AI teaching assistants in online discussion forums and peer-supported self-regulated learning

An AI teaching assistant that answers fast does not just answer: it changes who else speaks. Across four graduate courses and 28,123 coded forum sentences, an AI TA was the first responder in 97% of student-started assignment threads and replied in a median of 0.92 minutes against 490 minutes for peers, and students in AI-supported courses wrote fewer reply posts while starting threads at the same rate. Peer repliers carried roughly 70% of the self-regulatory language in their sentences against 27% for thread starters, so as their share of follow-up posts fell from 91% to 45% across a semester, aggregate regulatory language fell with it. The same responsiveness that makes an AI TA attractive appears to displace the peer dialogue through which co-regulation happens.

Key Findings

• The AI TA became the dominant responder where it was allowed to operate, and the shift was by design plus speed rather than by preference: the AI was first responder in 97% of student-started assignment threads in both high-exposure courses (OR = 62.7, 95% CI [27.3, 169.0], p < .001), with a median first response of 0.92 minutes against 490 minutes when a peer replied first and 364 minutes when a teacher did.
• Students in AI-supported courses posted less overall (15.85 posts on average against 23.05, p < .001), and the gap came entirely from replies: thread starts were statistically indistinguishable (4.88 against 4.93, p = .73), while reply posts fell from a mean of 18.2 to 10.9 (p < .001).
• Lower regulatory language followed from the lost peer replies rather than from any change in how individuals wrote. SRL-coded sentences dropped from 49% to 36% between the first and second half of the semester in the AI courses (t(67) = -6.03, p < .001) while rising slightly in the human-only course, but within each role the SRL rate stayed nearly constant: peer repliers 70.2% to 68.5%, thread starters 26.9% to 20%. Peer repliers' share of follow-up sentences, not their rate, collapsed from 91% to 45%.
• Having more AI TA responses in the threads a student started was associated with lower grades (beta around -1.38, p = .038), while a higher proportion of teacher responses went with higher grades (beta around 2.12, p = .009) and peer response proportion showed no relationship (beta around 0.02, p = .99). Where a student posted did not matter: the share of their activity in assignment threads did not predict grade (p = .282), even though that is precisely where the AI TA was active.
• Self-regulation predicted grade (beta = 2.63, 95% CI [0.49, 4.76], p = .017), and neither received co-regulation (p = .482) nor provided co-regulation (p = .660) predicted it on its own. Peer interaction moderated the relationship (interaction beta = 1.45, 95% CI [0.31, 2.60], p = .013), raising explained variance from R-squared 0.10 to 0.17.
• The grade penalty was concentrated among students best equipped to regulate. High early self-regulators who received above-median co-regulation outscored their below-median peers by 0.74 SD (p = .004); low self-regulators showed no comparable gap (0.36 SD, p = .369). Removing 17 high-self-regulation students whose threads were AI-dominated moved the AI TA coefficient from beta around -1.38 (p = .038) to -1.29 (p = .454).

The Responder Ecology of an AI-Populated Forum

The study treats a discussion forum as an interactional system rather than a pipe for answers, and names the actors within it a responder ecology: peers, instructors and human teaching assistants, and now an AI TA. The framework matters because prior work on teaching presence already showed that who responds shapes how a discussion proceeds, and instructor replies tend to close threads rather than extend them by supplying a definitive answer. An AI that answers in under a minute, consistently and for every thread, tightens that closing effect to its extreme.

JeepyTA, developed at the Penn Center for Learning Analytics and used across five universities and 29 courses, generated responses from category-specific prompts plus retrieved context, with optional instructor approval before publication. The courses analyzed used GPT-3.5 Turbo, GPT-4 and GPT-4 Turbo depending on when they ran. Four graduate courses formed the comparison: one human-only data science course with 40 students and 6,491 sentences, the same course with limited AI exposure where only 20% of students interacted with the AI TA (56 students), and two high-exposure AI courses where 89% (45 students, 4,756 sentences) and 95% (39 students, 8,687 sentences) of students were reached.

Why Regulatory Language Fell

The paper's most consequential analytic move is to separate self-regulation from co-regulation by position in the thread rather than by sentence content. Regulatory language in a thread a student started counts as that student's own regulation. The same language in a reply to somebody else's thread is co-regulation, one learner Scaffolding another's regulatory activity. Because the AI TA's documented effect was to displace peer replies, and peer replies are exactly where co-regulation lives, this coding choice turns a participation metric into a mechanism.

The result is that the drop in regulatory language is compositional. Nobody became a worse regulator in the sense of writing fewer regulatory sentences per post; the population of writers changed. With peer repliers supplying about 70% SRL-coded sentences and thread starters about 20 to 27%, the fall in peers' share of follow-up sentences, from 91% to 45%, drags the aggregate down mechanically. The finding also locates where the loss is sharpest: in AI-mediated threads the opening post carried less SRL than comparable openings in peer-mediated or teacher-mediated threads, and after the opening, SRL rose in AI threads while falling in teacher threads.

Who the Lost Peer Dialogue Hurts

The moderation analysis implies that peer dialogue and self-regulation are complements rather than substitutes, which cuts against a common assumption that strong self-regulators are the self-sufficient ones. If self-regulation pays off more when peers engage, then removing peers should cost the strongest regulators most, and that is what the median-split comparison shows. The authors are careful that the mechanism is not established: why high self-regulators were the sensitive group is not visible in the data, and unmeasured prior knowledge or Motivation could account for it. They also note the observational design cannot separate the AI TA's presence from the courses that chose to deploy it.

What this means for practice

• Treat an AI TA's response latency as a design variable. The authors point to the wait-time literature, in which delaying an authoritative answer raises the quantity and complexity of student contributions, and propose manipulating AI TA latency in asynchronous forums as the natural next experiment.
• Interrogate the responder mix rather than the answer quality. The grade association ran through who populated a student's threads, and the proportion of a student's own activity in assignment threads predicted nothing, so a course can look healthy on AI response quality while the peer dialogue that carries co-regulation drains away.
• Watch the peer reply series over a semester, not just its level. Peer shares fell from 91% to 45% between halves in these courses, a drift invisible in a single end-of-term snapshot and easy to miss in a dashboard that reports totals.
• Do not assume that the students who regulate best need the least support. In this dataset they were the ones whose grades moved most with the presence of peer co-regulation, so designing for self-sufficient high performers may be designing for the group most exposed to losing peers.

Limitations

• The design is observational, so the associations reported may reflect course-level or student-level confounds rather than effects of the AI TA, and the authors state plainly that causality cannot be established.
• Forum discourse is a proxy for regulation that captures only what students articulate in writing, leaving internal regulatory activity unobserved.
• The sample is graduate-level data science and educational design, where students may hold relatively advanced self-regulatory skills, which limits generalization to other levels and domains.
• Statistical power is modest at the student level: some regressions rest on roughly 62 to 67 students, which tempers the confidence intervals around the moderation and the concentration result.

Citation

Davis, R. L., Liu, X., Pankiewicz, M., Kissmehl, A., Engwall, O., Baker, R. S., & Viberg, O. (2026). AI teaching assistants in online discussion forums and peer-supported self-regulated learning. Proceedings of the European Conference on Technology Enhanced Learning (EC-TEL 2026), paper 354.