A Dartmouth Study Found 90.2% Engagement on 'Optional' Coursework — Here's the AI-Graded Assessment Service You Can Sell From It.

A workshop paper out of Dartmouth just handed away a productizable service, and it's not "build an AI tutor."

What the study actually found

Jonah Bard's paper, "Balancing Efficacy and Engagement in Interactive Texts," reports on Phosphor, a platform deployed across three sections of Introductory Statistics (MATH 010) at Dartmouth in Spring 2026 — 151 students enrolled, 143 finishing.

Phosphor embeds quizzes directly into lesson content. Multiple-choice questions are auto-graded; constructed-response questions are "graded by Claude Sonnet 4.6 against instructor-defined, question-specific rubric criteria." The platform is "presented as an entirely optional, ungraded alternative to traditional readings," and "quizzes permit unlimited retries."

The results, quoted directly from the abstract: "Full dosage of the Phosphor material is associated with an increase in final exam performance of between 0.71 SD (adjusting for prior exam scores) and 1.30 SD (unadjusted)." And on adoption: "the platform was adopted by 90.2% of enrolled students, far exceeding typical reading-compliance rates" — against a baseline the paper reports as "student- and instructor-reported baselines of 10-15% for this course."

The detail most people will skip past

The paper ran a natural experiment across its three course modules. Module 1 used mixed multiple-choice and constructed-response quizzes, and each additional lesson completion tracked with roughly 1.6 more percentage points on the midterm (p < 0.001, R² = 0.123). In Module 2, after "widespread student feedback that the CRQ auto-grader was rigid and discouraging," the team switched to multiple-choice-only quizzes — and the dosage-to-score relationship among users with at least one completion dropped to essentially flat (R² = 0.001). Module 3 restored constructed-response questions.

In other words: the format that felt easier to students was also the format that stopped predicting learning outcomes. The engagement driver (optional, ungraded, retry-friendly) and the outcome driver (effortful, LLM-graded, typed responses) are two different levers, and this paper is one of the few places you'll find both isolated with real numbers.

Worth noting before you pitch this: the HN discussion of the paper raises real caveats — no traditional control group, a possible selection effect where already-motivated students both engage more and score higher, and the fact the study format itself changed mid-course in response to feedback. Read it before you oversell this to a client.

The service to build

1. Find a client whose audience already doesn't consume what they publish — this is nearly every course creator, corporate training team, SaaS onboarding flow, and certification program. Reading compliance is a near-universal problem; you don't need this exact study to know that, but now you have citable numbers for the pitch.

2. Build quiz infrastructure directly into their existing content: short, ungraded, optional, unlimited-retry quizzes per section or lesson, mixing multiple-choice with constructed-response.

3. Wire constructed-response grading through an LLM against rubric criteria written per question — this is the part that actually needs engineering, and the part that plugin-style quiz tools skip.

4. Ship a lightweight completion dashboard reporting both an exposure figure (upper bound) and a completion figure (lower bound), the way the paper's Table 1 does, so the client can report real numbers to their own stakeholders.

5. Retainer the relationship on rubric tuning, quiz-bank expansion as content grows, and periodic spot-checks of grading quality.

Bottom line

The paper's real finding isn't "AI tutors work." It's that removing friction from the decision to participate (optional, ungraded, unlimited retries) and preserving effort in the task itself (typed, rubric-graded answers) are both required, and most existing quiz tools only do one of the two. That gap is the service.

Source: https://intextbooks.science.uu.nl/workshop2026/files/itb26_s1s2.pdf (presented at iTextbooks'26, Seoul, June 28, 2026)