There’s a lazy argument circulating in AI education circles that goes something like this: deepfakes are too good now, humans can’t tell the difference, so why bother teaching detection? I find that argument intellectually convenient and pedagogically dangerous.
Weigelt et al.’s 2026 study in Computers & Education pushes back on it. Their study uses eye-tracking technology to compare gaze patterns between students who received a critical thinking intervention and those who didn’t, and the results tell a story that’s as much about how we teach AI literacy as it is about deepfake detection itself.
How Weigelt et al. Designed the Deepfake Detection Study
The setup is a between-groups experiment with 76 first-year Bachelor’s students at the Holon Institute of Technology in Israel. The experimental group (n=24) went through a critical thinking intervention before the detection task. The control group (n=44) received no training and went straight to the test. Both groups viewed 12 human portrait images (6 AI-generated, 6 authentic) on a screen while an eye-tracker recorded their gaze patterns.
The intervention itself was classroom-based and taught students to analyze images at two levels: denotative (what do you literally see?) and connotative (what interpretation does the image carry?). Students also learned current detection strategies and practiced with 18 training images. It wasn’t a quick tip sheet. It was structured pedagogical instruction designed to shift how students process visual information.
What makes this study different from most deepfake detection research is the eye-tracking methodology. Weigelt et al. tracked fixation duration and fixation distribution across defined areas of interest: eyes, nose, mouth, and background. That data gives us something self-report surveys never can: a direct window into the cognitive processing strategies students used when making their judgments.
Trained Students Don’t Just Look Longer. They Look Differently
The eye-tracking results confirmed both of the study’s first two hypotheses. The experimental group showed significantly longer total fixation durations and a broader distribution of fixations compared to the control group. Trained students spent more time examining each image, and their gaze spread across a wider range of facial and background areas.
The background region turned out to be the most revealing difference. It showed the largest significant gap between groups in 10 out of 12 images. Untrained viewers concentrate almost entirely on central facial features, the eyes, nose, and mouth, which are typically the most polished parts of a deepfake. Trained students, by contrast, expanded their visual scanning to include ears, hair, neck, and background elements where AI-generated artifacts are more likely to show up.
I’ve been tracking the conversation around critical AI literacy closely, and Roe, Furze, and Perkins (2025) made a strong case that AI literacy needs to go beyond knowing how tools work and toward actively questioning what AI produces. Weigelt et al.’s eye-tracking data gives that argument empirical worth. The intervention didn’t just tell students to be more critical. It measurably changed the cognitive processing strategies they used to evaluate visual content.
The Accuracy Improved. The Overconfidence Didn’t
The experimental group did perform significantly better on actual detection accuracy. They got more correct identifications across the 12 images, and the chi-square results confirmed the difference was statistically significant. The training made a real difference in outcomes.
But the accuracy gains were uneven. Both groups still showed low accuracy on individual images, and performance varied considerably from one image to the next. The intervention helped at the aggregate level, but it didn’t turn students into reliable deepfake detectors image by image. Weigelt et al. are candid about this: the sophistication of current GenAI face-generation tools means even trained observers struggle with highly convincing fakes.
There’s also a bias effect worth naming. The experimental group showed a significantly greater tendency to classify images as deepfakes, while the control group leaned toward labeling images as authentic, even when they were fake. The authors connect this to the well-documented AI hyperrealism phenomenon, where people default to perceiving AI-generated faces as real, sometimes rating them as more authentic than actual human photos.
The most striking finding in this paper is the metacognitive gap. Both groups dramatically overestimated their own performance. The experimental group predicted about 68% accuracy and scored 52%. The control group predicted 66% and scored 35%. The training improved actual performance but did nothing to improve students’ awareness of their own limitations.
Weigelt et al. connect this to the Dunning-Kruger effect, and I think that connection is right. I covered a similar metacognitive blind spot in Fan et al.’s (2025) work on metacognitive laziness with generative AI. The pattern is consistent across studies: AI interactions tend to inflate students’ confidence in their own abilities, even when the underlying skills haven’t caught up.
What This Means for Teaching AI Literacy
This is where the study earns its place in the AI literacy conversation. Weigelt et al. argue that dismissing deepfake detection as a “lost cause” misses the point entirely. The goal of a pedagogical intervention like theirs isn’t to produce perfect human detectors. It’s to disrupt passive acceptance, to trigger that first cognitive step of questioning whether an image is authentic before moving to verification. That shift from passive consumption to active scrutiny is itself a critical thinking outcome, regardless of whether the student gets every image right.
I agree with that framing, and it connects to what Hillman, Holmes, and Duarte (2025) found in their rapid review of AI literacy frameworks for the Royal Society: the most effective approaches to AI literacy education go beyond technical knowledge and into critical evaluation, ethical reasoning, and metacognitive awareness.
Weigelt et al.’s study adds a piece that most frameworks are missing: attention to the visual domain. Most AI literacy work focuses on text-based outputs. Very little addresses the growing flood of synthetic images, and almost none uses objective measures like eye-tracking to assess whether interventions actually change how students process visual information.
The overconfidence finding also points to a gap in how we currently teach AI literacy. If a deepfake detection intervention improves actual skills but leaves students equally overconfident in their abilities, we haven’t finished the job. Gerlich (2025) flagged a version of this same problem in the context of cognitive offloading, where students who rely on AI tools don’t recognize the degree to which the tool is doing the thinking for them. The parallel here is striking: students who get better at spotting deepfakes still don’t realize how often they’re wrong.
Weigelt et al. recommend multi-faceted approaches for future interventions: training in recognizing uncertainty, cultivating appropriate skepticism through guided practice with feedback, and building contextual analysis skills that go beyond just looking at visual features. Those recommendations are practical, and they’re grounded in what the data actually showed.
The sample is small (n=24 in the experimental group), the focus is limited to portrait images, and the lab setting doesn’t tell us how these skills transfer to social media feeds or news articles. The authors flag all of this clearly. But the contribution is real: objective evidence that a targeted critical thinking intervention can reshape how students cognitively process AI-generated content. That’s something no survey or self-report study can give us.
We won’t make students infallible at spotting deepfakes. It will make them slower to trust what they see, and that’s exactly what critical thinking looks like in the age of generative AI.
References
- 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
- Hillman, V., Holmes, W., & Duarte, T. (2025). A rapid review of AI literacy frameworks, with policy recommendations. A report prepared for the Royal Society. London: The Royal Society. https://royalsociety.org/-/media/policy/projects/ai-in-education/hillman-et-al-a-rapid-review-of-ai-literacy-frameworks.pdf
- Roe, J., Furze, L., & Perkins, M. (2025). Digital plastic: A metaphorical framework for Critical AI Literacy in the multiliteracies era. Pedagogies: An International Journal. Advance online publication. https://doi.org/10.1080/1554480X.2025.2557491
- Weigelt, H., Segev, E., Kurtz, G., Kahana, O., & Raz Fogel, N. (2026). Enhancing students’ critical thinking literacy in a generative AI context: Eye movement patterns of deepfake detection. Computers & Education, 244, 105529. https://doi.org/10.1016/j.compedu.2025.105529