Socioeconomic inequalities in linked family and classroom reports of AI evaluation

Abstract Evaluating information generated by artificial intelligence (AI) is becoming an everyday educational task. Families and classrooms offer different settings for learning this practice, yet their socioeconomic distribution is usually examined separately. This exploratory study analysed linked parent and student reports from the Programme for International Student Assessment 2025 across sixteen education systems. Among 94,262 paired respondents, the equal system mean proportion combining a lower or unfamiliar parental AI evaluation rating with never or almost never practising AI evaluation in lessons was 15.45%. This combination occurred among 24.69% of students in the lowest socioeconomic resource quintile and 8.27% in the highest, a difference of 16.42 percentage points (95% confidence interval 15.45 to 17.39). The contrast was positive in all sixteen systems. In 93,687 records, lower parental ratings were associated with a 3.87 percentage point lower probability of reporting practice at least sometimes after adjustment for school fixed effects, student backgrounds and the full parental rating of ordinary online information (95% confidence interval 3.02 to 4.73 points lower). Estimates changed little with science performance adjustment or response weighting. Allowing arbitrary missing focal responses widened the resource contrast to an identification interval of 0.69–30.40 points, with sampling uncertainty at the lower endpoint including zero. The findings identify a socioeconomic concentration of joint low reports among respondents and position linked family and classroom experiences as a useful focus for equitable AI education.