Your Principals Are Already Using AI to Write Teacher Evaluations. Here’s Why That Should Worry You.

It's happening more often than most districts realize. Administrators are sitting in a classroom, taking observation notes, and then pasting those notes into ChatGPT or Google Gemini to help draft their teacher evaluations.

They're not doing it to be reckless. They're doing it because the current evaluation process is broken. A single formal observation can easily take hours from start to finish. You sit in the classroom, furiously scribble notes, then spend more time mapping those notes to a complex rubric (some of which run dozens of pages), and then draft a formal report. Multiply that by 30 or more teachers, three to six times a year, and it's no surprise administrators are looking for shortcuts.

A February 2025 RAND Corporation survey of principals found that nearly 60% reported using AI tools during the 2023–2024 school year, with writing-related tasks among the most common use cases.¹ And while the RAND data doesn't break out evaluation writing specifically, the pattern is clear: principals are turning to consumer AI tools for the exact kind of time-intensive writing that evaluations require.

The problem isn't that they want help. The problem is how they're getting it.

The FERPA question nobody is asking

When an administrator copies classroom observation notes into a consumer AI tool, those notes often contain student-identifiable information: names, behaviors, learning needs, specific incidents. Even though the evaluation itself is a personnel record, the observation notes fed into the AI tool can contain student PII, and that's where the FERPA risk lives. Under FERPA (the Family Educational Rights and Privacy Act), personally identifiable information extends beyond just student names. It includes any information that, alone or in combination, could allow a reasonable person in the school community to identify a student.²

FERPA generally requires written consent before a school discloses education records to third parties. Under the statute, “education records” means records directly related to a student and maintained by an educational agency or institution, or by a party acting for the agency.² There is an exception that allows disclosure to “school officials” with “legitimate educational interests,” but this exception requires specific conditions: the party must be performing an institutional service, the school must have direct control over how the data is used and maintained, and a written agreement must be in place.³^,4

Consumer AI tools used on personal accounts often won't meet these conditions absent a district contract, data processing agreement, and administrative controls over how the data is handled. The practical test is whether the district has a contract and controls (managed accounts, retention policies, training-use defaults, audit logs) that make the AI provider function like a school official service provider rather than a public writing app. There's typically no written agreement between the district and the AI provider. The district has no control over how the data is processed, stored, or potentially used for model improvement. The data may be retained well beyond what any educational purpose requires. Depending on the specific provider and tier, consumer data may even be used for training future AI models, meaning fragments of student information could theoretically surface in outputs to other users.

FERPA does exempt “sole possession notes,” defined as personal memory aids kept by a single staff member and not shared with anyone else.² But the moment an administrator pastes those notes into an external AI tool, they've shared them with a third party, and the sole possession exemption likely no longer applies.

This article is not legal advice, and districts should consult qualified counsel about their specific circumstances. But the risk profile here, under both FERPA and applicable state student privacy laws, is significant enough that every district should be asking hard questions about what their administrators are doing.

Consumer AI tools and your data

The distinction between consumer and enterprise AI products matters a lot, and it's one that most administrators don't know about.

When you use a free or standard paid account on most major AI platforms, the provider's terms of service may permit retention and use of your conversations for service improvement and, in some cases, model training. The specific policies vary by provider, tier, and settings. Some providers offer opt-out toggles, but these aren't always the default, and many users never change them.⁵^,6^,7

Enterprise and education-specific versions of these same tools are different. They typically exclude data from model training, offer administrative controls over data retention, and can support FERPA compliance through proper data processing agreements. OpenAI, for example, launched a free “ChatGPT for Teachers” product in November 2025 that includes a Student Data Privacy Agreement and does not use data for training. The free period runs through June 2027.⁸

But how many principals are using the verified education version versus just opening up the consumer app on their phone? In our experience working with Oregon school districts, the answer is almost always the consumer version. They signed up for it themselves. There's no district agreement in place. There's no data processing addendum. There's no audit trail. And there's no way for the district to know what data has been shared.

Unions are paying attention

Both major teacher unions have taken clear positions on AI in evaluations.

The NEA's 2024 policy statement on AI in education states that “AI-informed analyses and data alone should never be used for high-stakes or determinative decisions” affecting educators, including “employee evaluations.” The policy further states that such decisions must “rely primarily on the professional expertise and judgment of humans.”⁹^,10

The AFT's 2024 convention resolution declared that the impact of AI in the workplace is a mandatory subject of bargaining in unionized settings and that critical decision-making must remain with human professionals.¹¹

These aren't abstract policy papers. They're positions that unions will use at the bargaining table and in grievance proceedings. In Ithaca, New York, the local teachers' association identified protecting teacher work from AI outsourcing as a key bargaining priority in 2024–2025 negotiations.¹² In Orange County, Florida, the teachers' association filed a November 2024 lawsuit over the district's refusal to negotiate evaluation procedures, a dispute with direct implications for any AI integration into that process.¹³

The legal picture around AI in employment decisions is shifting quickly. New York City's Local Law 144, in effect since July 2023, requires that employers using automated employment decision tools obtain an independent bias audit within the past year, publish a summary of the results, and provide required notices to candidates and employees.¹⁴^,15 Colorado's AI Act, signed in 2024, will impose impact assessment requirements for high-risk AI used in employment decisions, though the implementation timeline has been subject to legislative revision.¹⁶^,17 State-level obligations vary widely, but the direction is clear: more regulation is coming, not less.

