AI Tools With Student Privacy Protections: Nuubi, Claude, ChatGPT, and Gemini Compared
If you review AI tools for a university, you have probably noticed that every vendor now says some version of "we take privacy seriously." On its own, that sentence tells you very little. The details that matter are specific: whether student writing can train a model, who can read conversations, how long data is kept, and which plan each promise attaches to. This post compares four tools that a higher-education reviewer is likely to encounter (Nuubi, Claude, ChatGPT, and Gemini) against one checklist, using each vendor's own documentation. Sources are listed at the end.
This article is published by Nuubi, one of the four tools compared below.
A Short Checklist Does Most of the Work
A short checklist covers most of what a reviewer needs to know. Does student data train AI models by default on this plan? Who owns the data, and under which agreement? Who inside the institution can see student conversations? How long is data retained, and who controls deletion? Does the vendor support FERPA obligations in writing? (https://www.nuubi.com/blog/student-data-ai-tools-ferpa-questions) For every answer, name the plan. Consumer and institutional plans under the same brand often have different terms.
One question goes beyond privacy: can the instructor who teaches the course see how students are using the tool, and step in?
OpenAI's ChatGPT Edu Keeps Student Data Out of Model Training by Default
OpenAI's institutional offering is ChatGPT Edu. According to OpenAI, data from ChatGPT Edu is not used to train OpenAI models, the institution controls how long data is retained, and deleted conversations are removed from OpenAI's systems within 30 days unless legally required (https://openai.com/enterprise-privacy/). Conversations are encrypted in transit and at rest, and OpenAI processes Edu data under its Student Data Privacy Agreement.
The consumer plans under the same brand work differently. For Free, Plus, and Pro accounts, OpenAI may use content to train its models unless the user opts out in settings (https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance). A student who pastes coursework into a personal account is therefore covered by those consumer terms, and the Edu commitments do not apply.
On FERPA, look at the contract rather than the marketing. For ChatGPT Edu, OpenAI processes data under its Student Data Privacy Agreement, and it publishes an implementation guide for institutions that handle student data (https://cdn.openai.com/osa/openai-sdpa-educational-implementation-guide.pdf). The guide puts real duties on the school, from configuring single sign-on to confirming that its use complies with privacy laws. A reviewer should check the signed agreement and the institution's configuration, rather than treat the product name as proof of FERPA compliance.
Anthropic's Claude for Education Puts the University in Charge of Its Data
Claude for Education is Anthropic's institutional plan. Anthropic's privacy center says the university's organization owns and manages the account and its data, and that Anthropic's consumer terms and privacy policy do not apply to Claude for Education, where Anthropic acts as the data processor (https://privacy.claude.com/en/articles/11732894-who-owns-and-manages-the-data-of-my-claude-for-education-account). Admin capabilities that could expose student conversations, including data exports and audit logs, are off by default, and the university's Primary Owner must request them from Anthropic.
On training, Anthropic's education page states that it only uses data to train its models when given permission (https://claude.com/solutions/education). For consumer plans, Anthropic says chats are used to improve its models when the user opts in to model improvement, submits feedback, or a conversation is flagged for safety review (https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training). Those terms should not be confused with the university's agreement.
On FERPA, Anthropic's public materials for Claude for Education do not name the law. Its education page says Claude meets the compliance requirements higher education institutions have for student data, and Anthropic acts as the data processor under the university's agreement. Anthropic does make an explicit FERPA claim for a different product: Claude for Teachers, its K-12 offering, advertises FERPA-aligned protections and a FERPA-configured setting (https://support.claude.com/en/articles/15926041-claude-for-teachers-your-data-and-our-terms). Higher-education reviewers should request a FERPA-specific position in writing for Claude for Education.
Google's Gemini Is Covered by the Workspace for Education Terms
For institutions on Google Workspace for Education, the Gemini app is a core service under the Workspace for Education terms. Google says the service supports compliance with FERPA, COPPA, and HIPAA (https://blog.google/products-and-platforms/products/education/gemini-enterprise-grade-data-protection/). Its privacy hub states that chats are not reviewed by humans or used for training outside the institution's domain without permission, and that prompts count as customer data under the Cloud Data Processing Addendum (https://knowledge.workspace.google.com/admin/generative-ai/generative-ai-in-google-workspace-privacy-hub). Retention is set by administrators: Gemini conversation history can auto-delete after 3, 18, or 36 months, with 18 months as the default. As with the other vendors, these commitments attach to the education edition and do not extend to a student's personal Google account.
