AI Transparency

What the AI does

Tutorify uses large language models for three things: answering student questions grounded in the materials of their own course, generating draft assessments from instructor-supplied content, and producing rubric-based draft grades and feedback for instructor review. It also summarizes engagement and mastery patterns into course analytics.

It does not make admissions decisions, does not proctor examinations, does not monitor students outside the coursework they submit, and does not produce a final grade on its own.

The instructor decides

Every output that reaches a student passes through an instructor first. Generated assessments are editable and require approval before publication. Draft grades and feedback are proposals: the instructor reviews, adjusts, and approves before anything is written to the gradebook or synced to the LMS.

This is a design commitment, not a configuration option, and it is the reason an academic-integrity or grade-appeal conversation ends with a human who made the decision. Where regulation distinguishes between an AI system that assists a human and one that decides, Tutorify is firmly in the first category.

Which models, and where they run

Tutorify runs open-weight models on infrastructure we operate ourselves. Course content and student work are not sent to a third-party inference provider, to a consumer AI product, or to anyone’s personal AI account. There is no external API in the inference path.

This matters beyond the privacy claim. Because we host the models, the terms governing your data cannot be changed by a vendor mid-contract, there is no third party retaining your prompts even briefly, and there is no downstream chain of inference sub-processors for your counsel to trace.

Institutions evaluating Tutorify can ask us for the specific model family, version, and licence in use for their deployment, and we will confirm it in writing as part of the agreement. If we ever introduce a managed inference service for some deployments, we will say so here and notify institutional customers before it takes effect.

What the model is given

A request to the model carries the course material relevant to the question, the student’s submission where the task is grading or feedback, and the instructions that define the task. Personal data is minimized before the request is made: identifiers that are not needed to complete the task are removed rather than passed along.

Model inputs and outputs are not used to train any model, ours or anyone else’s. We do not sell data, and we do not use student data for advertising or for building profiles.

Telling students

Students are told when they are interacting with an AI assistant rather than a person, and are told that AI-assisted grading has been reviewed by their instructor. We support institutions that want additional or differently-worded disclosure to meet their own AI policy.

Limitations we will state plainly

• Language models can produce fluent text that is wrong. Grounding answers in course materials reduces this substantially but does not eliminate it.
• Draft grades reflect the rubric they were given. A vague rubric produces vague grading, and the instructor’s review is what catches it.
• Model performance varies across subjects, question types, and languages. It is generally weaker on highly specialized or very recent material that is thinly represented in the course content provided.
• Automated assessment of writing can reflect biases present in training data. This is a live area of evaluation for us and one reason instructor review is mandatory rather than optional.

If you are running a pilot, we would rather help you design an evaluation that surfaces these limits in your own courses than have you discover them later.

Changes

Material changes to how models are used, which providers serve inference, or what data is sent to them are communicated to institutional customers before they take effect, and this page is dated so you can see when it last changed. Questions to privacy@tutorify.ai.