AI in Education: Helpful Tutor, Study Buddy or Threat to Learning?

In just a few years, Generative AI (GenAI) has gone from novelty to near-universal in UK universities and worldwide. The Higher Education Policy Institute’s (HEPI) 2026 survey of 1,054 undergraduates found that 95% of students now use AI in some form and 94% use GenAI to support assessed work (Stephenson and Armstrong, 2026). So, the question is no longer whether students use AI, but whether it helps them learn, and whether our assessments are still fit for purpose. Research at The Open University, across the UK and at leading universities worldwide gives a surprisingly consistent answer: it depends on how AI is designed and used by both students and educators.

Image 1: ChatGPT chat with AI.

Personalised support at scale

In 1984, Benjamin Bloom showed that one-to-one tutoring dramatically improves student performance (Bloom, 1984). However, the problem has always been (and still is) the cost of one-to-one tutoring. The OU’s Knowledge Media Institute (KMi) is testing whether AI can close that gap. Its AI Digital Assistant (AIDA) and Smart Assessment and Guided Education with Responsible AI (SAGE-RAI) projects are exploring whether AI can offer this kind of personalised support affordably and at scale. AIDA, built with the Institute of Educational Technology (IET), has been piloted in real learning contexts with evidence gathered from hundreds of students and staff in the OU (Coughlan et al., 2025). KMi researchers have also proposed an AI-Enhanced Personal Learning Environment in which AI helps learners set goals, reflect on progress and adapt their learning strategies, with early evidence from pre-service teachers (Mikroyannidis and Armakolas, 2025). In other words, AI used to build self-regulation rather than replace it.

Further evidence that this can work comes from Harvard University. In a randomised controlled trial with 194 physics undergraduates, students learned significantly more in less time with an AI tutor than in an in-class active learning session (Kestin et al., 2025). It is, however, worth noting that this was not off-the-shelf ChatGPT: the tutor was a purpose-built bot developed at Harvard University with pedagogical design embedded.

Similar work is under way in UK schools. In a 2025 exploratory trial on the Eedi maths platform, an AI tutor supervised by human tutors produced a 5.5 percentage-point increase in students correctly answering novel questions, compared with expert human tutoring alone (LearnLM Team, Google and Eedi, 2025). A larger trial, aiming to recruit 1,200 students across ten UK secondary schools, was registered in 2026; results have not yet been published (Brazão, 2026).

Design matters

The same technology can also undermine learning. A team at the University of Pennsylvania ran a field experiment giving nearly a thousand high school maths students one of two GPT-4 tutors: a standard ChatGPT-style interface or one designed to safeguard learning. When access was taken away, students who had used the standard version performed worse than those who never had access, with a 17% reduction in grades. The safeguards in the tutoring version largely mitigated these negative effects (Bastani et al., 2025).

Massachusetts Institute of Technology (MIT) Media Lab’s ‘Your Brain on ChatGPT’ study points the same way. Researchers used electroencephalography (EEG) to record the brain activity of 54 subjects writing essays with ChatGPT, a search engine, or no tools. The ChatGPT group showed the least brain connectivity, and most of them struggled to quote from their own essays shortly afterwards (Kosmyna et al., 2025). But students who wrote unaided first and only then used ChatGPT adapted well, which suggests timing matters. The study is a preprint that, at the time of writing, had not been peer-reviewed, and its lead author has asked the media not to describe the findings with terms such as ‘brain rot’ or ‘brain damage’. To be continued!

The lesson from Pennsylvania, Harvard and MIT is the same: AI that does the thinking for students can undermine learning, while AI that makes students think can help. How can we, as educators, ensure that?

Assessment, Assessment

GenAI also creates a serious assessment problem. In the ‘Bob or Bot’ study by the OU’s School of Computing and Communications, markers assessed 90 exam scripts without knowing some were written by ChatGPT. Every ChatGPT script for the undergraduate modules passed, and all the introductory ones earned a distinction (Richards et al., 2024).

The University of Reading found the same in psychology. Researchers secretly submitted GPT-4 answers for 33 fake students. The AI submissions went undetected in 94% of cases and typically scored half a grade higher than real students (Scarfe et al., 2024).

Detection isn’t a reliable fix, unfortunately!

In a study of 10,725 OU assessments, male students, younger students and those with lower prior qualifications were more likely to be flagged as using AI (Gooch et al., 2024). The authors stress that flagging is not proof of misconduct, which makes this a fairness concern rather than evidence of cheating. A cross-disciplinary OU study found that training markers to spot AI considerably improved detection but also caused a large rise in false positives and substantially increased workload (Hardie et al., 2024). Students feel this tension: HEPI reports growing anxiety about misconduct accusations (Stephenson and Armstrong, 2026).

So what?

Across many recent studies the message is consistent: design AI for learning, redesign assessment, and teach AI literacy rather than trying to police AI use. Institutions are moving. The OU now uses a three-tier policy: AI prohibited (Category 1), AI permitted responsibly at the student’s discretion (Category 2), or AI required as part of learning outcomes (Category 3) (Auton and Kouadri Mostefaoui, 2026). Oxford has given all students and staff access to ChatGPT Edu, with training courses on getting started with generative AI (University of Oxford, 2025). HEPI recommends a structured introduction to AI for first-year students, clear guidelines with both AI-free and AI-supported assessment, and tools provided for all students (Stephenson and Armstrong, 2026).

