Research Article
Perceived Learning vs. Engagement in AI-Assisted Homework: A Comparative Study of ChatGPT Use Across High School, University, and Teachers in Sonora, Mexico (2024–2025)
Synthesis: Torres-Peralta and colleagues surveyed 1477 high school students, university students, and teachers in Sonora, Mexico, comparing what students say they learn from ChatGPT with what they report doing with its output. They built a Learning Engagement Index (LEI, 0–1) from three behaviors — reading, modifying, adding one's own ideas — and set it against a single 0–10 self-rating. Adoption and self-rated learning both rose with level, yet the two barely moved together (pooled correlation 0.052). The authors read this as a warning for assessment validity, not a verdict on learning: take-home work no longer proxies for mastery.
Key Findings
- Adoption climbs with level, for students and teachers alike. University students and instructors reported more use than their high school counterparts.
- Perceived learning rises only modestly with level. Self-rated learning was slightly higher at university, a real but small difference.
- Engagement is common and also higher at university. Most respondents said they read and adapt AI output rather than copying it.
- Perceived learning does not track engagement. Students who edit AI output are not the students who report learning more; the association is essentially absent.
- Teachers remain the first stop for help. Asking a teacher was still the most common response to being stuck, with ChatGPT a complementary rather than substitutive option.
- Ethical framing matures with level and stays thin at the bottom. University students were more likely to name disclosure and integrity concerns than high school students.
What the survey measured
Between October 2024 and May 2025 the authors collected 1477 responses from public and private institutions in Hermosillo, Ciudad Obregón, and Guaymas: 602 high school students, 765 university students, 42 high school teachers, 68 university teachers. Students answered Likert items on perceived learning, trust in accuracy, and time savings, plus open-ended prompts on advantages, risks, and misuse; the coded text makes it mixed-methods work, though headline claims are quantitative.
The novelty is the LEI. Three self-reports of what a student does with a ChatGPT answer are coded, normalized to 0–1, and averaged, leaving an analytic sample of 1108 (1105 with a valid learning rating). Still self-reported, it moves past adoption counts toward engagement patterns self-regulated learning theory flags: read, modify, integrate.
Where the level gap comes from
University respondents reported larger time savings — 32.6% claimed 1–3 h per task, 13.5% more than 3 h — while high school respondents clustered at 15 min–1 h (43.2%). University students described more differentiated use — ChatGPT for structuring, Google for exploration, teachers for validation — while high school narratives were instrumental: simpler wording, a quick idea, a way to finish.
The two levels were surveyed at different times, so a 16.3 percentage point student adoption gap and a 35.8 percentage point teacher gap cannot be attributed cleanly to academic maturity; exposure, normalization, policy, and the calendar all moved during the window. The authors call it a gradient, not a developmental progression.
Reading the perception–engagement gap
The weak association is the finding to keep. Perceived learning was one 0–10 item; the LEI, three behavior items. Neither is a learning outcome, and the authors say so: no linkage does not mean students fail to learn or that their self-ratings are wrong. It means confidence that the tool helped poorly predicts whether they read, reshaped, or extended what it produced — and so poorly predicts critical thinking about it.
The ethics data point the same way. High school students judged use by whether the result looked correct and whether copying was visible; university students raised verification, authorship, and bias. Few at either level had a vocabulary for offloading risk, and those who did described it behaviorally — laziness, difficulty starting without the tool. The Copilot experiment supplies the mechanism: faster, more complete work with identical self-rated learning.
What this means for practice
- Instructors: Grade the process, not just the product. Collect the prompt, draft answer, and annotated revisions so the three LEI behaviors become visible, not inferred from a polished page.
- Instructional designers: Make read–modify–integrate required, credited steps, and design transfer items a fluent model answer cannot satisfy alone.
- Administrators: Treat homework completion as formative evidence, not mastery evidence, and say so in policy; teachers at both levels reported acting without guidance.
- Researchers: Pair perceived-learning items with behavioral or performance measures before claiming benefit; the LEI is a usable start but needs outcome validation.
- Policymakers: Fund connectivity and devices alongside AI guidance; students described uneven access, and the authors warn unmanaged integration widens gaps.
Limitations
- Both constructs are self-reported; the LEI records what students say they do, not what they were observed doing.
- No direct learning outcome was measured, so the gap says nothing about actual mastery.
- The two levels were surveyed at different times, confounding level with exposure and timing.
- The Copilot experiment had 16 participants on one Python task, so it is not generalizable.
- Sampling was regional and school-based, so the rates are not national estimates for Mexico.
Citation
Torres-Peralta, R., Cirett-Galán, F., Heras-Sanchez, M. del C., Lerma-Molina, K., & Espinoza-Flores, I. (2026). Perceived Learning vs. Engagement in AI-Assisted Homework: A Comparative Study of ChatGPT Use Across High School, University, and Teachers in Sonora, Mexico (2024–2025). Future Internet, 18, 122.