In the fall of 2023, a high school in Turkey turned nearly one thousand students into participants in an experiment. Over four 90-minute math practice sessions, researchers deployed two AI tutors—both powered by GPT-4. One was a standard chat interface: ask anything, get anything. The other was custom-built: its system prompt preloaded the full solution to each problem and teachers’ summaries of common errors, and it was explicitly instructed never to give answers directly. The practice results were stunning: students using the standard chat version scored 48% higher than the control group; those using the custom version scored 127% higher. Then came the critical step: researchers removed the AI and administered a no-AI exam to all students. Those who had used the standard chat version scored 17% *lower* on the exam than the control group—students who had never used AI at all. Every point gained during practice vanished—and then some. In contrast, the custom-version group performed essentially on par with the control group: no gain, but also no loss. This randomized controlled trial, led by a University of Pennsylvania team and published in PNAS, states its conclusion bluntly in the paper’s title: “Unguarded Generative AI Harms Learning.”
Why should this matter to everyone who cares about learning—parents, students, self-learners? Because it punctures a widely held intuition: the more powerful the tool, the better the learning outcome. These two tutors ran on the same model—yet their outcomes pointed in opposite directions. The difference lies not in intelligence, but in design.
## A Forty-Year-Old Dream
“The dream of every student having a great tutor” is not new—it’s a decades-old challenge in education research. In 1984, Benjamin Bloom, an educational psychologist at the University of Chicago, published a now-classic paper introducing the “2-sigma problem.” Synthesizing two doctoral students’ controlled experiments, he found that students receiving one-on-one tutoring *plus* mastery learning (with formative assessments and immediate correction at every step) scored, on average, two standard deviations higher than students in conventional classrooms (about 30 students per teacher)—placing them above 98% of the control group; roughly 90% of tutored students reached performance levels achieved only by the top 20% of the control group. Using mastery learning alone—without one-on-one tutoring—cut the effect in half: one standard deviation, placing students above 84% of peers.
Bloom’s assessment was clear-eyed: one-on-one tutoring was “too costly for most societies to implement at scale.” So he reframed it as a challenge: Could we find methods to bring group instruction close to one-on-one effectiveness?
The closest answer before large language models emerged came in 2016. Kulik of the University of Michigan and Fletcher of the Institute for Defense Analyses conducted a meta-analysis of 50 controlled evaluations of Intelligent Tutoring Systems (ITS): median effect size was 0.66 standard deviations—lifting students from the 50th to the 75th percentile—exceeding the ~0.4 standard deviation typical of human tutoring programs reported in the literature (note: Bloom’s 2σ measured the *combined* effect of optimal one-on-one + mastery learning; the meta-analysis measured diverse ordinary tutoring programs—these figures are not contradictory). Yet 0.66 remains far from 2σ—and older ITS relied on hand-coded knowledge models, failing completely outside pre-defined problems.
Large language models made this old dream suddenly seem feasible: dialogue is dynamic, content isn’t limited to preloaded question banks, and marginal cost approaches zero. The catch? Realizing the dream is far more finicky than imagined.
## First, the Good News: When Designed Right, Effects Are Stunning
In fall 2023, on Harvard University’s largest introductory physics course, the Kestin and Miller team ran a crossover randomized controlled trial: 194 students each attended one session with an AI tutor and one with face-to-face active learning—order randomized. The control group wasn’t a low-bar class: 89% of students rated this course as using *more* active learning than other STEM courses they’d taken. Results: median learning gain for the AI-tutor group was more than double that of the face-to-face group—statistically highly significant. Effect size: linear regression yielded 0.63 standard deviations—authors note this was suppressed by ceiling effects in the post-test; using quantile regression (more robust to ceilings), estimates ranged from 0.73 to 1.3. The AI group was also faster: median time 49 minutes vs. 60 for face-to-face; time spent showed no correlation with post-test scores. Students felt it too: engagement ratings were 4.1 vs. 3.6 (5-point scale); 83% judged the AI tutor’s explanations as equal to—or better than—those of live instructors.
But this AI tutor wasn’t just ChatGPT handed to students. Every component was tightly constrained by instructional design: step-by-step solutions for each problem were pre-written into the prompt to prevent hallucinated answers; the platform enforced sequential progression—no skipping steps; the system prompt embedded operational rules for active learning, cognitive load management, and growth mindset. The authors state plainly: evidence-based pedagogical best practices “must be explicitly and carefully built into every application.” Crucially, personalized immediate feedback and self-paced progression—two features impossible to sustain reliably in a 30-student classroom—are where AI wins—not because it’s smarter.
