Illinois Gov. JB Pritzker on July 11, 2026, signed Senate Bill 2909, making the state the first in the nation to categorically prohibit evaluators from using artificial intelligence tools to assign numerical scores or qualitative ratings in teacher performance evaluations. The law takes effect Jan. 1, 2027.
What the law prohibits and permits
The enrolled bill text states that an evaluator may not use an AI tool to assign a numerical score or a qualitative rating such as "excellent," "proficient," "need improvement," or "unsatisfactory" for any component of a teacher's evaluation or any evaluation task requiring professional judgment. AI may still be used to support the evaluator in administrative tasks. Separately, the law bars teachers from using AI to generate evidence of professional practice that will be used by an evaluator during the evaluation process, though teachers may use AI for administrative support.
The legislation also mandates mutual disclosure: if an evaluator uses an AI tool, the name and specific purpose of the tool must be disclosed to the teacher being evaluated. If a teacher uses an AI tool, the same information must be disclosed to the evaluator.
The bill does not define the term "artificial intelligence tool," leaving interpretation to the joint committee that oversees each district's evaluation plan under the Performance Evaluation Reform Act (PERA). That committee, composed of equal representation selected by the district and its teachers or their exclusive bargaining representative, will determine how AI tools may be used in accordance with the new restrictions. The law creates no new enforcement mechanism, penalty, or private right of action, according to the legislation tracker Regulon.
Senate sponsor Christopher Belt said that teachers should be judged on actual observations and professional judgment, not by AI software, and that educators deserve a transparent and fair evaluation process that reflects their actual classroom work and protects their privacy. House sponsor Mary Beth Canty stated that she supports exploring AI as a tool for basic organization and streamlining simple tasks, but that the technology is not capable of handling judgment-based tasks this complex.
The bill passed both chambers unanimously, and was enrolled on May 27, 2026, before being sent to the governor.
How other states are approaching AI in schools
Illinois's categorical ban on AI in teacher evaluations is narrower in scope but more absolute than legislation enacted or proposed in several other states during the 2025-2026 session. Most other state measures focus on student-facing AI, not on educator evaluation.
Oklahoma enacted SB 1734, the Oklahoma Responsible Technology in Schools Act, signed in May 2026. That law requires AI tools in schools to be educator-directed with a human-in-the-loop, prohibits AI from being the primary basis for high-stakes decisions such as grading or discipline, mandates annual parent disclosure, and requires district AI policies by the 2027-2028 school year. Unlike Illinois's law, Oklahoma's applies broadly to student-facing AI rather than specifically to educator evaluation.
Texas introduced HB 5282 in the 2025 regular session, which would have prohibited AI from scoring constructed-response assessment items unless the tool was trained on representative samples including disadvantaged students, demonstrated validity and reliability consistent with NAEP standards, and underwent independent bias evaluation. That bill was referred to committee but did not pass, and it targeted student assessments, not teacher evaluations.
Virginia enacted legislation in its 2026 session directing the Department of Education to develop guidance for safe, ethical, and equitable use of AI in instructional settings, requiring local school boards to adopt policies, and establishing an AI Innovation in Education Pilot Program with annual reporting on effectiveness, safety, equity, and risks. The guidance includes prohibitions on relying solely on AI for certain high-stakes decisions as defined by the department.
Maryland enacted SB 720 in the 2026 regular session, which requires the State Department of Education to issue AI guidance, mandates local school systems to adopt AI policies, requires designation of AI coordinators, and directs the development of a rubric for evaluating AI tools selected for use by local systems. Morgan State University or another four-year institution is tasked with supporting certification of compliant AI tools.
South Carolina introduced Bill 5253 in its 2025-2026 session that would prohibit AI from replacing licensed teachers in core instruction or assigning final grades, ban AI from making automated disciplinary or placement decisions without meaningful human review, require parental opt-in consent for student AI use, and mandate public disclosure of approved AI tools. As of February 2026 the bill remained in committee.
What research says about AI in teacher evaluation
A growing body of peer-reviewed research published in 2025 and 2026 has examined the reliability, accuracy, and limitations of using large language models (LLMs) to evaluate classroom teaching. The findings are mixed, with several studies pointing to substantial limitations in current AI tools.
A 2026 study in Computers and Education: Artificial Intelligence compared 8,618 AI-generated evaluations from eight different LLM endpoints against consensus ratings from certified TEACH experts. The study found substantial stochastic variability across repeated evaluations, with no model achieving uniformly high reliability across instructional elements. Agreement with expert ratings remained moderate at best, and reliability and accuracy did not co-vary systematically. The authors concluded that automated classroom observation cannot be treated as a uniform capability and highlighted structural limitations of current text-based AI observation pipelines.
A 2025 preprint study on arXiv reported that custom LLMs built on sentence-level embeddings could achieve human-level and even super-human performance on classroom observation instruments, surpassing average human-human rater correlation above 0.65. However, the alignment with teacher value-added measures held at the aggregate level but not consistently at individual item level, suggesting the models have not yet achieved full generalization. The study also noted that even expert human ratings have low reliability, unknown accuracy, and are expensive to conduct.
A 2026 meta-analysis in Frontiers in Psychology synthesizing 35 pre-test-post-test designs and 37 post-test studies found that AI-driven teaching interventions yield significant positive effects on teaching effectiveness broadly, with a post-test effect size of g_p = 0.586 and a pre-post gain of g_delta = 0.136. But the study noted high heterogeneity driven primarily by AI type and results direction, and it focused on AI's impact on teaching effectiveness generally, not specifically on AI's use in evaluating teachers.
A 2026 paper from the Association for Computational Linguistics (ACL Findings) cautioned that automated scoring systems should not replace human raters in high-stakes settings, where unchecked deployment may create self-reinforcing feedback loops of biased predictions. The authors recommended positioning AI as decision-support tools that augment rather than replace professional judgment.
A 2026 working paper from the Annenberg Institute at Brown University added a further layer of context. Using five years of Chicago Public Schools data, researchers found that classroom composition causally affects teacher performance ratings: a one standard deviation increase in a classroom quality index led to a 0.07 standard deviation increase in observation ratings and a 0.13 standard deviation increase in student survey scores, even for the same teacher. A policy simulation showed that adjusting ratings for classroom characteristics would meaningfully change rankings, with Black teachers benefiting most, gaining about 8 percentile points. The finding that human observation systems already contain evaluator bias provides context for concerns that AI might amplify similar biases.