-purpose model is often the first thing that goes wrong, and it usually happens through individual staff using consumer tools rather than through procured systems. Policy and provision have to arrive together.
2. Assessment validity
If a task can be completed by a model in seconds, the grade certifies the model, not the student. This is a design problem, not a policing problem. See assessment redesign.
3. Equity
AI capability tracks device access, connectivity and household literacy. Without deliberate design, it widens gaps. Low-bandwidth and SMS-based delivery is not a lesser version of the product for lower-income markets; it is a requirement.
4. Deskilling
If early-career teachers never plan a lesson unaided, planning expertise does not develop. Systems should make the pedagogical reasoning visible rather than hiding it behind a generate button.
5. Accountability
Someone must be answerable for what was taught and how it was graded. The record should show who approved what.
An implementation sequence that works
| Stage | Focus | Typical duration |
|---|---|---|
| 1 | Policy, data protection position and staff acceptable-use agreement | 4–6 weeks |
| 2 | Teacher workload tools: planning, marking, reporting | One term |
| 3 | Observation and CPD, with an explicit developmental framing | One to two terms |
| 4 | Assessment redesign for tasks AI can complete | One academic year |
| 5 | Student-facing tools, with supervision and age-appropriate limits | Ongoing |
Schools that invert this — starting with student-facing AI — almost always end up retrofitting the policy work under pressure.
Questions to ask any AI education vendor
- What instructional model does your output follow, and can I see it?
- Which curriculum standards is content mapped to, and how is mapping verified?
- Is student data used to train models? Under what contract terms?
- Where is data stored and processed, and under which jurisdiction?
- What does the teacher approval record look like?
- What happens to the school’s data at contract termination?
- What is your published false positive rate on any detection or prediction feature?
Regional context
AI adoption in education is not a single global story. Regulatory posture, connectivity, examination culture and teacher supply differ enough that the same product needs a different implementation in each market.
- United States — district procurement, ESSA reporting, FERPA and state-level AI guidance.
- United Kingdom — DfE guidance, Ofsted expectations, UK GDPR and safeguarding duties.
- Nigeria — scale, teacher supply pressure, examination stakes, connectivity variance.
- Ghana — NaCCA standards-based reform, GES reporting, mother-tongue instruction in early grades.
Related reading
For the specific approach Edves takes to instruction, see Pedagogy with AI. For definitions of the terms used throughout, see the education glossary. For original data, see Edves Research.