AI Training Data and Student Work
A plain-language guide to ai training data and student work: the record, the purpose, the authority and the safeguard.
- Record
- student input to an AI tool in ai tools and student privacy
- Data
- AI Training Data and Student Work
- Safeguard
- a pre-adoption review and a contract term on training use
- First check
- Limit the sharing to the minimum the purpose requires.
The short answer
For ai training data and student work, think in terms of a chain: a student, a student input to an AI tool, an objective, an authority, a receiver, a safeguard and a retention date. Break any link and the sharing fails, even if each other link is sound. This guide walks that chain link by link, so the reader can see exactly where a proposed sharing retains and where it breaks.
The record comes first and the requirement follows it, and any determination made in the opposite order should be rechecked before it is acted on.
Accessibility tools can collect personal data about disability, and their confidentiality review must be as careful as their accessibility review.
A verbal sharing counts the same as a written one for privacy purposes, even though it leaves a thinner trail for later review.
A sharing is lawful solely when its purpose, its authority and its receiver all line up, and a gap in any one of those three is a reason to pause.
What this covers
First-generation students might have less familiarity with privacy notifications and opt-outs, and institutions should not treat silence as informed agreement.
Releases to other schools where a student seeks to enrol are permitted, but the student should be told that such releases occur, via the annual notice.
Video surveillance holds images of identifiable students, and its storage, access and disclosure rules should be documented before cameras are installed.
What counts as directory information varies by institution, since each institution designates its own list within the federal framework.
Network logs hold connection information that can reveal student activity patterns, and access to those logs should be logged in turn.
Backups are copies too, and a deletion procedure that ignores backups leaves the record in place under another name.
How to work through it
Run penetration tests on student systems, and treat the findings as a work list with owners and dates, not as a report to file.
Limit the disclosure to the minimum the objective needs, and treat extra information as extra risk rather than extra helpfulness.
Put vendor terms in the contract, not in a sales deck, and check that the contract names the data, limits its use and requires removal on exit.
Test the actual data flow of an integration, rather than trusting a diagram, since integrations routinely pass more than intended.
Audit surveillance systems for access, storage and disclosure, because cameras generate files continuously, even when nobody watches.
Separate duties so that the person who approves a sharing is not the solely person who can see that it happened.
Prepare for a regulator inquiry by retaining inventories and logs current, because inquiries rarely arrive with a convenient timetable.
Audit vendor compliance, and do not accept a vendor's self-assessment as a substitute for evidence where the data is personal.
A worked illustration
The cost of fixing the badly done adoption, in renegotiation, migration and lost trust, is far higher than the cost of the review that was skipped.
Consider a tutor who wants a second opinion from a colleague regarding support for a student, and who first checks whether a summary without identifiers would serve.
In that illustration the project is paused, the information uses are documented in plain language, and the dashboard is rebuilt around a stated objective.
In that illustration the registrar explains the position politely, in writing, and offers the parent a route via the student's own consent.
Consider a college that receives a media request for a student's disciplinary file and routes it to the privacy officer instead of answering from the first department.
Where guides go wrong
Its value is in the structure it teaches: name the record, state the objective, settle the authorization, limit the sharing, protect the data, and retain the record.
Redisclosing a file received under an exception, without checking whether the second disclosing is covered, is a mistake that spreads the original disclosure.
Excluding students from privacy oversight is a mistake, since students notice problems that staff have normalised.
Disclosing on the strength of a letter alone, without a court order or subpoena process, is a mistake that may be avoided by a single phone call to the privacy officer.
Leaving access in place following a role change is an error that compounds quietly, because the former access looks legitimate in every log.
Sources and evidence on the record
- Family Educational Rights and Privacy Act regulations, 34 CFR Part 99, text on eCFR.
- US Department of Education, Student Privacy Policy Office, studentprivacy.ed.gov - FERPA guidance, FAQs and model notices.
- US Department of Education, ed.gov - guidance, Dear Colleague letters and enforcement information.
Evidence basis: 34 CFR Part 99, Department of Education guidance, FTC AI guidance Applies when: When a college holds, uses or discloses a student input to an AI tool in the context of ai tools and student privacy