The most interesting question about AI in language learning isn’t whether the tool helps students write better sentences. It’s what happens to the student in the process. Du and Alm’s 2024 study in Education Sciences takes that question seriously, using Self-Determination Theory (SDT) to examine how ChatGPT shapes the psychological needs of English for Academic Purposes (EAP) students in New Zealand.
I want to walk through what they found, because the tension at the center of this research matters for anyone thinking about how to bring AI into language classrooms responsibly.
The study is qualitative, grounded in semi-structured interviews with 24 postgraduate international students at a New Zealand university. Du and Alm used SDT as their theoretical frame, focusing on three basic psychological needs: autonomy, competence, and relatedness. They analyzed the interview data through thematic analysis using MAXQDA, with two independent coders to keep the interpretation reliable.
It’s worth noting the timing. This paper was published in 2024, which means the data was almost certainly collected when ChatGPT was still relatively new, during that early wave of discovery and experimentation. The AI field has shifted dramatically since then.
Models are far more capable, institutions have had years to develop policies, and the research base has grown significantly. That doesn’t make these findings irrelevant, but it does mean we’re reading a snapshot of a moment that has already passed.
Where ChatGPT Delivered: Autonomy and Competence
On autonomy, the findings are clear and unsurprising. Students told Du and Alm they could learn at their own pace (10 out of 24 participants), take ownership of what they practiced and when (n=11), and work without the fear of being judged for mistakes (n=13). That last point came up especially among students who felt anxious about asking questions in traditional classrooms. ChatGPT gave them a space to practice without the social cost of looking incompetent.
Competence was the strongest category in the entire dataset. Personalized feedback and error correction had the highest coding frequency (n=41), followed closely by context-specific examples (n=40).
Students described getting help with grammar, vocabulary, pronunciation, and academic writing conventions that felt immediate and tailored to their particular struggles. Nineteen students reported increased self-efficacy, a kind of confidence loop where feedback led to improvement, which led to a willingness to take on harder tasks.
I’ve seen similar patterns in the research on AI-enhanced feedback that Hawkins, Taylor-Griffiths, and Lodge (2025) covered in their work on feedback literacy. The mechanism is consistent: when students get timely, specific responses to their writing, their confidence grows. The question is always what happens to that confidence when the AI isn’t there anymore.
And that’s exactly where Du and Alm’s data starts to complicate the picture. Eleven students raised concerns about ChatGPT’s accuracy for complex academic tasks. Eight noted that the tool couldn’t adjust its language level as they improved. The feedback that felt helpful at week two sounded the same at week ten.
That’s a real limitation, and one that speaks to a pattern I’ve covered in posts on cognitive offloading (Gerlich, 2025) and metacognitive laziness (Fan et al., 2025): AI tools can build a sense of competence that depends entirely on the tool’s continued presence.
The Relatedness Problem in AI-Assisted Language Learning
The most revealing findings in this study come from relatedness, the SDT need that deals with social connection and belonging. And the data here is genuinely split.
Thirteen students described feeling a sense of companionship with ChatGPT. They felt less isolated in their language learning, particularly those who were shy or socially anxious. One participant told Du and Alm they felt “more socially connected” through their AI interactions.
But fifteen students reported the exact opposite: a lack of social belonging. AI interactions, they said, couldn’t replicate the warmth and reciprocity of talking to another person. Twelve of those students said their human interactions had actually decreased because of ChatGPT. They were turning to the tool for help they would have previously sought from classmates or instructors. The face-to-face conversations dropped. The collaborative learning dried up.
One participant raised a point that is especially interesting: peer pressure. If everyone uses ChatGPT, the competitive dynamics shift. Students start worrying about falling behind peers who use it more strategically.
Du and Alm connect this to introjected regulation in SDT, where behavior gets driven by guilt or anxiety, not genuine interest. Only one student mentioned it, so the finding is thin. But the logic is solid, and I wouldn’t be surprised to see future studies confirm it at scale.
Du and Alm put it clearly: “the varied experiences of relatedness suggest that AI tools like ChatGPT may satisfy this need for some learners while thwarting it for others” (p. 13). That’s a clear-eyed reading of messy data, and I respect the researchers for not smoothing over the contradiction.
What This Means for AI Pedagogy
The core insight from this study is that the three psychological needs don’t move in the same direction when AI enters the picture. Du and Alm found that ChatGPT could support autonomy and competence at the same time, a student feels free to ask questions and improves their grammar in a single session, but those gains sometimes came at the expense of relatedness.
For some students, the autonomy and competence benefits were strong enough to carry the overall motivation. For others, the loss of human connection undermined everything else.
Du and Alm frame this as a possible shift in how the three needs interact in AI-mediated environments: “the weights of these needs in determining overall motivation might shift, with autonomy and competence potentially compensating for reduced relatedness in some cases” (p. 13).
That’s a significant theoretical claim, and I think it’s right, at least partially. But it also has a practical edge. If autonomy and competence are the easy wins for AI tools, and relatedness is the casualty, then the pedagogical design work needs to focus precisely on what AI can’t do.
That means pairing ChatGPT with collaborative tasks, peer review exercises, in-class discussions about what students are learning from the tool, and structured opportunities for human interaction.
The authors recommend exactly this, and I think it’s the right instinct. The mistake would be treating ChatGPT as a standalone learning companion and letting it replace the social infrastructure of a classroom.
I also think we need to contextualize these findings carefully. The sample is 24 students in one program at one university. It’s qualitative and exploratory, which means it’s useful for generating hypotheses but not for drawing broad conclusions. The authors describe the study with full transparency, and they call for larger, mixed-methods follow-ups across different settings. That’s exactly what’s needed.
The technology will keep evolving. The pedagogical questions about what students actually need to thrive, autonomy, competence, connection, and a reason to care, those won’t change.
References
- Du, J. D., & Alm, A. (2024). The impact of ChatGPT on English for Academic Purposes (EAP) students’ language learning experience: A Self-Determination Theory perspective. Education Sciences, 14(7), 726. https://doi.org/10.3390/educsci14070726
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544
- Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006
- Harkins-Brown, A. R., Carling, L. Z., & Peloff, D. C. (2025). Artificial intelligence in special education. Encyclopedia, 5(1), 11. https://doi.org/10.3390/encyclopedia5010011