The correlation between usage frequency and self-reported exam performance was ρ = −0.02 (P = 0.81), and with self-rated academic standing ρ = −0.01 (P = 0.93). In other words, students who felt AI was turbocharging their learning were not reporting better grades than classmates who rarely touched the tools. The study, published open access in Discover Education on 24 October 2026, recruited 103 medical students across all four class years at an institution with 422 enrolled students, distributed through class-wide GroupMe messaging groups with reminders in the second and fourth weeks.

After excluding two students who did not consent and six who left required items incomplete, the analytic sample stood at 103, a response rate of 24.4%. All four class years were represented in comparable numbers, with 26 first-year, 23 second-year, 27 third-year, 25 fourth-year, and 2 research-year respondents, split almost evenly between pre-clinical (47.6%) and clinical (52.4%) training phases. The majority of respondents were female (61.2%), and the largest racial or ethnic groups were White or Caucasian (55.3%) and Asian (19.4%).

No incentives were offered, and the survey received exempt status from the Central Michigan University Institutional Review Board under the revised Common Rule. The survey instrument was developed by the investigators and administered electronically via REDCap. AI usage frequency was captured with a single item on a 5-point ordinal scale ranging from Never to Always, asking how often students currently use AI tools to assist with studying or coursework.

Perceived academic benefit was assessed with three Likert-scale items covering distinct facets: whether AI use has positively impacted exam grades, whether it helps students understand complex medical concepts faster than traditional methods, and whether it builds confidence in clinical reasoning during case simulations. These three items showed strong internal consistency, with Cronbach's α = 0.84, and were averaged into a Perceived Benefit Composite score. Academic performance was measured two ways: a single self-reported tier collapsing all exam averages across all blocks and all years of medical school into categories from below 70% to Honors at 90% or above, and a self-rating of overall academic standing compared with peers.

The usage numbers alone are a signal of how thoroughly generative AI has penetrated medical training. More than half of respondents, 53.4%, reported using AI tools Often or Always for studying. When the Perceived Benefit Composite scores were compared across three usage groups, the contrast was dramatic: students in the Often/Always group reported a mean score of 3.9 (SD = 0.6) on the 5-point scale, compared with 2.3 (SD = 0.9) among Never/Rarely users, a difference of 1.6 points that was highly significant (P < 0.001).

But the same table told a very different story about attainment. The proportion of students achieving Honors did not differ significantly across usage groups (P = 0.68), and neither did the proportion rating themselves Above Average (P = 0.95). The study applied a Bonferroni correction across its five primary statistical tests, setting the adjusted significance threshold at P < 0.01, and the usage-benefit association cleared that bar decisively while the usage-performance associations did not approach it.

The authors frame the study through the Technology Acceptance Model, which holds that perceived usefulness is a central determinant of technology adoption. Applied here, the model predicts that students who perceive greater academic benefit will use AI more frequently, and it makes no claim that this perceived usefulness corresponds to objectively measurable gain. The research team, working within a post-positivist paradigm of quantitative measurement, was explicit that its two central variables represent different things: usage frequency is an indicator of adoption, and the Perceived Benefit Composite is an indicator of perceived usefulness rather than verified academic improvement.

The study was also shaped by a practical observation: faculty had noticed that generative AI had become a routine part of how students prepare for coursework, while institutional guidance on academic integrity, transparency, and data stewardship specified acceptable use at the course level but said nothing about how often students actually use these tools as personal study aids. What could explain the gap between feeling helped and reporting better outcomes? The authors are careful to say the near-zero correlation is at least as likely to reflect a measurement artifact as a true gap between perception and attainment.

Their performance measure is a single self-reported tier that collapses every exam across every block and every year of medical school into one ordinal category, subject to recall bias, social desirability bias, and inaccurate estimation. A student whose grades improved meaningfully since beginning to use AI would still be asked to report a single composite category, and the survey did not distinguish performance before and after AI adoption. The measure also captures only one kind of academic outcome: examination scores say nothing about literature searching, writing efficiency, communication skills, clinical reasoning during patient encounters, or preparation for Objective Structured Clinical Examinations, all plausible domains of AI benefit that the survey never assessed.

Still, the authors consider mechanisms that could produce a genuine perception-performance gap, and one stands out for its psychological resonance. AI tools may provide a subjective sense of understanding or efficiency that does not translate to measurable gains on traditional exams, a phenomenon consistent with fluency illusions, in which ease of information processing is mistaken for depth of learning. Students may feel more confident or prepared after consulting AI, but that confidence may reflect familiarity with the material rather than genuine mastery.

