Tutoring interfaces

A tutoring interface is the surface through which a tutor presents material and collects the learner’s answers. For an AI tutor the main contrast is between a free-form chat conversation and a structured interface: a lesson page or tutoring-system screen with steps, answer forms, quizzes and records of earlier attempts. The practical question is whether the interface changes learning measured afterwards without the AI, or whether only the teaching strategy behind it matters. The question is open, because the nearest evidence compares tools that differ in prompts and design as well as interface (Chen et al., 2026b; Brender et al., 2025; Bastani et al., 2025). The evidence is stronger for individual features such as prompted self-explanation, and weaker or contested for confidence ratings, closed-book quizzes and learner-visible progress records. How AI tutors teach in general is the subject of AI tutoring.

History

The research did not cover the history of tutoring interfaces beyond the classification of computer tutors by interface granularity (see Core concepts).

Core concepts

Terms for tutoring interfaces

The terms are not standardized, and “structured interface” has no fixed definition in the sources. VanLehn (2011) classifies computer tutors by the granularity of the interaction at the interface: other computer tutoring (computer-aided instruction, computer-based training) is answer-based, while most intelligent tutoring systems (ITS) are step-based or substep-based (VanLehn, 2011). A “structured interface” or “step-based tutor” in this article corresponds to the latter. A 2026 paper defines conversational tutors as systems that hold natural-language dialogue, in contrast to ITS that act as adaptive problem sets (Vanacore et al., 2026). An open learner model is a learner model that is shown to the learner to support reflection (Kerly et al., 2007).

What interface granularity is worth

VanLehn’s review reported intelligent tutoring systems at d = 0.76 against no tutoring, close to human tutoring at d = 0.79 (VanLehn, 2011). Human tutoring’s effect in that review was much lower than the d = 2.0 once believed (VanLehn, 2011).

LLM help can be delivered inside such a structured tutor rather than in a chat. In a randomized study (N = 274), LLM-generated hints given inside a problem-solving tutor produced significant learning gains against no help, with no significant difference from human-authored hints; the human-authored condition alone did not reach significance against no help (Pardos & Bhandari, 2024). That study shows hints work inside a step interface; it does not compare the same hints with chat delivery.

Questions, controversies, and new developments

Lesson page versus chat

Three studies come closest, and none isolates the interface. In a graduate robotics course (n = 58), a structured ChatGPT interface produced no difference in performance or learning from free ChatGPT use; the prompting behaviours linked to higher learning gains did not carry over once students were no longer constrained to the structured platform. The source does not state whether post-tests were taken without AI (Brender et al., 2025).

In a staggered-access study of a chatbot plus an embedded proof-review tool, earlier access raised homework performance but not exam scores. Heavier chatbot use and answer-seeking were negatively associated with midterm performance, and chatbot use, not proof-review use, was negatively associated with a later exam (Chen et al., 2026b). The authors state that these are associations under naturalistic conditions, because students chose how much to use each tool (Chen et al., 2026b).

In a field experiment with high school mathematics students, those who had used a plain ChatGPT-like tutor scored worse, by 17 percent, on a closed-book, closed-laptop exam than students who never had access. A tutor with learning-safeguard prompts largely removed the harm, leaving exam performance statistically indistinguishable from the control, so about zero effect rather than a gain (Bastani et al., 2025). Both tutors were chats; the contrast tests prompt guardrails, not page versus chat (Bastani et al., 2025).

Do learners follow a structured interface?

One 2026 preprint compared course chatbots and found answer-seeking was more common on an unrestricted chat platform than on pedagogically constrained tools, though still the majority on both (66.5 percent on constrained tools, 92.0 percent on unrestricted chat). Roughly half of students ignored the AI tutor’s guidance, and rates of resistance to scaffolding were nearly identical across platforms (0.227 against 0.223) (Kobler et al., 2026). The authors interpret students as treating scaffolding as a barrier to answers, and report that use of optional AI tools concentrates around deadlines (Kobler et al., 2026). The study is observational and the platforms came from different courses.

Splitting a lesson across two places

A design with a chat and a separate page splits the lesson in two. General multimedia-learning evidence says separated, related information costs learners: a meta-analysis (Ginns, 2006) supported the split-attention effect with a weighted mean effect size of d = 0.85 (Beege et al., 2019). Findings are mixed, and the harm depends on intrinsic load: when experienced intrinsic load is low, the extra load does not inhibit learning (Beege et al., 2019).