For districts, the risk isn't just that a union files a grievance. It's that the evaluation itself could be challenged as procedurally deficient. Most collective bargaining agreements specify evaluation procedures in detail. If an administrator used AI to draft an evaluation without disclosure and without union agreement, that evaluation could be vulnerable to challenge on procedural grounds, particularly for tenured teachers whose continued employment is a protected interest.

AI-generated evaluations carry real bias

The accuracy and fairness concerns with AI-generated evaluative language are well documented. A study published in The Lancet and reported by Scientific American found that ChatGPT generated recommendation letters with significant gender differences, using more communal and appearance-related language for women and more agentic, professional language for men.¹⁸ A 2025 Stanford Graduate School of Business study found pervasive age-gender bias in large language model outputs, with women systematically described in ways that reflected age-based stereotypes.¹⁹

In K–12, this matters. Roughly three out of four K–12 teachers in the U.S. are women, according to federal labor data.²⁰ If language models show systematic bias in professional evaluative writing, districts should assume similar risks when those same models are used to draft teacher evaluations. Even when an administrator doesn't intend it, biased AI-generated language could create patterns of disparate treatment across a district's evaluation data.

Beyond bias, there's the hallucination problem. AI tools sometimes generate confident, specific-sounding claims that are simply wrong. OpenAI's own research has documented that hallucination is a structural feature of how current language models work.²¹ When an AI tool is asked to elaborate on brief observation notes, it may produce plausible but inaccurate details about what happened in a classroom. If those fabricated details end up in a formal evaluation, the evaluator has signed their name to something that didn't actually happen.

And then there's the problem of generic language. AI tools produce text by predicting likely word sequences, which means they lean toward safe, templated phrases. “Demonstrates effective classroom management.” “Uses questioning techniques to engage students.” These phrases could describe any teacher. They're the opposite of what makes evaluations meaningful: specific, evidence-based observations tied to what actually happened in the room.

The governance gap

While AI adoption among principals is widespread, policy coverage is extremely thin. An EdWeek Research Center survey from summer 2025 found that a significant portion of educators report their district has no AI policy, and even among districts with some guidance, very few address how AI can be used for evaluating or managing staff.²²

Georgia's Department of Education stands out for issuing January 2025 guidance that explicitly categorizes AI use for educator evaluations as prohibited, using a “traffic light” system that red-lights AI for evaluations, IEP goals, and subjective grading.²³ A growing number of states, now numbering around two dozen, have published K–12 AI guidance, though most focus on student-facing use rather than administrative HR functions.²⁴

Here in Oregon, the state issued early AI guidance in 2023 through ODE, emphasizing equity, bias awareness, and data privacy.²⁵ But as ODE spokesperson Liz Merah has noted, there are no existing legal requirements around AI in schools, meaning ODE provides guidance and support rather than mandates.²⁶ Meanwhile, COSA (the Coalition of Oregon School Administrators) actively offers COSA BOT.AI, a subscription tool that allows administrators to generate draft evaluation feedback from observation notes.²⁷ The tension between the teacher unions' explicit opposition to AI in evaluations and the administrator association's active promotion of it captures a fault line that districts will have to deal with head-on.

What districts should do this month

Before getting into what we've built, here are four steps any district can take right now to reduce their exposure:

Put a one-page rule in writing. No student-identifiable information in consumer AI tools. Period. If your district doesn't have an AI policy yet, this is the minimum starting point.

Require staff to use only district-approved AI workspaces for any content that may contain student PII. If your district hasn't approved a specific tool, that means no AI tools for this purpose.

Stand up a vendor review process. Before approving any AI tool for evaluation-adjacent work, review the data processing agreement, retention terms, training-use defaults, admin audit logs, and breach notification procedures.

Decide, with labor counsel, whether AI-assisted drafting in evaluations must be disclosed and/or bargained. Don't wait for a grievance to find out.

There's a better way

This is exactly the problem we built Elevate to solve.

Elevate by Tandem Education is a purpose-built observation platform designed to support FERPA-aligned workflows with district-level controls, contracts, and data processing agreements in place. Instead of administrators pasting sensitive classroom data into consumer AI tools they signed up for on their own, Elevate provides a controlled, district-provisioned environment where AI serves as a co-pilot for the evaluation process.

