AI governance

How MyTA uses AI responsibly in higher education

1. Our approach to AI

MyTA uses AI to assist instructors, not to replace them. The AI drafts replies, feedback, and grades for you to review. Replies, announcements, and reminders wait for your approval, and grades stay drafts until you publish them in Halo. Section 4 describes the two things MyTA does in Halo without a click.

2. AI models we use

MyTA uses commercial AI models, accessed through each provider’s API:

Anthropic Claude: The default for MyTA’s AI features. Larger Claude models handle the chat assistant, essay grading, and drafting. A smaller, faster model handles summaries, discussion grading, and planning.
Google Gemini or OpenAI: An institution can choose either one instead of Anthropic. The chat assistant still writes its replies with Anthropic. The Telegram assistant runs on Google Gemini by default.
Voyage AI: Turns the text of inbox threads and the chat assistant’s notes into search vectors so MyTA can find related items.

Some tasks use no AI model: sorting student messages by intent and flagging sensitive messages both use keyword rules. Ascent Labs does not build, train, or fine-tune its own AI models.

3. Student data and AI models

Ascent Labs does not use student data to train AI models. How long an AI provider keeps request data, and whether it uses that data, is set by that provider’s terms.
Grading, AI detection, and drafting replies and discussion posts replace student identifiers with placeholders like “[Student A]” in student text before it is sent to the AI. Each institution’s administrators choose which identifiers are replaced.
Grading a criterion about formatting can send the original file as written. The privacy policy has the details.
The replacement is not AI-powered: it uses roster-based matching and regex patterns. It can miss an identifier written in a form it does not recognize, and the request still goes ahead when that happens.
Other features send student names or student text to an AI provider as written. These include discussion reminders, student outreach messages, the chat assistant, and the Telegram assistant. Reminders and outreach need the student’s name to address them, and the chat and Telegram assistants need names to answer questions about a student. Search indexing through Voyage AI also receives student message text as written. Building a teaching voice profile and learning an instructor’s grading preferences apply the replacement while the instructor’s “Automatic PII redaction on examples” setting is on, which is the default.

4. Human oversight

You review drafts in the web app, where you can approve, edit, or reject each one. You can also approve reply drafts from Telegram.

These always wait for your approval:

Replies to student messages. Approving a reply sends it to the student.
Announcements and reminders.
Marking students complete on a discussion.
Discussion posts and replies that MyTA drafts for you. Saving one puts it in Halo privately, and students see it only after you publish it.
Saving grades to Halo, apart from the discussion grading described below.

Two things happen in Halo without a click:

Discussion grading: In a discussion where grading is on, which is the default, MyTA uses AI to grade student posts that don’t have a grade in Halo yet. It saves each grade, with any feedback, to Halo as a draft. To stop this, turn Grading off on the discussion.
Substantive marks: When Auto-Evaluate Substantive Replies is on, MyTA marks replies that meet the reply word minimum as substantive in Halo. This check uses rules, not AI. The setting is off unless an admin or instructor turns it on, and you can turn it off for a course or a single discussion.

Turning off the Discussion Compliance Agent doesn’t stop either one. MyTA doesn’t publish grades, so grades it saves to Halo stay hidden from students until you publish them there.

5. Fairness & bias mitigation

AI grading is anchored to rubrics or assignment instructions provided by the instructor, not to general AI knowledge or training data biases.
Instructor voice profiles ensure feedback matches the instructor’s own communication style, not a generic AI tone.
Course-specific policy settings (word count minimums, grading intensity, feedback level) are configurable to reflect institutional standards.
AI grades stay drafts until you publish them in Halo, so you can check each one first.

6. What AI cannot do in MyTA

AI grades are drafts. A grade is final only when you publish it in Halo.
The system does not send AI-written messages to students on its own. Students who allow notifications get automatic push notifications when a grade is posted or a discussion deadline is near.
The system cannot modify LMS settings or administrative functions
Messages that mention topics such as self-harm, harassment, accommodations, or academic integrity are flagged for the instructor by keyword rules, not AI, and MyTA does not draft a reply to them automatically. Flagged messages are still included in search indexing through Voyage AI, and the Telegram assistant can list them.

7. AI usage monitoring

MyTA logs each request it makes to its AI provider (Anthropic, Gemini, or OpenAI) with the action type, model used, and token counts. This data is available to institutional administrators through the admin dashboard, supporting transparency and cost tracking. Requests to Voyage AI and Mem0, and the Telegram assistant’s requests, are not in this log.

8. Questions & documentation

For questions about MyTA’s AI practices, contact Ascent Labs at privacy@courseops.ai. Institutional IT teams can request more detailed technical documentation.