July 5, 2026
Read this, maybe? AI said so
New AI tutor achieves 0.71-1.30 SD effect size in Dartmouth course [pdf]
College kids ignored the textbook—then 90% showed up for the AI version and the comments went wild
TLDR: A Dartmouth study found an optional AI study tool was tied to much better exam results and, shockingly, got about 90% of students to use it when almost nobody read the textbook before. Commenters are torn between calling that adoption number the real breakthrough and warning that motivated students may have made the results look better than they are.
A Dartmouth experiment just dropped the kind of stat that makes education nerds spit out their coffee: an optional AI-powered study tool was linked to much better final exam scores, and more than 90% of students actually used it. That last part is what sent commenters into full main-character mode. The paper says normal textbook reading was basically a ghost town—around 10% to 15%—with one student summed up as, basically, “literally no one does that.” So when an ungraded AI tutor got near-total participation, the community reaction was less “nice research” and more “wait, this is the real bombshell.”
The comment section split into familiar internet factions. Team Optimist called it wildly promising, especially because it got students to engage at all. One commenter practically shrugged at the exact score boost and said the adoption numbers matter more: if the old system teaches a tiny group everything and the new one gets almost everyone learning something, that’s a huge win. Team Skeptic immediately hit the brakes, pointing out the obvious drama: this wasn’t a randomized test, so maybe the most motivated students were always going to do better anyway. Translation: did the AI help, or did the try-hards just try hard harder?
Then came the side quests. One person praised the paper’s looks—yes, the typesetting got fan mail. Another dropped the gloomy hot take that education tech never makes money, even if it works. And a big practical question hovered over the whole thread: sure, statistics is easy to grade, but what happens when the AI tutor wanders into literature or civics and has to judge fuzzier answers? In other words, the comments weren’t just impressed—they were already preparing the sequel, the backlash, and the meme captions.
Key Points
- •The paper introduces Phosphor, a digital learning platform that embeds LLM-graded formative assessments into course readings.
- •In a deployment with 151 Dartmouth College students across three introductory statistics sections, full usage was associated with final exam gains of 0.71 to 1.30 standard deviations.
- •Phosphor was offered as an optional, ungraded alternative to traditional readings and achieved 90.2% adoption.
- •The article argues that unrestricted external AI use can hinder learning, citing prior research on GPT-4 and broader increases in student generative AI use.
- •The study suggests embedded constructed-response quiz formats may be an important factor in improving outcomes within AI-augmented instructional tools.