In a study published in Discover Artificial Intelligence, researchers Yara Mohammed, Hassan Qandil, and Brady Lund surveyed 56 students in a mechanical engineering Machine Elements course after they completed a 30-minute assignment on fatigue failure under variable loading. The study stands out because it does not simply ask students whether they like AI. Instead, it grounds its survey in Ellis's classic model of information-seeking behavior, applying a six-stage subset of that framework: starting, chaining, browsing, differentiating, monitoring, and verifying.
Each stage corresponds to a distinct activity in how people find and process information, from getting a handle on a task to verifying what they have gathered. By mapping survey items onto these stages, the researchers could compare how students reported using AI at different points in the problem-solving process rather than treating AI use as a single undifferentiated behavior. The sample itself was a convenience sample drawn from students who completed the survey after the in-class assignment, which focused on fatigue theory, crack initiation at surface discontinuities, selection of fatigue analysis methods, fluctuating stress calculations, and the modified endurance limit.
Of the 56 respondents, 47 reported at least some use of generative AI during the task, ranging from "a little" to "extensively," while nine reported none at all. Across the full sample, the mean reported extent of AI use was 2.77 on a five-point scale, with a standard deviation of 1.26. Among the reported users, Gemini was the most commonly selected platform, chosen by 61.7 percent, followed by ChatGPT at 44.7 percent, with Copilot selected by just one respondent.
The clearest pattern in the data emerged when the researchers examined agreement levels across the Ellis-linked items. Agreement was highest for the stages the authors map to monitoring and browsing. Among the 47 reported AI users, 83 percent agreed or strongly agreed that AI made the problems easier to solve, and 74.5 percent agreed that AI helped them understand the concepts involved.
Just over half, 53.2 percent, said they used AI to explore design solutions or approaches, and 51.1 percent said they used it to check or validate their calculations or concepts. At the bottom of the list, only 46.8 percent agreed that they used AI to understand the problem requirements in the first place. Students in this sample perceived AI as most valuable for ongoing support while working through a task and for obtaining explanations of unfamiliar concepts, and least valuable for the initial interpretation of what a problem is actually asking and for the final validation of technical work.
The perception items beyond the Ellis framework reinforced the generally positive but measured tone. Among reported users, 72.3 percent agreed or strongly agreed that using AI helped them learn the material more deeply, and 63.8 percent agreed that AI improved the quality or accuracy of their final submission. Meanwhile, 59.6 percent agreed that they applied their own judgment rather than copying the AI's answer.
A descriptive comparison between light users, who reported using AI "a little" or "somewhat," and heavy users, who reported using it "often" or "extensively," added a cautionary dimension. Heavy users more frequently agreed that AI made the problems easier to solve, at 100 percent versus 72 percent, improved their submission quality, at 83 percent versus 52 percent, and deepened their learning, at 89 percent versus 62 percent. They were also far more likely to say they would have struggled without AI, at 78 percent versus 31 percent.
Most tellingly, half of the heavy users agreed that they relied on AI instead of thinking through the problem themselves, compared with none of the light users, and only 33 percent of heavy users reported applying their own judgment rather than copying AI, versus 76 percent of light users. Because the groups were small and unequal in size, the authors emphasize these are descriptive patterns, not statistically significant or causal effects. The open-ended responses gave the numbers a human texture.
Students repeatedly described AI as a way to get faster access to information, with one respondent noting that AI "gave quick responses to questions" without requiring a search through multiple websites. Others praised concept clarification, with one student saying Google's AI overview "provided me with definitions to terms I wasn't immediately familiar with," and formula recall, with another reporting that AI helped them "recall equations that were buried in the slides." Perhaps most encouraging for educators, several students framed AI as a checker rather than an answer machine: "I didn't ask AI for the correct answer; instead, I asked what steps I was missing," one wrote, while another said, "AI helps me go in the right direction when I'm stuck." The authors connect these clarification-oriented uses to the idea of "theoretic stories," structured explanations that help students connect concepts, though they note the study measured perceptions rather than the actual quality of the explanations students received. The researchers are careful about what their data can and cannot show.
The findings come from a single course, a single fatigue-failure assignment, and a convenience sample of 56 students, so they may not generalize to other engineering subjects, institutions, or student populations. Everything rests on self-reports, which may not match students' actual behavior, especially given that some students counted indirect exposure through Google's AI-generated overview as AI use. Perceived benefits such as deeper learning and improved submission quality were not backed by direct measures of performance, assignment grades, or AI interaction logs.
The internal consistency of the survey's composite categories was also modest, with Cronbach's alpha of 0.61 for AI usage items and 0.54 for perceived usefulness items, and the qualitative coding was performed by a single author without intercoder agreement checks. No inferential statistics were conducted, so the study is explicitly exploratory. Its central message is nuanced: in this sample, students did not describe AI as a replacement for thinking but as an accelerant for it, a tool for clarifying concepts, retrieving formulas, and confirming they were on the right track, while remaining comparatively hesitant to delegate problem interpretation and technical verification.
At the same time, the divide between light and heavy users hints that the line between support and dependence may depend on intensity of use. The authors call for future work that pairs survey responses with assignment submissions, AI interaction logs, and interviews across larger samples and multiple institutions, so that educators can eventually distinguish what students believe about their AI use from what they actually do, and whether the perceptions of deeper learning hold up against real measures of engineering competence.
Subject of Research: Engineering students' perceptions of generative AI use in learning, trust, judgment, and information seeking
Article Title: Engineering student perceptions of generative AI use in learning trust judgment and information seeking
Article References: Mohammed, Y., Qandil, H., & Lund, B. (2026). Engineering student perceptions of generative AI use in learning trust judgment and information seeking. Discover Artificial Intelligence, 6(1), Article 1004. https://doi.org/10.1007/s44163-026-02086-2
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02086-2
Keywords: generative AI, engineering education, information seeking, Ellis model, student perceptions, mechanical engineering, trust and judgment, survey research, ChatGPT, Gemini, academic integrity, overreliance
News Source: Denise Maddox. (October 28, 2026). Engineering Students See Generative AI as a Learning Aid, but Trust Their Own Judgment First. Scienmag.
