Editorial — We hold that lecture-hall AI assistants must ship with open corpus disclosures—training data summaries, licensed textbook corpora, and live retrieval indexes—because students asked to rely on tutoring bots cannot judge answers whose provenance is hidden behind vendor NDAs.
Campuses are buying black boxes
Dozens of universities raced to deploy classroom copilots after vendors promised safer models with campus-controlled hosting. Contracts often guarantee FERPA compliance yet classify retrieval stores as trade secrets. A bot that cites no source can still shape exam prep for thousands of undergraduates in introductory economics, biology, and writing.
Faculty senates already police plagiarism; they cannot police synthetic paraphrases drawn from unknown forums.
Why open corpus beats vague “safety”
Transparency does not require publishing raw scrapes. It requires machine-readable manifests: which publisher collections are licensed, which open-courseware sets are indexed, which date ranges are excluded, and whether user prompts train future weights. Students could see when an answer leans on a decade-old forum post versus a peer-reviewed text.
NIST’s AI risk framework treats documentation as a control, not a nice-to-have. Campuses should treat it as a procurement requirement, not a press release.
The objection from vendors
Suppliers warn that corpus maps aid adversarial prompt attacks and aid competitors. Security through obscurity failed in every prior ed-tech wave; it will fail here. Red-teamers need manifests to test bias; so do disability services offices verifying whether assistive deployments treat dialect fairly.
If a model cannot survive disclosure, it should not grade participation in a required course.
Limits of this argument
We are not demanding that every research university open-weight frontier models. We demand disclosure at the boundary where credit-bearing instruction meets automated guidance. Research labs can keep separate stacks under stricter review.
What boards should do
Trustees should withhold renewal dollars until vendors provide quarterly corpus delta reports and independent bias audits tied to named document sets. Congress, revisiting Title IV accountability, should tie assistive AI subsidies to public manifests stored with accreditors.
Students deserve to know whether the lecture hall bot learned from their own campus library—or from a scraped stack overflow thread. Until disclosures are standard, “campus-approved AI” is marketing, not pedagogy.
Authority in higher education still runs through curriculum committees, not venture slides. Make them see the corpus, or keep the assistant out of the hall.
Accreditation and credit
Regional accreditors reviewing learning outcomes should ask whether AI tutors appear in syllabi as required tools. If yes, corpus manifests belong in the evidence file next to textbook ISBNs. If no, deans should explain why IT deployed hallway kiosks that students treat as mandatory.
Disability services offices report rising requests to compare bot answers with human note-takers; without corpus maps, those comparisons become he-said-she-said disputes that delay accommodations.
A practical floor
We support pilots that publish retrieval snippets with every answer—footnotes, not marketing. Vendors that refuse should lose RFP points, even if their models score well on benchmark trivia.
Parents paying tuition deserve the same visibility as faculty: if a bot shapes grades, its reading list is part of the syllabus.
Library alliances
University presses and library consortia already negotiate corpus licenses for journal archives; those contracts should become the template for classroom bots, with redacted manifests published alongside e-reserve policies. When vendors refuse, libraries should block SSO integration rather than smuggle black-box tools through department credit cards.
Student government leaders from several public flagships said they would back boycott petitions against undisclosed tutors; trustees should listen before enrollment deposits renew.