Good news comes in cheaper versions too. In summer 2024, a World Bank team ran a six-week trial across nine public high schools in Benin City, Nigeria: first-year students attended computer lab sessions twice weekly, using Microsoft Copilot (GPT-4 powered) to study English aligned with the national curriculum—guided by teachers. Results: composite scores rose 0.31 standard deviations; English-specific scores rose 0.24; even the official semester English exam—covering material far beyond the six-week intervention—rose 0.21. Translating learning gains, six weeks equaled 1.5–2 years of regular schooling; every $100 invested generated 3.2 “equivalent school years”—ranking among the most cost-effective education interventions ever documented. Benefits accrued across ability levels—with largest effects for girls and higher-starting students. For the first time, the cost barrier to “every student has a tutor” has truly fallen.
## Now, the Bad News: Same Model, Opposite Outcomes
Back to Turkey. Why does the same generation of model yield Harvard’s +0.63 to +1.3 standard deviation gains on one side—and a 17% *drop* after tool removal on the other? Comparing experimental designs side-by-side reveals the answer.
The standard chat version failed due to “answering on demand.” Researchers reviewed student chat logs: unguarded, many students treated GPT-4 as an answer machine—asking directly for answers, copying them, moving on. Practice scores soared—because answers were always correct—but the *learning event* of problem-solving never occurred. The custom version’s prompt architecture blocked this path: knowing each problem’s solution, it could pinpoint exactly where a student got stuck—and offer hints, not answers. Behavioral evidence followed: custom-version students sent significantly more messages, and over successive sessions learned to use the tool more substantively—asking follow-ups, requesting explanations—not just demanding answers.
The Harvard team explicitly warned of this in their discussion: educators should avoid deploying AI in contexts where students might use it as a “crutch” to bypass thinking; AI tutors “should not replace face-to-face instruction,” but rather prepare all students equally *before* class—freeing classroom time for higher-order practice.
## Why “Feeling Like You’ve Learned” Is a Terrible Gauge
Both experiments tap into a mechanism named decades ago by educational psychology. UCLA’s Robert Bjork introduced “desirable difficulties” in 1994: spaced practice, interleaved topics, using testing *as* learning—conditions that worsen *immediate* performance and slow things down, yet strengthen long-term retention. His core warning: people naturally assume “conditions that improve classroom performance must improve long-term learning”—and this assumption “can be spectacularly wrong.” Performance and learning trace two separable curves.
Desirable difficulty has a boundary condition often overlooked by those citing it: difficulty is only “desirable” when students *can* handle it; beyond that capacity, it becomes pure frustration. This is precisely the line between a good tutor and an answer machine (this synthesis is the article’s judgment): a good AI tutor removes *undesirable* difficulty—getting stuck with no one to ask, waiting, lacking feedback, being misled by hallucinated answers—while preserving *desirable* difficulty—retrieval, generation, trial-and-error. GPT Tutor’s prompt engineering and Harvard’s PS2 Pal platform essentially translate Bjork’s lab-based, human-delivered interventions into replicable software design. Difficulty lands within the student’s zone of capability—and thus yields learning. If AI performs the “reach-and-grasp” act itself, report cards look instantly better—but the learning event vanishes on the spot.
This also explains why “feeling like you’ve learned” is the most unreliable gauge: fluent dialogue, instant answers, readable explanations—all generate *familiarity*. And Bjork repeatedly stresses: familiarity is most easily mistaken for understanding.
## Four Takeaways
First, when evaluating any AI learning tool, ignore performance *during* practice—focus solely on what remains *after* closing the tool. In the Turkish experiment, the group whose practice scores rose 48% was precisely the one whose exam scores fell 17% once AI was removed. Practice scores can be faked; learning outcomes won’t perform for you.
Second, when choosing an AI tutor—for yourself or your child—choose the one that “asks first,” not the one that “answers instantly.” A tool willing to say, “What do you think the next step should be?” is vastly more valuable than one that complies on demand—its value lies in preserving desirable difficulty.
Third, if you’re an educator: the model is just the engine; pedagogy is the steering wheel. The same GPT-4, fitted with guardrails and scaffolding, becomes the Harvard experiment’s excellent tutor; left unconstrained, it becomes the Turkish experiment’s crutch. And design principles are already spelled out concretely across two trials: pre-load solutions to prevent hallucinations; enforce step-by-step progression to prevent skipping; give hints—not answers; retain the human role (the Nigerian trial included teacher guidance).
Fourth, if you’re an individual using AI for learning or work: for looking up facts, polishing phrasing, or practicing dialogues—go fast. But before outsourcing *thinking itself*, pause and ask: Is this thinking a skill I need to train—or a burden I can set aside? Your answer determines whether the same tool becomes a tutor—or a crutch.
Have you ever experienced the “understood everything in the moment, forgot it all upon closing the app” feeling while using AI to learn? Or conversely, been pushed to real learning by a tool that refused to give direct answers? Please share your experiences in the comments.
The deepest irony here is this: the potential of AI tutors is real—the 2σ problem, unresolved for forty years, now has its first cost-feasible candidate solution. But the key to unlocking that potential lies not in the model—it lies in the hands of designers, and in the hands of users. The more powerful the tool, the heavier the weight carried by those two “hands.”
Source: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895486