AI summaries were linked to weaker reading—not proved to cause it
Among 608 undergraduates in one Chinese city, heavier self-reported reliance on AI summaries correlated with lower comprehension. Motivation weakened the association, but the one-session design cannot establish direction.
The short version
- What happened
- The cross-sectional study involved 608 students in 13 classes across three universities in Changchun, China.
- Who it affects
- Students need study designs that preserve engagement with the original text, while teachers need evidence before treating every use of an AI summary as harmful.
- What remains unproven
- Cross-sectional design, self-reported use and processing, one-city sample, researcher-developed test, unobserved tool behaviour and incomplete handling of class clustering.
Keep in mind
Cross-sectional design, self-reported use and processing, one-city sample, researcher-developed test, unobserved tool behaviour and incomplete handling of class clustering.
Source analysis · 1 linked source. Primary analysed source date: 16 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The association is real; the direction is unresolved
Students who said they relied more heavily on AI-generated summaries also scored lower on a researcher-built deep-reading test, but the study cannot show that the tools caused the difference. Among 608 undergraduates at three universities in Changchun, China, summary reliance correlated with comprehension at −0.24 and with self-reported elaborative processing at −0.41. Students who reported connecting, restating and questioning ideas more deeply tended to score better.
Participants came from 13 intact classes and completed everything in one supervised session. They had 25 minutes to read two unfamiliar expository passages of about 1,100 and 1,300 words and answer 20 open or short-answer questions, followed by 15 minutes of questionnaires. The sample was 53.6% female with a mean age of 19.9, and its subject mix leaned towards commerce.[1]
Motivation changed the statistical pathway
The authors modelled a pathway in which summary reliance related to less elaborative processing, which in turn related to lower comprehension. With 5,000 bootstrap resamples, the indirect association was negative for students at low and average reading motivation, but not statistically detectable at high motivation. This is consistent with motivated readers using summaries without abandoning effortful engagement, but it is not direct observation of how any student used a tool.
The measures deserve attention. AI-summary use and elaboration were self-reported habits, not activity logs. The comprehension test was developed for this project, reviewed by two teachers and piloted on 42 other students. Two raters double-scored 20% of responses with high agreement; the remainder had one rater. The authors did not retain class identifiers, so clustering was handled through sensitivity calculations rather than direct multilevel modelling.
Most importantly, the design was cross-sectional. Students with weaker reading skills may turn to summaries more often, a third factor may influence both, or the proposed sequence may operate differently over time. Statistical mediation in a single sitting does not establish a causal chain, even when the coefficients and bootstrap intervals are precise.[1]
The distinction that matters
What the study measured—and what it did not
All measures were collected in one session, so the arrows describe a statistical model rather than proven causal order.
Reported use
Six-item AI-summary reliance scale
Correlated −0.24 with comprehension
Proposed pathway
Less elaboration linked use to scores
Absent at high motivation
Missing test
No random assignment or usage logs
Cause and direction unresolved
Teachers can test a safer division of labour
The most useful implication is narrower than banning summaries. A course can require students to read first, write their own account, compare it with the machine version and defend discrepancies using the source text. That keeps the tool available while making verification and elaboration visible. The current paper suggests that motivation and active processing are promising intervention targets; it does not show that this sequence improves learning.
The evidence check is limited to three institutions in one city, a commerce-heavy sample and one assessment session. Tool names, prompts, source quality and verification behaviour were not logged, and measurement invariance across gender, year and institution was untested. No funding or commercial conflicts were declared. Confidence would rise with randomised classroom comparisons, longitudinal testing, actual tool-use logs, validated comprehension measures and outcomes across subjects and countries. For now, the defensible conclusion is association, not cognitive damage.[1]