Self-explanation prompts

A meta-analysis of prompted self-explanation found a moderate overall benefit (g = .55) across 69 effect sizes (Bisra et al., 2018). The effect showed no statistically detectable difference by medium (digital, print, video), interactivity (computer-based instruction, ITS, simulation, none) or presence of diagrams, and was found in digital and print alike (both g = .55); the authors’ absence of a detected difference is not proof of equivalence (Bisra et al., 2018). A page or tutoring system is therefore not required for the effect.

Whether the prompt’s format matters is unsettled. The meta-analysis could not decide: multiple-choice prompts (only 2 studies) showed an effect not different from zero, and the authors call for direct comparisons of multiple-choice, fill-in and open-ended prompts (Bisra et al., 2018). In a 92-person calculus study, menu-based and open-ended (LLM-feedback) self-explanation did not differ from no self-explanation on post-test performance. The open-ended condition wrote better explanations than control on “Not Enough Information” transfer problems, but multiple-choice accuracy was not significantly better, and learners in that condition did fewer practice problems (Chen et al., 2026c). A secondary report of game studies says open-ended focused self-explanation gave higher learning than menu-based in one study, while Johnson and Mayer (2010) found no difference (Andres et al., 2023).

Confidence ratings before feedback

In one lab study, a confidence-weighted multiple-choice format, which builds uncertainty into the answer, beat standard multiple-choice practice on a final cued-recall test of new but related questions (8.11 against 6.54 of 20 items in Experiment 2). Adding a separate 0-100 confidence rating to standard multiple choice did not help (6.41 against 6.54, t(72) = .45, p = .65) (Sparck et al., 2016). When learners rate confidence and then get feedback, high-confidence errors are more likely to be corrected (the hypercorrection effect), but correction fades after a week, and high-confidence errors can return if the correct answer is forgotten (Butler et al., 2011).

A single AI-powered metacognitive calibration tool improved learning gains over a control group by 8.9 percent in a randomized trial (N = 133), with the effect significantly mediated by learning behaviours. The tool bundles several features, so the gain cannot be attributed to confidence rating alone (Lee et al., 2025b).

Closed-book quizzes

Results conflict. In a field study of two parallel university courses, one closed-book and one open-book (not randomised by student), students who answered practice questions closed-book outperformed open-book students on a surprise quiz in week 8 and on related questions of the later exam. Samples were small (about 46 at the quiz, 51 at the exam), with delays of 1 to 7 weeks between practice tests and the surprise quiz and 8 weeks to the exam (Rummer et al., 2019).

In a doctoral thesis’s experiments, open-book and closed-book testing produced equivalent retention on a final test after two days. The format of the initial test did not change long-term retention; what mattered was whether learners expected a closed-book final test, and a closed-book expectancy led to better delayed performance than an open-book expectancy (Agarwal, 2009). Closed-book practice is thus not clearly better as a format, though learners who expect a closed-book final test retain more on a delayed test. Separately, a reanalysis argues that the classic advantage of retrieval practice over concept mapping is a methodological artifact (extra memorization time), so the evidence for retrieval over other study activities is itself contested (Mayrhofer et al., 2023).

Attempt histories and progress records

Evidence that learner-visible progress records improve learning is weak. A narrative synthesis of 26 open-learner-model studies found generally positive but mixed results: gains in monitoring and self-assessment accuracy, some achievement gains and several nulls, depending on design; effects were not pooled (Robles Mucho et al., 2026). A systematic review of 38 studies found no evidence that learning analytics dashboards improve academic achievement, and many studies confounded dashboard use with engagement (Kaliisa et al., 2023).

Broader connections

Most of the effect-bearing features here are techniques rather than interface properties: prompted self-explanation works in print as well as on screen (Bisra et al., 2018), so a chat can in principle host it. The Self-study materials article covers the same techniques in static material. The evidence on guardrails suggests that what the tutor is told to do can matter as much as where the learner meets it (Bastani et al., 2025).

Further reading

• Bisra et al. (2018), the meta-analysis of prompted self-explanation: the strongest evidence here, with format left open (Bisra et al., 2018).
• Bastani et al. (2025), the field experiment on unguarded versus safeguarded chat tutors, measured on a closed-book exam (Bastani et al., 2025).
• Chen et al. (2026), a comparison of conversational and embedded LLM feedback in a proof course; associational (Chen et al., 2026b).
• VanLehn (2011), the review of human tutoring, ITS and other tutoring by interface granularity (VanLehn, 2011).
• Sparck et al. (2016), confidence-weighted testing against a separate confidence rating (Sparck et al., 2016).