Does digital confidence show readiness for AI learning?
An analysis of 118,764 Grade‑8 students across 35 education systems found that very high self-reported ICT confidence did not translate into higher tested digital competence. ICILS 2023 measured neither AI use nor AI literacy, so the profiles are possible foundations—not a readiness test.
The short version
- What happened
- The analytic sample included 118,764 Grade-8 students with complete scores on four ICT-disposition scales, representing 87.6% of the relevant ICILS 2023 sample across 35 education systems.
- Who it affects
- Teachers should pair student confidence with practical demonstrations of verification, judgement and safe AI use.
- What remains unproven
- The study used cross-sectional secondary data and cannot establish developmental or causal relationships.
Keep in mind
The study used cross-sectional secondary data and cannot establish developmental or causal relationships.
ICILS 2023 measured neither AI use nor AI literacy, so the profiles are indirect candidate foundations only.
The 12.4% excluded for missing indicators were less advantaged and scored substantially lower, producing positive selection.
Pooled profiles across 35 systems are statistical summaries and were not reproduced identically in every system.
Self-efficacy, attitudes and expected future use do not capture verification, ethics, privacy or human-AI interaction skill.
Source analysis · 1 linked source. Primary analysed source date: 9 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The direct answer: confidence and tested competence diverged
Very high digital confidence did not identify the strongest tested digital competence in this international dataset. The 7.9% of students assigned to an 'efficacy-dominant' profile reported general ICT self-efficacy 2.16 standard deviations above the pooled mean and specialised self-efficacy 1.66 standard deviations above it. Their mean score on the separately tested computer and information literacy scale was 486.7—only 1.7 points above the 485.0 mean for the much larger average profile.
That 1.7-point contrast had a standard error of 1.38, an effect size of 0.017 and p=0.19. In other words, it was indistinguishable from zero. Students in the attitude-oriented profile scored highest at 495.4, while the low-disposition profile averaged 472.1. Even the widest profile difference was about 23 points, or 0.23 of the ICILS international standard deviation. The paper's useful warning is that confidence can be informative without being a proxy for demonstrated capability.[1]
What the 118,764-student analysis actually measured
The study reanalysed ICILS 2023, retaining 118,764 Grade-8 students across 35 education systems who had complete scores on four questionnaire scales. That was 87.6% of the relevant sample. Models involving student characteristics used 115,537 students with complete predictor data. The four profile indicators captured confidence in general digital tasks, confidence in specialised tasks such as programming or configuration, positive attitudes toward ICT in society, and expectations of future ICT use in study or work.
Tested computer and information literacy was kept outside the profiles as an external criterion. ICILS represents proficiency through five plausible values on an international scale with mean 500 and standard deviation 100. The authors combined those values under standard survey methods, applied the study's sampling weights and used 75 jackknife replication zones for uncertainty. Equal senate weights made every participating education system contribute the same total weight rather than letting the largest populations dominate.[1]
Four statistical profiles emerged
Weighted latent-profile models compared one to six groups. The selected four-profile solution placed 55.9% of students in an average profile, 25.2% in an attitude-oriented profile, 11.0% in a low-disposition profile and 7.9% in the efficacy-dominant profile. The attitude-oriented group combined positive views of technology and high expected future use with only modest self-efficacy. The low-disposition group was below the pooled mean across every indicator.
These are model-derived clusters, not natural or permanent kinds of learner. Their boundaries depend on the included scales, covariance assumptions and pooled sample. The authors deliberately avoid naming either intermediate group 'high preparedness'. Independent estimation recovered the same general configuration in about two thirds of education systems, which supports some cross-system relevance but falls well short of universal student types.[1]
Why this is not an AI-readiness assessment
ICILS 2023 contained no measure of AI literacy or AI use. It did not ask about generative-AI experience, detecting fabricated output, judging provenance, protecting data, reasoning about bias, ethical decision-making or managing a human-AI workflow. The authors describe the four dispositions as candidate foundations that may matter for AI-mediated learning. The data cannot show that a profile learns more effectively with a tutor, writes better with a model or uses AI more safely.
The analysis is also cross-sectional. It cannot show whether confidence changes competence, whether competence changes confidence or whether schooling and home access shape both. Calling the efficacy-dominant result a confidence-competence mismatch is a reasonable interpretation of the measured pattern, not proof that those students are overconfident in every setting. A dedicated AI assessment and direct observation of learning would be required for that conclusion.[1]
What teachers and education systems should do
Schools should not allocate advanced AI access, support or responsibility from a student's confidence alone. Short practical checks can ask learners to verify a source, compare model output with evidence, explain uncertainty, protect sensitive information and revise a flawed answer. Students who sound technically assured may still need structured practice; quieter students may demonstrate stronger judgement. Assessment should therefore combine knowledge, observable performance and reflection.
The profiles can guide questions, not labels. The low-disposition group may need access, confidence and foundational skills, while the attitude-oriented group may benefit from translating enthusiasm into practice. Yet support must be based on local assessment and student voice. A statistical class should never become a permanent record, a gate to opportunity or a justification for lower expectations. Teachers need time, validated instruments and inclusive curriculum rather than another predictive score imposed on pupils.[1]
Selection bias and international pooling matter
The 12.4% of students excluded for missing profile indicators were not random omissions. They scored close to one international standard deviation lower in tested competence and came from less advantaged backgrounds. The retained sample is therefore positively selected, the low-disposition group is probably underrepresented and reported socioeconomic gradients may be conservative. This is especially important if policymakers hope to use the analysis for equity planning.
Pooling 35 systems also averages over different curricula, infrastructure, languages and inequalities. Equal system weights support an international pattern but do not estimate the number of students in each profile worldwide. National and local teams should inspect their own survey design, missingness and subgroup results. The paper reports that gender contrasts were larger and more consistent than socioeconomic contrasts on a common probability scale, but those associations are not biological explanations and should not harden into stereotypes.[1]
Funding, disclosure and what would change the assessment
The Regional Government of Castilla y León supported the work under grant SA217P23 and, according to the authors, had no role in design, analysis, interpretation or writing. The authors reported no commercial or financial conflict. One author was a Frontiers editorial-board member; the paper states this did not affect review or the decision. The team also disclosed using Anthropic's Fable 5 to assist database preparation, R-code development, drafting and editing, with human review and responsibility retained.
The assessment would change with longitudinal studies that measure AI literacy directly and observe real AI-supported learning. Researchers should validate profiles within individual systems, include students missing from standard assessments, and test whether targeted support improves outcomes without reinforcing inequality. Dedicated instruments should cover verification, uncertainty, data protection, fairness and human accountability. For now, the strong conclusion is not who is ready for AI; it is that self-confidence cannot safely stand in for measured competence.[1]