Nuubi Is Built to Integrate Into a Course
Nuubi is built to integrate into a course. Its privacy protections follow that structure:
Course materials are used as context for answers, never to train AI models, by Nuubi or the backend LLM API. Relevant course content and the student's question are sent to the backend LLM API to generate a response. Synced materials stay scoped to that exact course: Nuubi does not pool them across courses, institutions, or tenants, or retrieve them for answers in another course. Using content for a response is different from training a model on it (https://www.nuubi.com; https://www.nuubi.com/blog/ai-privacy-safety-security-students).
Answers cite course materials that faculty control, so the source behind an answer stays visible inside the course.
Access is scoped to the course. Private drafts and feedback stay visible only to the student, enforced on the server and not just hidden in the interface, and usage reporting stays at the course level.
Student questions and activity are sent to the backend LLM API anonymously, with role labels (admin or member) in place of names. This keeps student names out of the model input, which helps prevent bias based on names, and follows the principle of least privilege.
Hosting is in the United States, and Nuubi states FERPA compliance on its public pages (https://www.nuubi.com).
Nuubi retains data for eight years by default because higher-education institutions may need records from earlier academic years. Institutions, students, and faculty can request earlier deletion; the default is not a requirement to keep every record for eight years. Under current terms, the backend LLM API may keep prompts and responses in abuse-monitoring logs for up to 30 days by default, with legal and safety exceptions. No training does not mean no storage or no provider access; authorized staff and contractors may access API data for limited support, abuse-review, and legal purposes (https://platform.openai.com/docs/guides/your-data; https://openai.com/enterprise-privacy/). Institutions can request Nuubi's completed HECVAT 4 and VPAT.
Privacy Controls Are Not the Same as a Human in the Loop
Everything so far concerns data: where it goes, who can read it, how long it is kept. A reviewer can check every one of those boxes and still miss what happens while the tool is in use. On the general assistants, many oversight controls sit with administrators. ChatGPT Edu lets workspace admins pull conversation logs through its compliance platform. Claude for Education keeps data exports and audit logs off by default, and the university's Primary Owner must request them from Anthropic. Gemini gives admins control over conversation-history retention. It also gives teachers access to transcripts for assigned Classroom notebooks and Guided Learning activities when Gemini activity is on (https://support.google.com/edu/classroom/answer/16534159?hl=en; https://support.google.com/edu/classroom/answer/18866217). That teacher access is an important exception to the admin-level controls.
Nuubi puts course-level oversight at the center. Instructors see the questions and exchanges in their own courses as they happen, can step in at any point, and review the AI-fluency coaching students receive. In this course-level human-in-the-loop model, the AI answers, the instructor can see the exchange and step in, and the loop closes inside the course, not through an IT request. The general assistants are organized around a workspace or an institution. Google's Classroom tools also provide teacher visibility for assigned activities, but not for a student's Gemini use generally.
The Plan Matters More Than the Logo
All three general assistants offer strong protections on their institutional plans, including no training by default, written agreements, and admin controls. All three also have consumer plans under the same brand where the defaults differ. When a student uses a personal account for coursework, the institution's contract does not follow them there. For a reviewer, the practical question therefore covers the vendor, the specific plan students will actually use, and whether the institution's agreement covers that plan.
These Questions Close Most Gaps Before You Choose
A short list of questions closes most gaps. Which exact plan do these commitments attach to, and where is that in writing? Does student content train models on that plan by default? Who at the institution can read student conversations, and is that access on or off by default? Who sets retention, and what happens when a student or a course is deleted? Finally, what does the instructor of a course see compared with an IT administrator?
Nuubi Fills a Gap at the Course Level
The institutional plans provide substantial data protections. Course-level oversight is a separate consideration: workspace administration is not the same as an instructor seeing AI-supported work within a course. Google's assigned Classroom activities provide some of that visibility. Nuubi integrates it into course questions, exchanges, and AI-fluency coaching.
