AI is starting to appear in teacher contracts, with provisions on job protection, instructional use, and in some cases evaluation and discipline. The next step is more specific language about how AI may be used during observations and in evaluation write-ups. Districts should agree on those rules with their teachers' association before a pilot, rather than wait for a dispute. Many of the protections that belong in an agreement also belong in the design of the tool.
A disclosure: I'm the CTO at Tandem, which builds an AI observation tool. These are the safeguards I would want districts and teachers' associations to ask of us, too.
Where AI contract language stands today
Early agreements show several approaches to AI: protecting jobs, preserving professional judgment, and giving educators a role in adoption. In St. Paul, according to the AFT, the local's clearest win “came around evaluation and discipline”: educators “cannot be disciplined, involuntarily transferred or non-renewed based solely on AI-generated data or metrics.” In July 2026, Detroit's new teachers' contract stated that AI cannot replace teacher professional judgment, barred AI from student grading, discipline, and IEP decisions, and required the district to publish a list of approved AI tools. Those are rules for instructional use, not for teacher evaluation.
Education Week's August 2025 survey of AI in labor contracts found other models. Rockdale District 84 in Illinois has a four-person union-district committee, set up in its 2024–28 agreement, that advises the school board on AI adoption and teacher training. In 2025, negotiations in Ithaca, New York, stalled partly over whether AI could replace staff. At its July 2025 convention, the NEA voted to develop model bargaining language on AI. The same report cited RAND's finding that little more than one in ten districts had policies on teachers' use of AI, with fewer addressing AI in hiring, training, and evaluating educators.
St. Paul's provision governs what happens after an evaluation. None of the provisions described here says how AI may be used while the evaluation is being written, and that is where clearer rules are needed.
Why evaluation needs its own rules
Some administrators already use AI in evaluation. In March 2026, GovTech reported that the practice is “emerging in pockets of K-12 systems, often at the discretion of individual administrators,” though its scope is hard to quantify. One New York assistant principal described uploading observation notes and the Danielson rubric to ChatGPT and asking it to “suggest a rating and give aligned feedback,” with no set policy governing that use. Rob Weil, CEO of the AFT-founded National Academy for AI Instruction, called AI for evaluation premature and said “using it without transparency is a real problem.”
Where AI-specific rules are missing, educators and administrators are left to interpret existing evaluation procedures without clear guidance about the new tools. We wrote about the FERPA risk of consumer AI in evaluations in February 2026. A negotiated agreement can specify which tools are permitted, what safeguards apply, and how their use is documented.
What a good AI evaluation agreement covers
These are proposals for a district and its association to work from, drawn from the contracts above and from how we think observation tools should be built. Most go beyond what the cited contracts contain. They aren't legal language, and your counsel and bargaining team will shape the final wording.
A person makes every judgment
AI does not independently determine ratings, evaluation conclusions, or employment actions. The evaluator forms and documents their own judgment from verified evidence and remains responsible for the final evaluation. The tool does not prefill ratings before the evaluator records their own assessment, and any rubric connections or draft language it suggests are identified as suggestions and checked against their sources.
St. Paul provides one starting point: AI-generated data alone cannot justify specified adverse employment actions. The language proposed here goes further by addressing how judgments are formed during the evaluation itself.
Educators know before the observation
Before an observation, the educator is told which AI tool will be used, what it captures, whether the observation is for coaching or formal evaluation, who can access the resulting records, and how long those records are kept. The agreement also spells out any notice or consent procedures that apply. Notice to the teacher doesn't settle questions about capturing students' speech; those need their own review with counsel.
Educators can see and correct the evidence
The educator can review the transcript, source evidence, and AI-generated material relied on in preparing the evaluation, with appropriate protections for student information. They can flag errors, add missing context, and submit a written response that stays attached to the evaluation record. The agreement says when that review happens and how disputed evidence is handled before the evaluation is final.
Observation data has defined uses
The agreement defines the permitted uses of observation data and prohibits unrelated reuse, including model training, employee profiling, and unrelated monitoring. Audio retention is minimized, and if the workflow doesn't retain audio, that commitment covers the vendor and every provider it uses. Our recommended default is a workflow that doesn't retain classroom audio at all. Transcripts and AI-generated material follow a defined retention schedule, with access limited to authorized people and exceptions spelled out for required retention or pending disputes. This protects students as well as teachers, because transcripts may contain student names, statements, and sensitive information.
Only approved tools
Evaluators use only district-approved tools covered by contractual privacy and security commitments. Classroom audio, transcripts, observation notes, and evaluation records don't go into unapproved AI services. Detroit's published list of approved tools is a useful model for visibility, even though its contract addresses instructional use rather than evaluation.
Joint review and a way to enforce it
A labor-management committee reviews the proposed evaluation use before adoption, every year, and whenever the tool, its data practices, or its role in evaluation changes significantly. The agreement says what information the committee receives, how it makes recommendations, and how unresolved concerns are escalated. Rockdale District 84's committee is a precedent for joint oversight of AI adoption; the review schedule and evaluation-specific role proposed here would be additions.
The agreement also says how unauthorized AI use, disputed evidence, and failures to follow the notice or access terms are handled through the district's evaluation and dispute-resolution procedures. Without that, a teacher has no recourse when a safeguard is skipped.
What this means for vendors
For a vendor, these terms are a useful test of the product. A vendor should be able to explain how its tool supports each commitment, and where district procedures or contract terms are needed to make it enforceable. Some protections depend on product design: whether ratings are prefilled, whether evidence is visible and correctable, whether audio is retained. Others, like notice and dispute handling, depend on district practice, and the agreement should name both.
Transcription can help protect teachers by grounding feedback in a reviewable account of what was said. That claim is more credible when teachers helped shape the rules and can inspect and challenge the evidence. Involving them gives a district a stronger basis for trust than asking them to accept a tool after it has already been chosen.
Where to start
Check whether your current agreement says anything about AI in observation or evaluation. If it doesn't, raise it with your association before piloting a tool for either. Bring the vendor's written answers about data handling, ratings, and evidence access to that conversation, because most of them map directly onto the terms above.
If you'd like to talk through how an AI observation tool can work within the terms your teachers need, get in touch.
Frequently asked questions
Can teachers' unions bargain over AI in teacher evaluation?
Whether a particular AI-related evaluation proposal is negotiable depends on state law, the bargaining relationship, and the current agreement. Unions are already negotiating AI protections: according to the AFT, St. Paul educators won language prohibiting discipline, involuntary transfer, or non-renewal based solely on AI-generated data or metrics. Districts and associations should review proposed terms with counsel.
What should AI language in a teacher contract cover?
Proposed terms should address the evaluator's responsibility for independent judgment; advance notice of AI use; access to and correction of evidence; approved tools; permitted data uses and retention; joint review of adoption and significant changes; and procedures for disputed evidence or unauthorized use. The final provisions should fit the district's evaluation process and applicable law.
Do districts have policies on AI in teacher evaluation?
Policies are still developing, and there is no current national figure. In August 2025, Education Week reported RAND's finding that little more than one in ten districts had policies on teachers' AI use, with fewer addressing AI in hiring, training, and evaluating educators. In March 2026, GovTech described some administrators using AI in observations without formal policies and noted that the extent of the practice was difficult to quantify.