R3 4.4 April 27, 2026 AI Research Roundup
Eight studies on AI, learning, and cognition that you might want to add to your summer reading list.
In the ancient AI past - meaning about a year ago - I was cautioning against drawing too many conclusions about what “research says” about generative AI and learning. Scholarship takes time, or at least it used to. As in all things AI, the speedup is palpable.
In some ways, this moment reminds me of the surge of research that followed the emergency pivot to online, hybrid, and otherwise technology-mediated teaching during COVID-19. In the years that followed, we saw wave after wave of studies: some reflecting on the unique experience of being in higher education during that period, some attempting to take advantage of the “natural experiment” created by the sudden shift in modality, and others simply recording the outcomes of what instructors tried in the classroom out of necessity. The flood of scholarship might not have been a bad thing. That moment in history was unique for many reasons, but in higher education especially, tools and approaches that had been in the background for years suddenly moved to center stage. And many of us really did invent new techniques and gain new skills, all on a shockingly fast time scale.
This is not to say that those years were full of silver linings. COVID was a life-limiting and life-threatening ordeal, one that left us with burnout, grief, and diminished resources that are still very much with us. But that earlier flood of research is what comes to mind now, as generative AI gives us another fast-moving, field-wide disruption - this time without the context of a society-wide emergency, but with plenty of urgency all its own.
Now, even as we continue to grapple with the aftermath of the pandemic’s worst moments, there’s another research deluge, one that is much needed but also risks becoming overwhelming. For those of us in faculty development, pedagogy, instructional technology, and the learning sciences, the challenge is to keep an eye on what is emerging without getting swept away by the sheer volume of it. We are all, in one way or another, trying to pick out the work that is most relevant, most carefully designed, and most useful.
Eventually, we may see not just practical recommendations emerge from this literature, but useful, well-supported conceptual frameworks: ways of thinking about AI that help us make sense of new studies as they appear. I wouldn’t say that kind of theoretical cohesion has emerged yet. But AI-related research now crosses into almost every area of behavioral science I follow - memory, attention, motivation, reasoning, and more. Some of it illuminates how to best support big objectives like critical thinking or self-regulated learning; some of it tests ways of designing AI-supported learning environments to meet practical goals; some of it offers resources and research tools we might adapt for scholarly projects of our own.
At their best, these studies can also help us make decisions. Many of us are in leadership positions for AI initiatives - or are being asked to advise departments, programs, and colleagues - where we find ourselves at one crossroads after another. Current research will not neatly answer every question we face, but it can help us navigate the sea of possible answers.
With that, here’s a sample of the recent research I’ve been reading. Take it as a set of highlights that arose organically during my own attempts to monitor developments in the field rather than a systematic review, and keep in mind that I didn’t include reviews, meta-analyses, or opinion-style pieces. When putting it together, I concentrated mostly on empirical studies, especially from the kinds of journals I tend to read in my own field of cognitive psychology and neighboring areas of the learning sciences. Some of these studies are directly useful for pedagogical practice; others simply struck me as interesting. All could make good additions to a summer reading list for those of us keeping tabs on the unfolding research story of generative AI, learning, and cognition.
Hou, C., Zhu, G., Liu, Y., Sudarshan, V., Leng, J., Chong, L., & Tan, H. (2026). The effects of critical thinking intervention on reliance behaviors, problem-solving quality, and creativity during human–generative AI collaborative learning. Computers & Education, 247, 105576. https://doi.org/10.1016/j.compedu.2026.105576
Critical thinking is, to put it mildly, a hot topic in cultural discourse about AI. So even though this study doesn’t attempt to conclusively demonstrate effects of AI on critical thinking overall, any linkage they can establish is worth checking out. (And, I would note, at this point no other line of research in the psychology of critical thinking has established this kind of clear causal link.) This study assessed impacts of a critical thinking intervention that fit within an approximately two-hour problem-based learning assignment, incorporating structured dialogue and similar approaches to encourage students to consider a problem from multiple angles, consider counter-arguments, and also to reflect on potential biases that AI might introduce during the process. Students assigned to complete the curriculum (compared to control classes that did not) produced more creative solutions, and were somewhat less likely to engage in thoughtless or non-critical reliance on AI output. Notably, the intervention curriculum did not improve scores on a global measure of critical thinking, although as authors note, short-term interventions rarely produce measurable improvements in critical thinking.
Weigelt, H., Segev, E., Kurtz, G., Kahana, O., & Fogel, N. R. (2026). Enhancing students’ critical thinking literacy in a generative AI context: Eye movement patterns of deepfake detection. Computers & Education, 244(June 2025). https://doi.org/10.1016/j.compedu.2025.105529
This one takes on what you could call visual critical thinking – the ability to discriminate real images from AI-generated ones. Researchers examined patterns of visual analysis and attention using eye tracking (a well-established methodology that yields precise data on how people process visual scenes). A short instructional module significantly altered how students examined images for signs of authenticity, even though there were only modest and nonsignificant gains in accuracy. This is a concerning finding, but as the authors put it: “While the rapid progression of GenAI might lead some to argue that the human capacity to distinguish authentic from synthetic media is a ‘lost cause,’ this perspective neglects the essential role of visual analysis within the broader framework of critical evaluation. The primary goal of pedagogical interventions, such as the one examined in this study, is not to create infallible human detectors but to foster this initial stage of cognitive engagement.”