During a classroom observation, Elevate transcribes audio in real time and applies automated redaction to likely student identifiers in the transcript, with the administrator reviewing and confirming the scrubbed output. By default, audio is processed and not retained, though districts can configure retention settings based on their own policies and union agreements. The administrator adds timestamped notes alongside the live transcript, capturing the visual and contextual observations that audio alone can't provide. After the observation, Elevate's AI suggests how the evidence connects to the district's evaluation rubric and generates a draft report grounded in verbatim teacher quotes from the actual observation.

The administrator reviews, edits, and approves everything before it becomes part of the formal evaluation. The AI never makes the final call. It handles the time-consuming parts of the process so the evaluator can focus on being present in the classroom and providing meaningful feedback that helps teachers grow.

On the privacy side, Elevate was built with data protection baked in from day one. Student data does not route through consumer AI accounts. District data is not used for model training under our customer agreements. The platform supports whatever evaluation framework your district uses, whether that's Danielson 2022, a district-customized rubric, or something else entirely. And because it's a district-provisioned tool, there's an institutional agreement in place, an audit trail, and administrative oversight of how the tool is being used.

For district leaders, Elevate isn't just an efficiency tool. In early pilot use, administrators have reported completing observations that previously took hours in under 30 minutes. But more than speed, it's a risk management tool. It gives your administrators the AI assistance they're clearly looking for, while keeping your district on the right side of FERPA, union agreements, and evolving state regulations.

Your principals are already using AI. The question is whether you want them doing it on their own, on consumer platforms, with no guardrails, or whether you want to give them a tool that was built for this specific job.

We'd love to show you how it works. Get in touch to learn more or schedule a demo.

This article is for informational purposes and does not constitute legal advice. Districts should consult qualified legal counsel regarding FERPA compliance and the use of AI tools in personnel evaluation processes.

References

RAND Corporation, “Uneven Adoption of Artificial Intelligence Tools among U.S. Teachers and Principals in the 2023–2024 School Year,” Research Report RR-A134-25, February 2025. eric.ed.gov. See also EdWeek summary.
34 CFR 99.3 (FERPA definitions, including “education records,” “personally identifiable information,” and sole possession notes exclusion). ecfr.gov.
34 CFR 99.31(a)(1) (conditions for disclosure to school officials without consent). law.cornell.edu.
U.S. Department of Education, “FERPA Exceptions Summary.” studentprivacy.ed.gov.
OpenAI Terms of Use and Privacy Policy (consumer tier). openai.com.
Anthropic Consumer Terms and Privacy Policy. anthropic.com.
Google Gemini Apps Privacy Notice. support.google.com.
OpenAI, “A free version of ChatGPT built for teachers,” November 2025. openai.com.
National Education Association, “Artificial Intelligence in Education: Full Statement Text,” 2024. nea.org.
National Education Association, “Five Principles for the Use of Artificial Intelligence in Education.” nea.org.
American Federation of Teachers, “Resolution on Artificial Intelligence,” 2024 Convention. aft.org.
14850.com, “Contract talks resume between Ithaca teachers and district after months-long pause,” 2024–2025 reporting. 14850.com.
Orlando Sentinel, “Orange County teachers union sues school district over teacher evaluation negotiations,” November 18, 2024. orlandosentinel.com.
NYC Local Law 144: Automated Employment Decision Tools, effective July 2023. davidrichlaw.com.
New York State Comptroller, “Enforcement of Local Law 144: Automated Employment Decision Tools,” December 2025. osc.ny.gov.
Colorado SB 24-205 (AI Act). See K&L Gates analysis: klgates.com.
Seyfarth Shaw, “Colorado Postpones Implementation of AI Law as California Finalizes New Employment Discrimination Regulations.” seyfarth.com.
Scientific American, “ChatGPT Replicates Gender Bias in Recommendation Letters.” scientificamerican.com.
Stanford Report, “Researchers uncover AI bias against older working women,” 2025. news.stanford.edu.
Bureau of Labor Statistics, “Labor Force Statistics from the Current Population Survey,” Table 11. bls.gov.
OpenAI, “Why language models hallucinate,” September 2025. openai.com.
Education Week, “How School Districts Are Crafting AI Policy on the Fly,” October 2025. edweek.org.
Georgia Department of Education, AI Guidance for K–12 Schools, January 2025. Referenced via AI for Education state tracker.
AI for Education, “State AI Guidance for Education” (tracker). aiforeducation.io.
Oregon Department of Education, “Developing Policy and Protocols for the use of Generative AI in K–12 Classrooms,” 2023. oregon.gov.
Oregon Capital Chronicle, “Oregon's governor made a deal with Nvidia to get AI education in Oregon schools. What does it mean?” July 23, 2025. oregoncapitalchronicle.com.
COSA BOT.AI product page, Coalition of Oregon School Administrators. cosa.k12.or.us.