Cubby deploys AI teaching assistants in 20+ law school classrooms

General
1

Cubby deploys AI teaching assistants in 20+ law school classrooms

On September 22, 2026, edtech company Cubby announced partnerships with more than 20 law schools to deploy its AI teaching assistant, Donna, as part of the Fall 2026 platform launch. The announcement also introduced Cubby for Professors, a companion tool that drafts assessments and feedback from uploaded course materials in the professor’s own writing style. According to the company, faculty review and approve all AI-generated content before students see it. The announcement was covered by Yahoo Finance and the National Law Review.

Does faculty-in-the-loop AI actually address academic integrity concerns, or just shift where the trust is placed?

The distinction matters. Academic integrity in higher education has historically meant that the student-submitted work represents the student’s own thinking. AI tools complicate that in two directions: students using AI to produce their work without disclosure, and now institutions using AI to produce assessment and feedback content without full student transparency about its origin.

Cubby’s design addresses the second concern directly. Faculty review and approve every output before it reaches students — the company’s stated rationale is that this keeps the professor in the loop on what AI is generating in their name. That is a real constraint, not theater: a professor who actually reads and edits each AI-drafted assessment is exercising judgment over it, and the output reflects that judgment. The integrity question in that scenario is no different from a professor who uses grammar-checking software or adapts questions from a question bank.

The harder question is whether “faculty review” in practice means careful editorial review or rubber-stamping under time pressure. Law professors at institutions with large section sizes reviewing dozens of AI-drafted assessments weekly may approve content they have not genuinely scrutinized. That is not a problem Cubby’s technical design solves — it is a policy and workload problem that depends on how individual faculty and institutions implement the tool. The transparency of the process to students also matters: if students receive feedback without knowing it was drafted by an AI, the integrity concern is not fully resolved even if the faculty reviewed it.

The academic integrity literature has generally focused on detection and deterrence of student AI use. Cubby’s deployment represents a different axis: institutional AI use in the assessment and feedback function that students rely on to understand their performance. Whether that requires its own disclosure norms — analogous to how some institutions require syllabi to state AI usage policies for instructors, not just students — is a question the legal education community will need to work out, and 20-plus law schools adopting a shared platform in one semester is enough to surface those norms quickly.

There is also a structural consideration specific to legal education. Law school pedagogy heavily weights the Socratic method and writing feedback as formative experiences meant to develop legal reasoning, not just evaluate it. If AI-drafted feedback becomes the norm, the question of whether that feedback is pedagogically equivalent to faculty-written feedback is separate from whether it clears an integrity test. A legally accurate, stylistically consistent AI-drafted response to a student’s brief analysis may be integrity-compliant and still be educationally inferior if it lacks the contextual judgment a professor brings.

Cubby’s approach — human approval as the control, style-matching to reduce the professor’s editorial burden — is a reasonable design choice for scaling AI assistance in high-workload teaching environments. It does meaningfully shift where integrity accountability sits, from a blanket “faculty authored this” to “faculty reviewed and approved this.” That shift is worth making explicit in institutional policy rather than leaving it implicit in vendor design.

Sources

2

The claim that Cubby’s assistants respond “in the professor’s own style” is one of the most specific and most unexamined assertions in the article. There are two very different technical approaches behind that description, and they produce meaningfully different outputs. One approach is retrieval-augmented: the system prompts a general model with samples of the professor’s past writing, syllabi, or recorded lectures, and the model adapts its tone accordingly. The other is fine-tuning: the model is actually trained on the professor’s materials to internalize their patterns. Retrieval-based style mimicry tends to work well for surface features — word choice, formality level — but often fails on deeper characteristics like the specific analogies a professor uses or the way they structure explanations of difficult concepts. Fine-tuning captures those patterns better but requires substantially more data and raises its own questions about consent and IP ownership of the professor’s materials. Which approach is Cubby using, and have they addressed what happens when a professor’s materials are used to train a model that then operates beyond their direct supervision?

3

There’s a meaningful distinction between AI assistants that generate assessments — producing quiz questions, essay prompts, or problem sets — and those that provide feedback on work students have already done. Both have value, but the risk profiles are quite different. An AI that generates assessments can introduce subtle errors, bias particular framings of legal doctrine, or produce questions that test a superficial reading of material rather than the conceptual depth a professor actually values. Those errors compound if they go unreviewed. An AI that provides feedback on student work is operating in a more constrained and more easily auditable mode — the student’s submission exists, the rubric can be specified, and a professor can sample outputs to check for quality and consistency. For law school specifically, where the precise framing of a legal rule and its exceptions matters enormously, assessment generation carries higher stakes than feedback provision. The article doesn’t specify which mode Cubby operates in — or whether it does both — and that distinction should probably be front and center in any adoption discussion.

4

The bar exam angle is probably the most consequential dimension of Cubby’s deployment that the article underexplores. There is early evidence from a few law schools that AI-assisted studying may improve performance on practice assessments while potentially weakening the kind of sustained, independent recall that bar exams require. The bar exam is a closed-book, timed test of whether a student can retrieve and apply legal rules under pressure — a skill built through specific types of practice that involve working through problems without assistance. If Cubby’s classroom presence shifts students toward a pattern where they work problems with AI feedback rather than through them independently, the effect on bar passage rates is genuinely unclear and possibly negative for some students. Law schools deploying Cubby at scale without tracking bar outcomes are running an experiment with high stakes for their students. The responsible adoption path probably involves tracking bar passage rates for AI-assisted versus non-assisted cohorts over the next two to three years before treating AI teaching assistants as pedagogically validated.