The Comparison at a Glance
Cells for the three general assistants come from the vendor documentation in the sources, current as of November 2026. Nuubi's cells come from its public pages and its Technical Security Summary, available to institutions on request.
| Question | Nuubi | ChatGPT Edu | Claude for Education | Gemini for Education |
|---|---|---|---|---|
| Does student data train models on the institutional plan? | Never | No | Only with permission | Not without permission |
| What does the model receive? | Relevant course materials, questions and conversation context; attachments when used; role labels (admin or member) in place of student names | What the user types or uploads | What the user types or uploads | What the user types or uploads |
| How are course materials used and scoped? | Context only, never training; restricted to the exact course, not pooled across courses, institutions or tenants. | Context under the institutional plan; check workspace and sharing settings. | Context under the institutional plan; check the university agreement and account controls. | Context under Workspace terms; check sharing permissions and assigned Classroom activities. |
| Who at the institution can read student conversations? | The instructor of the course, by design | Workspace admins, through an audit log | No one by default; access must be requested | Teachers can read chats in assigned Classroom activities when Gemini activity is on |
| How is access scoped? | Private drafts and feedback visible only to the student; reporting is course-level only | Each user sees their own conversations | Each user sees their own conversations | Each user sees their own conversations |
| How long is data kept, and how is it deleted? | Nuubi: 8-year default, earlier deletion on request. Backend LLM API: prompts and responses may be kept in abuse-monitoring logs for up to 30 days by default under current terms; legal and safety exceptions apply | Admin-controlled; deleted within 30 days | Until deleted; back-end deletion within 30 days, with safety, feedback and legal exceptions | 18-month default; admin retention settings include auto-delete after 3, 18, or 36 months |
| FERPA position in writing | States FERPA compliance on public pages | Student Data Privacy Agreement and implementation guide | Processor agreement; FERPA named only for its K-12 plan | States support for FERPA compliance |
| Consumer plan under the same brand with different defaults? | None | Yes | Yes | Yes |
| Instructor can see student AI use in their course | Yes, for course exchanges and coaching | No instructor role | No instructor role | For assigned Classroom notebooks and Guided Learning, when Gemini activity is on |
Sources
OpenAI, Data controls in the OpenAI platform: https://platform.openai.com/docs/guides/your-data
Anthropic Privacy Center, Education account ownership and data management: https://privacy.claude.com/en/articles/11732894-who-owns-and-manages-the-data-of-my-claude-for-education-account
Anthropic Privacy Center, Consumer data and model training (March 16, 2026): https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training
Claude for higher education: https://claude.com/solutions/education
OpenAI, Enterprise privacy at OpenAI: https://openai.com/enterprise-privacy/
OpenAI Help Center, ChatGPT Edu at OpenAI: https://help.openai.com/en/articles/9377311-chatgpt-edu-at-openai
OpenAI Help Center, How your data is used to improve model performance: https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance
Google, Gemini gets enterprise-grade data protection for education (March 28, 2025): https://blog.google/products-and-platforms/products/education/gemini-enterprise-grade-data-protection/
Google Workspace Help, Generative AI in Google Workspace Privacy Hub (updated October 6, 2026): https://knowledge.workspace.google.com/admin/generative-ai/generative-ai-in-google-workspace-privacy-hub
Nuubi: https://www.nuubi.com
Nuubi, Beyond the Model: AI Privacy, Safety, and Security for Students: https://www.nuubi.com/blog/ai-privacy-safety-security-students
OpenAI, ChatGPT Edu: https://chatgpt.com/business/education/
OpenAI, Student DPA Educational Implementation Guide: https://cdn.openai.com/osa/openai-sdpa-educational-implementation-guide.pdf
Anthropic, Claude for Teachers data terms: https://support.claude.com/en/articles/15926041-claude-for-teachers-your-data-and-our-terms
Nuubi, FERPA Questions for Reviewers of AI Tools That Handle Student Data: https://www.nuubi.com/blog/student-data-ai-tools-ferpa-questions
Anthropic Privacy Center, Organizational data retention: https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data
Google Classroom Help, Create and assign Gemini Notebook and Gems in Google Classroom: https://support.google.com/edu/classroom/answer/16534159?hl=en
Google Classroom Help, Create and manage Guided Learning activities in Classroom: https://support.google.com/edu/classroom/answer/18866217