The strongest current evidence is for reducing teacher administrative load — planning, grading, reporting and family communication. Claims about direct learning gains from AI tutoring remain contested and depend heavily on implementation quality. Districts generally get more value applying AI to teacher workload first and 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 against a standard, differentiating for a known section, generating retrieval questions. Teacher edits and approves.
- Grading and feedback. First-pass scoring against a rubric with the teacher moderating. Usually the fastest measurable time recovery in a building.
- Reporting. Assembling attainment, attendance and behavior data into report card comments and state compliance returns.
- Family communication. Drafting messages, translating them, and routing across app, email, text and voice.
- Early warning. Flagging attendance and grade patterns that predict disengagement early enough to intervene.
Where the evidence is thinner than the marketing
Being straight about this matters more than it costs.
- AI tutoring as a substitute for instruction. Results vary enormously by subject, grade band and how much teacher oversight is retained. Treat headline effect sizes with caution and ask what the comparison condition was.
- Automated essay scoring at high stakes. Adequate for formative feedback; contested for accountability testing, particularly across dialects and multilingual learners.
- 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 audited, and in a system with documented disproportionality in discipline and special education identification, that is not an abstract risk.
- AI detection tools. False positive rates are high enough that they should never be the sole basis of an academic integrity finding. Several districts have reversed findings after the fact.
Five risks a district must actually manage
1. Student data leaving the district
The first thing that goes wrong is almost never the procured system. It is individual staff pasting student work, IEP text or behavior records into a consumer chatbot with no contract behind it. Policy and sanctioned provision have to arrive together — a ban without an approved alternative produces shadow use, not compliance. See FERPA and student data privacy.
2. Assessment validity
If a task can be completed by a model in thirty seconds, the grade certifies the model. This is a design problem, not a policing problem.
3. Equity
AI capability tracks device access, home connectivity and adult literacy support. Without deliberate design it widens gaps. Any tool that only works well with a modern device and reliable home broadband will under-serve exactly the students whose outcomes the district is most accountable for.
4. Deskilling
If a first-year teacher never plans a lesson unaided, planning expertise does not develop. Systems should make the pedagogical reasoning visible rather than hiding it behind a generate button — which is a large part of why Pedagogy with AI shows the model and the standard rather than just the output.
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 survives a privacy review
| Stage | Focus | Typical duration |
|---|---|---|
| 1 | Acceptable use policy, data privacy position, staff agreement, board briefing | 4–8 weeks |
| 2 | Teacher workload tools: planning, grading, reporting, communication | One semester |
| 3 | Observation and professional learning, framed explicitly as developmental | One to two semesters |
| 4 | Assessment redesign for tasks a model can complete | One school year |
| 5 | Student-facing tools with supervision and age-appropriate limits | Ongoing |
Districts that invert this — starting with student-facing AI — almost always spend the following year retrofitting governance under pressure, frequently after an incident.
Questions to ask any AI education vendor
- What instructional model does your output follow, and can I see it?
- Which state standards is content mapped to, and how is the mapping verified and updated when the state revises?
- Is student data used to train models? Show me the contract language, not the marketing page.
- Where is data stored and processed, and by which subprocessors?
- What does the teacher approval record look like?
- What happens to our data at contract termination, and what does export cost?
- What is your published false positive rate on any detection or prediction feature?
- Have you signed the student privacy pledge or an equivalent commitment, and will you sign our state’s data privacy agreement?
The last question filters more vendors than the rest combined.
The policy landscape
Federal guidance on AI in education has been advisory rather than regulatory, which leaves states and districts setting their own rules. A growing number of state education agencies have issued AI guidance for districts, and the picture changes frequently — check your own state agency’s current position rather than relying on any vendor’s summary, including this one.
What has not changed is the underlying law. FERPA, COPPA, PPRA and state student privacy statutes apply to AI systems exactly as they apply to any other vendor, and “it is AI” is not a category exemption.
Related reading
Pedagogy with AI for the instructional approach. FERPA and student data privacy for the compliance detail. Education glossary for definitions. Higher education for assessment redesign.