Gao, F., Xia, L., & Zhong, W. (2025). Stereotypes in artificial intelligence-generated content: Impact on content choice. Journal of Experimental Psychology: Applied. https://doi.org/10.1037/xap0000548
This series of studies focuses on both the qualities of AI-generated content, and the impacts of particular patterns of content on human users. The authors verify, through a systematic survey of AI-generated content, that such content does indeed replicate stereotypes that are prevalent in the human-generated content that AI is trained on (e.g., casting women in support and service roles). Then, they show that these stereotypes nudge users to select similarly stereotypical content when using AI, e.g. to generate images for a hypothetical ad campaign. Warning users that AI-generated content may contain stereotypes had limited effectiveness for countering this tendency, suggesting that while efforts to raise awareness (as would commonly be done in AI literacy modules) might be helpful, adjusting AI content itself would be the better strategy for mitigating its tendency to accentuate social stereotypes.
Chow, J. K., McGugin, R. W., & Gauthier, I. (2025). Domain-general object recognition predicts human ability to tell real from AI-generated faces. Journal of Experimental Psychology: General, 155(3), 629–648. https://doi.org/10.1037/xge0001881
Here’s another spotlighting the unexpected aspects of visual perception and AI, uncovering some surprising individual differences in the ability to tell apart real and AI-generated human faces. (As an aside – it’s surprising as well that AI faces have become as convincingly realistic as they have, given the exquisitely tuned abilities we have to perceive faces relative to other objects. Face perception is supported by its own dedicated area within the brain and we recognize faces differently than instances of other objects.) It turns out that some people are reliably better than others at detecting AI faces, an ability that traces back to a special, possibly inborn talent for discriminating fine details in visual objects in general. This could be something useful to keep in mind in AI literacy development, along with the fact that researchers came up with a new test specifically for testing a person’s ability to do this particular task well, the AI Face Test (AIFT).
Schmidt, L., Obergassel, N., & Roelle, J. (2025). AIming high: Do goal structures matter in learning with ChatGPT? Applied Cognitive Psychology, 39 (6). https://doi.org/10.1002/acp.70148
This one brings together two lines of research that faculty and faculty developers already care a lot about: generative AI and motivation. Specifically, it examines the effect of “goal structures” (essentially, the overarching objectives of engaging in an activity) on how students engage in a ChatGPT-supported learning task. Students learned social psychology concepts with ChatGPT under either a mastery goal structure (emphasizing individual progress and understanding) or a performance goal structure (emphasizing comparison and outperforming others). The mastery-oriented framing led to better acquisition of conceptual knowledge, while the performance-oriented framing increased pressure and anxiety and prompted more requests for peripheral details more related to superficial aspects of the material. One take-home: AI is not just a tool that students either use well or poorly in isolation; the broader motivational climate around the task matters as well. This type of fine-grained consideration may become more important as AI tutoring systems become widespread-to-ubiquitous.
Pan, S. C., Schweppe, J., Teo, A. Z. J., Indrajaya, A., & Wenzel, N. (2025). Using ChatGPT-generated prequestions to improve memory and text comprehension. Journal of Applied Research in Memory and Cognition. Advance online publication. https://dx.doi.org/10.1037/mac0000254
This one connects generative AI to one of the most reliable families of memory-enhancing strategies: prequestioning, or asking learners to try answering questions before they have studied the material. Prequestioning is promising, but in practice it can be hard for students to do on their own, because they need reasonably good questions before they know the material well enough to judge the questions’ quality, and the time demands of creating adequate numbers of questions might be prohibitive for instructors to take on. Researchers tested whether ChatGPT-generated prequestions could fill this gap. Participants answered AI-generated prequestions before reading an educational text, then took memory and comprehension tests. The results were encouraging: AI-generated prequestions improved learning compared with simply reading, worked about as well as human-generated prequestions, and outperformed an AI-generated outline. Importantly, the benefits appeared not only for test questions students had effectively previewed, but also for related questions they had not seen before. This makes the study a nice example of AI being used not to replace a learning process, but to make a well-supported teaching and learning technique easier to put into practice.
Joo, S., Han, I., & Park, I. (2026). The effect of metacognitive prompts using generative AI on cognitive load, task performance, and self-efficacy in online self-regulated learning. Educational Psychology, 1–24. https://doi.org/10.1080/01443410.2026.2618717
In this case, researchers studied fully online learning, with a focus on metacognition and self-regulated learning. The researchers designed an online learning environment that integrated ChatGPT with metacognitive prompts – essentially, user interactions emphasizing the user’s problem solving process and reflection, without giving away any answers. Similar to the above study on prequestions, the general idea was to take a technique that has historically been prohibitive to implement (customized questions and suggestions aligned to an assignment) and lowering that bar using generative AI tools. Results suggest that the AI-supported prompting environment improved students’ conceptual understanding, increased self-efficacy, and reduced perceived task difficulty. One interpretation is that AI-aided interactions, if carefully aligned with self-regulated learning principles, may help learners direct their mental effort toward the productive parts of the task while making the whole activity feel more manageable.
Li, X., Li, T., Yan, L., Li, Y., Zhao, L., Raković, M., Molenaar, I., Gašević, D., & Fan, Y. (2026). FLoRA: An advanced AI-powered engine to facilitate hybrid human-AI regulated learning. Computers and Education, 243, 105527. https://doi.org/10.1016/j.compedu.2025.105527
This article also focuses on self-regulated learning (SRL), discussing approaches to AI that take self-regulated learning into account by “reloading” SRL back onto the student as learning progresses (e.g., by prodding students to make their own decisions about what to study next or to do their own progress tracking). These principles are applied in the form of an AI learning assistant (FLoRA) integrated into Moodle, which tracks learning analytics connected to student SRL and produces personalized guidance messages. Case studies, including quantitative data on the frequency of SRL-associated actions during learning, are reported based on a selection of the ~20 IHEs that have implemented FLoRA.