Claude for Teachers: how to audit an education assistant before classroom use

On July 14, 2026, Anthropic introduced Claude for Teachers, one year of free access for verified US K-12 educators who sign up by June 30, 2027. It includes Pro-level features, Claude Code, Cowork, lesson-preparation skills and a connector to academic standards. It is intended for educators and education staff, not students. A dedicated school and district offering was still forthcoming when the announcement was published.

The change is not simply placing a chatbot in front of a teacher. Anthropic is trying to wrap the model in three inspectable layers: curriculum context, procedures for creating materials and terms designed for school data. Those layers help, but they do not make a response pedagogically correct or independently authorize disclosure of records. The useful skill is learning to audit an education assistant in three separate columns: what it does, what evidence evaluates it and what the contract actually permits.

Curriculum as context, not a certificate

The Learning Commons connector provides access to standards in all 50 states, the smaller competencies beneath them and their learning progression. The offering includes references such as OpenSciEd and Illustrative Mathematics and connections to outside tools. In practice, a teacher can request a plan tied to a standard or adapt material for different proficiency levels. That reduces the risk of starting from a completely generic request, but “aligned to a standard” describes a documentary relationship; it does not prove that an activity teaches that content well to a particular group.

Anthropic published the K-12 skills repository. It contains two main procedures: one builds standards-aligned lesson plans and the other differentiates an existing lesson into below-, at- and above-proficiency versions while addressing specific student needs. The second sets an important condition: the core content should remain consistent across levels. Scaffolding should open access to the intellectual challenge, not replace it with an easier task that no longer represents the same learning.

Visible code makes it possible to inspect instructions, examples and dependencies. It also shows that the skills depend on the Learning Commons knowledge graph; anyone running them outside the product must adapt the procedure without that connector. Repository transparency supports review of intended behavior and hidden assumptions. It does not establish what proportion of real outputs will be correct or how the system will perform with every local curriculum.

An open rubric is not yet evidence of impact

The repository includes evaluation rubrics developed by Anthropic and Learning Commons. They separate pedagogy, rigor, output and conversational-scaffolding criteria. Lesson planning has shared criteria plus subject-specific files for mathematics, English language arts, science and social studies; differentiation has one rubric for the material and another for clarification behavior. Each criterion is scored independently so that an average does not conceal a consequential failure.

That structure teaches a transferable practice. A useful evaluation does not ask only, “Is this lesson good?” It asks whether the lesson preserves grade-level demand, requires reasoning, anticipates points of difficulty, separates teacher- and student-facing material and requests missing context. The document itself recommends monitoring per-criterion pass rates because an aggregate can hide gaps. An attractive activity can fail on rigor; a rigorous one can still be impossible to use the next day.

The method matters too. The rubrics are designed to be supplied to a language model acting as judge, although they can also support human or deterministic evaluation. The document calls them a living framework and recommends calibrating criteria and judging with education practitioners or researchers. Publishing a rubric improves auditability, but does not eliminate the question of who scores the generator. Examples, criterion-level failures and human disagreements should be retained instead of only an overall grade.

What the terms promise about data

The account-specific documentation says Anthropic does not train models on the text, files or connectors an educator provides, or on the responses returned. The customer retains rights to inputs and, to the extent allowed by law, owns outputs. The company describes the customer as controller and Anthropic as a processor handling data to provide the service.

The US K-12 terms, effective July 14, add boundaries that do not fit in the product announcement. A teacher may not accept the terms for a school without legal authority or sufficient authorization; if a school already has a written agreement, that agreement governs. Where FERPA applies, the customer represents that it may designate Anthropic as a school official and will not process student data without that authority. The obligation does not disappear because a product page mentions compliance.

The same terms require the customer to assess whether outputs are appropriate before using or sharing them and to warn users that factual assertions may be false, incomplete, misleading or out of date. This is a decisive operational limitation. A draft quiz may save time; a grade, an accommodation for a disability or a sensitive communication requires review, context and human accountability. The product does not transfer that decision to Anthropic.

What the DPA covers and what the school still decides

The K-12 data processing agreement defines student data broadly: it may include education records, discipline, test results, grades, evaluations and identifiers. Anthropic commits not to sell or share those data as defined by applicable law, not to use them outside the direct relationship for purposes unrelated to the service and to maintain confidentiality. Subprocessors must accept substantially equivalent protections, and the customer receives notice before a new one accesses personal data.

The DPA also establishes an incident process: Anthropic will notify the customer in writing without undue delay and no later than 48 hours after becoming aware of a security breach. When the agreement ends, it provides for return on request and deletion within 30 days, with exceptions for legal duties, disputes or combating harmful use. These are concrete commitments that can be checked; “safe for schools” would be a much less informative label.

None of those clauses decides what a teacher should upload. The official FERPA regulation protects records linked to students and regulates disclosure of personally identifiable information. State law and district policy can add further requirements. A prudent protocol starts with minimization: use aggregate or pseudonymized data when sufficient, exclude unnecessary names and identifiers, confirm the authorized purpose and know how to delete the material. The right question is not, “Does the company say it complies?” It is, “Do we have authority, need and controls for this data and this task?”

Testing value without turning a demonstration into proof

Anthropic announced a pilot with Detroit Public Schools Community District to study effects on educator wellbeing and practice. That wording correctly signals that the evaluation was still to come. The launch provides no causal estimate of time saved and does not demonstrate improved learning. It mentions early teacher feedback and says the skills were evaluated for rigor, pedagogical alignment and classroom usability, but the announcement publishes no success rate that can be generalized across subjects and classrooms.

A school can evaluate the product through a ladder of risk. Start with reversible tasks containing no personal data: outlines, practice questions or variants of an explanation. Next, review a sample against the school’s own rubric and reference materials. Only after measuring errors, actual correction time and differences among subjects would wider use make sense. Decisions with consequences for a student require more review and governance than preparing a draft.

The outcome that matters is not how many pages Claude generates, but how much correct and usable work remains after standards, sources, cognitive demand and classroom fit are checked. Claude for Teachers offers unusual material for that audit: public procedures and rubrics, K-12 terms and a specific DPA. Their existence does not finish the evaluation; it makes evaluation possible. Separating function, evidence and contractual permission allows a school to use a tool without mistaking educational packaging for an educational guarantee.