What predicts ChatGPT use among Saudi medical students?
A peer-reviewed survey of 243 students at one Saudi university linked personal innovativeness—and intention to use the tool—to self-reported ChatGPT use. Privacy concern, technophobia and guilt were not significant predictors, but the study cannot show causation, competence, safe practice or better learning.
The short version
- What happened
- The February-to-April 2026 online survey received 243 completed questionnaires from 350 invited MBBS students at King Faisal University, a 69.4% response rate; recruitment was by convenience sampling.
- Who it affects
- Medical students need supervised practice in evidence checking, disclosure and confidentiality whether or not they are enthusiastic early adopters.
- What remains unproven
- A cross-sectional survey cannot establish whether innovativeness caused later use or whether prior use shaped students’ self-description.
Keep in mind
A cross-sectional survey cannot establish whether innovativeness caused later use or whether prior use shaped students’ self-description.
Convenience sampling at one Saudi university cannot represent all Saudi medical students, other health professions or other countries.
ChatGPT use, intention and every predictor were self-reported in the same questionnaire; no platform logs or independent behaviour measures were used.
The study measured adoption constructs, not AI literacy, safe handling of patient information, clinical reasoning, learning or patient outcomes.
Intention and reported use were empirically close, and standard diagnostics cannot eliminate common-method bias.
Source analysis · 1 linked source. Primary analysed source date: 9 October 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct answer: willingness to experiment tracked reported use
Among 243 medical students at King Faisal University, the factor most clearly associated with an intention to use ChatGPT was personal innovativeness: a self-reported willingness to try new technologies. Intention, in turn, was strongly associated with students saying they used ChatGPT. Privacy concern, technophobia and guilt about recent use did not have statistically significant relationships with intention once the other variables in the researchers’ model were considered.
That result is about adoption, not whether adoption is beneficial. The study did not test medical knowledge, clinical reasoning, source verification, confidentiality practice, academic integrity or patient outcomes. It did not observe prompts or collect platform logs. It therefore cannot establish that more innovative students used ChatGPT well, that the tool improved learning, or that privacy protections can be relaxed. For medical schools, the useful signal is narrower: voluntary uptake may concentrate among students already disposed to experiment, so access to supervised, critical practice should not depend on that disposition alone.[1]
Who was studied—and how
The researchers ran an anonymous online survey from February to April 2026 among undergraduate MBBS students at King Faisal University in Saudi Arabia’s Eastern Province. Recruitment was by convenience sampling. Of 350 students invited, 243 completed the questionnaire, producing a 69.4% response rate and exceeding the authors’ a priori minimum sample of 129. Participants were aged 19 to 25, with a mean age of 21.6; 124 were men and 119 women. All academic levels were represented, including the internship year.
A 30-student pilot, excluded from the main analysis, was used to refine wording. The final questionnaire adapted previously published scales for privacy concern, technophobia, guilt, personal innovativeness, behavioural intention and reported ChatGPT use. Contextual adaptation matters: for example, the guilt items were shifted from a social-media scale to recent ChatGPT use. Reliability statistics were acceptable in this sample, but adapting a scale does not automatically prove that it captures the same construct across institutions, languages or professional cultures.[1]
What the statistical model found
Using partial least squares structural equation modelling and 5,000 bootstrap resamples, the authors estimated a standardised coefficient of 0.579 between personal innovativeness and behavioural intention, with a 95% bootstrap confidence interval from 0.476 to 0.682. The coefficient from intention to self-reported use was 0.777, with an interval from 0.709 to 0.835. The indirect association from innovativeness through intention to reported use was 0.450. These estimates were statistically significant under the study’s prespecified model.
The model explained 39.5% of the variation in intention and 60.4% of variation in reported use within this dataset. A robustness analysis adding age, gender, academic year, grade-point average and urban or rural residence preserved the pattern, although coefficient magnitudes changed. Those figures describe how well the model fits and predicts its own survey constructs; they are not percentages of students whose behaviour was caused by innovativeness, nor do they validate self-reported use against actual activity.[1]
The non-significant barriers must not be misread
Privacy concern was not significantly associated with intention in the multivariable model: beta 0.099, p equals 0.191, with a confidence interval from minus 0.049 to 0.250. Technophobia was likewise non-significant at beta minus 0.020, p equals 0.758, and guilt at beta minus 0.109, p equals 0.073. These results mean the study did not detect the hypothesised relationships with sufficient precision in this particular sample and model. They do not demonstrate that the effects are zero in all medical students.
More importantly, a coefficient about intention cannot determine whether privacy, anxiety or academic-integrity rules deserve attention. Students may use a tool despite concerns, may not recognise a confidentiality risk, or may interpret institutional expectations differently. The questionnaire measured perceived privacy concern rather than observed handling of clinical information. Medical schools still need explicit rules that prohibit identifiable patient data in unapproved systems, explain acceptable study use and provide routes for students to seek guidance without penalty.[1]
Why the design cannot establish cause or safe practice
All predictors, intentions and use measures came from the same respondents in one survey at one time. That prevents temporal ordering: students who already use ChatGPT may describe themselves as innovative, rather than innovativeness causing later use. Shared measurement can also strengthen associations through response style or social desirability. The authors ran standard diagnostics for common-method bias, but correctly state that those checks cannot rule it out. Intention and reported use were also empirically close, with a heterotrait–monotrait ratio of 0.912, warranting caution about how distinct the two constructs were in this sample.
Convenience recruitment at one university further limits generalisation. King Faisal University draws students from multiple Saudi regions, but that does not make the sample nationally representative. More than half of respondents reported household income above 15,000 Saudi riyals a month, and 82.7% lived in urban areas. Conditions could differ at other medical schools, among other health professions, or where access, language, assessment policy and institutional guidance differ.[1]
What medical educators can do now
The practical response is not to maximise adoption. It is to make responsible competence available to students who are eager, hesitant or uncertain. A supervised curriculum can ask students to compare model outputs with primary clinical guidance, trace claims to sources, identify fabricated citations, document their prompts, and explain when an answer should not be trusted. Assessment should distinguish permitted assistance from work that must demonstrate independent reasoning, and it should preserve a non-AI route where access or disability adjustments require one.
Educators should measure what matters after training: accuracy of evidence checking, disclosure of tool use, handling of confidential information, reasoning quality and transfer to unfamiliar cases. Usage frequency by itself is a poor success metric. Because the study measured no learning or safety outcome, it offers no evidence that encouraging an innovative mindset will improve medical education. Its more defensible implication is that institutions should not mistake spontaneous uptake for readiness and should not leave less adventurous students without structured opportunities to build critical AI literacy.[1]
Funding, disclosure and what would change the assessment
King Faisal University funded the research through grant KFU265657. The authors declared no commercial or financial conflict of interest. The university approved the study under reference KFU-REC-2025-OCT-ETHICS3617, participants gave electronic consent and the authors state that raw data will be made available without undue reservation. The paper also declares generative-AI use for paraphrasing. Those disclosures aid interpretation but do not remove the study’s design limits.
Confidence would increase with preregistered longitudinal research across several Saudi public and private medical schools, probability-based recruitment and objective usage records collected with appropriate privacy safeguards. Stronger evidence would connect adoption to verified AI literacy, examination performance, clinical reasoning, error detection, privacy compliance and equitable access, while comparing supervised training with standard instruction. Until those data exist, the result should remain a local association: openness to experimentation tracked reported ChatGPT use, while competence and educational benefit remain unproven.[1]