The question isn’t AI or no AI but which AI, used when, and for what? A Socratic tutor that asks questions is a very different tool from a chatbot that hands over answers. Getting that distinction right is now one of the most important design challenges in education.

Soraya Kouadri Mostéfaoui, Senior Lecturer and Scholarship Lead for the School of Computing and Communications

Dr. Soraya Kouadri Mostéfaoui is Senior Lecturer in Computing, postgraduate research tutor and Scholarship Lead in the School of Computing and Communication at The Open University. Her research interests include the design and uptake of human-centric digital systems, with a focus on the relationships between people and the digital world, addressing issues surrounding AI, education, ethics and healthcare.

Image 2: Soraya Kouadri Mostéfaoui. 

References

Auton, S. and Kouadri Mostefaoui, S. (2026) ‘Responsible generative AI use in student assessments: the carrot vs the stick’, INTED2026 Proceedings. 20th International Technology, Education and Development Conference, Valencia, Spain, 2–4 March. Valencia: IATED, article 2142. Available at: https://doi.org/10.21125/inted.2026.2142. 

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025) ‘Generative AI without guardrails can harm learning: evidence from high school mathematics’, Proceedings of the National Academy of Sciences, 122(26), e2422633122. Available at: https://doi.org/10.1073/pnas.2422633122. 

Bloom, B.S. (1984) ‘The 2 sigma problem: the search for methods of group instruction as effective as one-to-one tutoring’, Educational Researcher, 13(6), pp. 4–16. Available at: https://doi.org/10.3102/0013189X013006004. 

Brazão, V. (2026) ‘Testing the efficacy of AI tutoring in secondary mathematics: a scaled RCT in UK classrooms’, AEA RCT Registry. Available at: https://doi.org/10.1257/rct.18079-1.1. 

Coughlan, T., Rienties, B., Bektik, D., Ullmann, T., Herodotou, C., Tessarolo, F., Edwards, C., Coughlan, E., Devine, P. and Domingue, J. (2025) AIDA on OpenLearn: design and evaluation report. Milton Keynes: Institute of Educational Technology, The Open University. Available at: https://oro.open.ac.uk/106176 (Accessed: 19 October 2026).

Gooch, D., Waugh, K., Richards, M., Slaymaker, M. and Woodthorpe, J. (2024) ‘Exploring the profile of university assessments flagged as containing AI-generated material’, ACM Inroads, 15(2), pp. 39–47. Available at: https://doi.org/10.1145/3656478. 

Hardie, L., Lowe, J., Pride, M., Waugh, K., Hauck, M., Ryan, F., Gooch, D., Patent, V., McCartney, K., Maguire, C., Richards, M. and Richardson, H. (2024) Developing robust assessment in the light of generative AI developments. Milton Keynes: The Open University and NCFE. Available at: https://www.scholarship-exchange.open.ac.uk/articles/report/Developing_robust_assessment_in_the_light_of_Generative_AI_developments_Research_paper_/27960792 (Accessed: 19 October 2026).

Kestin, G., Miller, K., Klales, A., Milbourne, T. and Ponti, G. (2025) ‘AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting’, Scientific Reports, 15, 17458. Available at: https://doi.org/10.1038/s41598-025-97652-6. 

Kosmyna, N., Hauptmann, E., Yuan, Y.T., Situ, J., Liao, X.-H., Beresnitzky, A.V., Braunstein, I. and Maes, P. (2025) ‘Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task’ [Preprint]. arXiv. Available at: https://doi.org/10.48550/arXiv.2506.08872. 

LearnLM Team, Google and Eedi (2025) ‘AI tutoring can safely and effectively support students: an exploratory RCT in UK classrooms’ [Preprint]. arXiv. Available at: https://doi.org/10.48550/arXiv.2512.23633. 

Mikroyannidis, A. and Armakolas, S. (2025) ‘Using generative AI to enhance personalised and self-regulated learning’, EdMedia + Innovate Learning 2025. Barcelona, Spain, 19–23 May. Association for the Advancement of Computing in Education (AACE), pp. 249–258. Available at: https://oro.open.ac.uk/103711/ (Accessed: 19 October 2026).

Richards, M., Waugh, K., Slaymaker, M., Petre, M., Woodthorpe, J. and Gooch, D. (2024) ‘Bob or Bot: exploring ChatGPT’s answers to university computer science assessment’, ACM Transactions on Computing Education, 24(1), article 5, pp. 1–32. Available at: https://doi.org/10.1145/3633287. 

Scarfe, P., Watcham, K., Clarke, A. and Roesch, E. (2024) ‘A real-world test of artificial intelligence infiltration of a university examinations system: a “Turing Test” case study’, PLOS ONE, 19(6), e0305354. Available at: https://doi.org/10.1371/journal.pone.0305354. 

Stephenson, R. and Armstrong, C. (2026) Student generative AI survey 2026. HEPI Report 199. Oxford: Higher Education Policy Institute. Available at: https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/ (Accessed: 19 October 2026).

University of Oxford (2025) Oxford becomes first UK university to offer ChatGPT Edu to all staff and students. Available at: https://www.ox.ac.uk/news/2025-09-19-oxford-becomes-first-uk-university-offer-chatgpt-edu-all-staff-and-students (Accessed: 19 October 2026).