How Alps uses language models

How language models help

Alps uses language models to help with tutoring, practice, feedback and grading, to help teachers create activities, and to show teachers how their class is doing.

What they never do

• Pathfinder, the help students get while they practice, is not a chat box: a student never types a question to a language model.
• Language models are never used to detect cheating, and Alps never decides whether a student cheated.

Teachers stay in charge

• Teachers see every grade and can change any of them. A teacher’s grade always takes precedence.
• Teachers can turn Pathfinder off for any assignment.
• A student who has a question about a grade asks their teacher, who can review and change it.

Student data

• Requests to language models carry no student names, email addresses or account identifiers. Student work is sent as the student wrote it, so it can include anything the student typed.
• We don’t use identifiable student work to train models, and we don’t permit model providers to train on it.
• We never sell student data or use it for advertising.
• Our list of service providers is available on request.

Our privacy policy has the full details, and districts can sign a student data privacy agreement with us.

How we manage the risks

A written policy guides how we use language models. It follows the four functions of the NIST AI Risk Management Framework:

• Govern. Our Security and Privacy Officer owns the policy and approves each model provider, and each new use of a language model, before it reaches a class.
• Map. Before release, we assess each use: its purpose, the data it needs, who sees its output, what could go wrong, and how teachers oversee it.
• Measure. We check the quality of what language models produce, including how their grades compare with teachers’ grades.
• Manage. We can switch off any use of a language model when we need to. We review reports of wrong or harmful output, and handle any that expose data or cause harm under our incident response plan.

2Sigma is a member of the EdSafe AI Alliance Industry Council.

Contact

Email privacy@alps.2sigma.us with questions about how Alps uses language models, to report output that looks wrong or harmful, or to request our AI Risk Management Policy and completed security questionnaires.