Frequent AI users may also differ from peers in study habits, motivation, or prior achievement in ways that confound simple correlations. The authors list several patterns that could produce the observed result: AI may genuinely provide little measurable benefit on traditional exams, it may benefit some courses but not others with the overall average insensitive to that variation, it may benefit domains the survey did not measure, or students who are already struggling may adopt AI more heavily, producing selection effects that mask a real benefit. The cross-sectional design and coarse performance measure cannot distinguish among these possibilities, and the authors urge readers to treat the absence of a usage-performance correlation as a hypothesis to be tested with better data, not as evidence that AI does not help.

Two further observations add texture. The strong correlation between self-reported exam performance and self-rated academic standing (ρ = 0.60, P < 0.001) should not be read as independent validation of either measure, since both are global self-assessments completed by the same respondent in the same survey and likely reflect a stable self-view rather than convergent measurement of an external outcome. This is consistent with meta-analytic evidence that medical students' self-assessments correlate only modestly with objective criteria, with a pooled correlation of approximately 0.21.

Separately, the usage-benefit association was of similar magnitude among pre-clinical (ρ = 0.71) and clinical students (ρ = 0.65), so the pattern does not appear specific to one stage of training. Reported AI use did differ by class year, with fourth-year students reporting Often or Always use less often than students in the first three years (28% versus 62%, P = 0.003), an exploratory comparison the authors cannot explain with these data but suggest may relate to differences in assessment format, time in clinical placements, and residency application activity during the final year. The limitations are extensive and the authors state them plainly.

The response rate of 24.4% means students with pre-existing interest in AI or greater comfort with technology may have been more likely to respond, likely overstating usage in the wider student body. The survey did not capture how students used AI, whether passive reading of AI-generated summaries or active problem-solving with AI as a Socratic tutor, which specific tools they used, for which courses, or at what stages of studying, even though the mode of use is plausibly a stronger predictor of outcomes than raw frequency. Frequency anchors such as Sometimes and Often were left to participant interpretation, a non-differential misclassification that on average attenuates observed associations toward the null.

The instrument was not formally pilot tested, the study was conducted at a single institution, and with N = 103 the study was powered to detect Spearman correlations of |ρ| ≥ 0.27 at 80% power, so smaller associations cannot be excluded. Item-level analyses showed all three benefit items were positively associated with usage, with the exam grades item the weakest (ρ = 0.53 against 0.63 for conceptual understanding and 0.64 for clinical reasoning confidence), and removing it left the composite association essentially unchanged. If future work using objective, granular performance data reproduces a perception-performance gap of the kind tentatively suggested here, the implications for medical educators would be practical rather than prohibitive.

Rather than discouraging AI use, educators might focus on helping students develop metacognitive skills to critically assess whether their AI-assisted studying is producing genuine learning gains. Concrete strategies could include structured reflection exercises in which students compare predicted exam performance with actual outcomes, self-testing protocols that provide objective feedback on knowledge retention, and comparative study designs in which students alternate between AI-assisted and traditional methods and track their performance. Students might pair every AI-generated explanation with a retrieval-practice step, such as closing the tool and attempting to reproduce the explanation unaided, or with a self-test drawn from faculty-validated question banks, so that the felt sense of understanding is checked against performance before an examination does it for them.

The authors call for future studies using institutionally verified course-level records, stratification by training phase, direct measurement of how students use AI, broader outcomes including OSCEs and standardized patient encounters, longitudinal and quasi-experimental designs, and multi-institutional replication across allopathic, osteopathic, and international programs. With more than half of medical students at this institution reporting near-constant AI use, the urgency of that agenda is hard to overstate.

Subject of Research: Generative AI use and self-reported academic outcomes among medical students

Article Title: Association of generative artificial intelligence use with perceived academic benefit and reported performance among medical students

Article References: Manam, S., Escobar, V., Zimmerman, C., Zimmerman, T., Fakhoury, P., Haddad, C., Dalman, D., Roukoz, R., & Dijkstra, P. (2026). Association of generative artificial intelligence use with perceived academic benefit and reported performance among medical students. Discover Education, 5(1), Article 1218. https://doi.org/10.1007/s44217-026-02350-8

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02350-8

Keywords: generative AI, medical education, medical students, survey research, artificial intelligence, Discover Education, Association, generative, artificial, intelligence, perceived, academic