Our AI policy.
How Coursim uses large language models – and the limits we put on them.
Which models
Coursim uses commercial large language models for most generative tasks. Specific models and versions rotate for quality and cost; the current deployed configuration is listed on our status page.
What we do not do
We do not train foundation models on student data. We do not fine-tune models on identifiable student work. Student prompts and responses are not used to improve any model, by us or by the model provider. This is contractually binding, and the provider is named in our DPA.
Content safety
Every AI-generated output passes our content safety filter before reaching a student. The filter screens for age-inappropriate language, sensitive topics inappropriate for the configured grade band, and known prompt-injection patterns. False-positive rate is published quarterly.
Sage, the tutor
Sage is constrained to pedagogical behavior: never gives final answers, uses scaffolded questioning, escalates to the teacher after sustained struggle. Sage cannot browse the web and cannot retrieve information from any student's account other than the one she is speaking with.
Teacher oversight
Every AI-generated lesson is reviewable and editable by the teacher before publishing to students. Nothing Coursim generates reaches a student without a teacher's review or an explicit teacher-set auto-publish rule.
Research
We publish quarterly reports on content safety performance, including false-positive rates, near-miss incidents, and process improvements. Reports are at our blog.
Questions about how we use models?
We publish our model list, our limits, and our evaluation methodology. Ask us anything a skeptical teacher or board member would.