Better performance during study doesn't guarantee memory afterward
The most relevant study here compared undergraduate programming students who had access to ChatGPT-4.5 with students using conventional web search without AI [5]. The ChatGPT group showed higher performance, but lower retention and reduced ownership of the work. That pattern matters because it separates two things people often blur together: how well you do while the tool is open, and how much stays with you once it's gone.
The retention check in that study came 48 hours after the session, not weeks [5]. So even the clearest signal of a performance-retention trade-off describes a short window. Whether the gap widens, holds steady, or closes over weeks is simply not something these studies can tell us.
Why unrestricted use might weaken retention
One plausible explanation is that unrestricted ChatGPT use lets a learner skip the effortful steps that build memory. Retrieval practice — trying to recall something before looking it up — is a strategy known to strengthen long-term memory [4]. When an AI can supply the answer on demand, those recall attempts may never happen, and the study session becomes smoother but less memorable.
A framework for comparing AI-assisted and traditional learning interventions treats retention, independent performance, and transfer as distinct outcomes rather than interchangeable ones [3]. That framing is useful here: a tool can lift performance on an assessment while leaving retention and the ability to work independently unaffected or worse. The framework itself is a way of organizing comparisons, not a demonstration that unrestricted use causes any particular retention outcome [3].
Structured AI use looks different from unrestricted use
The studies that report gains tend to embed ChatGPT inside a specific teaching method rather than leaving students to use it freely. In a randomized crossover study with undergraduate medical students learning about acute abdomen, ChatGPT-4-assisted personalized learning improved academic outcomes compared with traditional methods [1]. A quasi-experimental study with high school students found ChatGPT-integrated discovery learning improved mathematical literacy on three-variable linear equation systems relative to conventional instruction [2].
Both of those studies measured performance close to the learning period and did not assess retention weeks later [1][2]. The second paper's authors themselves flag long-term retention as something still to be examined [2]. So these results support the idea that guided, structured AI use can help learning in the moment — they don't show that it protects memory over time.
There is also a hint that how ChatGPT is used matters for memory specifically. A study of ChatGPT-assisted retrieval practice, which pairs AI feedback with the act of recalling material, describes retrieval practice as a route to long-term memory retention and reports benefits for exam performance [4]. That is consistent with the idea that AI helps when it supports recall rather than replaces it, though the study did not measure retention weeks out either.
What this means if you're studying with ChatGPT
The evidence points to a practical distinction rather than a verdict on ChatGPT itself. Unrestricted use — asking for answers as you go — is the condition associated with higher performance but lower retention in the one direct comparison available, and that comparison covered only 48 hours [5]. Structured use, where ChatGPT is built into a defined learning activity, is associated with better immediate outcomes than traditional instruction [1][2].
If retention weeks later is what you care about, the safest reading is that the question remains open. The studies that would answer it — measuring the same learners weeks after AI-assisted versus traditional study — have not been done in this evidence set, and at least one research team explicitly lists long-term retention as future work [2]. In the meantime, the pattern across these studies suggests that keeping some effortful recall in your study routine, rather than outsourcing every step to the tool, is the choice most consistent with durable memory [4][5].
Sources used in this answer
Impact of ChatGPT-assisted personalized learning on teaching acute abdomen to undergraduate medical students: A randomized crossover study
ChatGPT-4-assisted personalized learning improved academic outcomes in undergraduate medical students learning acute abdomen compared with traditional methods, but retention weeks later was not assessed.
Effectiveness of ChatGPT-integrated discovery learning on mathematical literacy in three-variable linear equation systems: A quasi-experimental study
ChatGPT-integrated discovery learning improved high school students' mathematical literacy on three-variable linear equation systems relative to conventional instruction, with long-term retention left for future study.
Generative AI, performance, and learning: A framework for comparing interventions across assessment regimes
The paper offers a framework treating retention, independent performance, and transfer as separate outcomes when comparing unrestricted ChatGPT use with a learner-first strategy, without providing weeks-later retention data.
ChatGPT-assisted retrieval practice and exam scores: Does it work?
ChatGPT-assisted retrieval practice was associated with exam performance benefits, framed around retrieval practice as a strategy for long-term memory retention, though delayed retention was not measured.
Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership
Undergraduate programming students using ChatGPT-4.5 showed higher performance but lower retention and reduced ownership than students using conventional web search, with retention measured after 48 hours.