Develop the measures

The model is a merge: decades of person-level research (beliefs, motivation, belonging) and environment-level research (classrooms, policies, the cultural story about who and what AI is for), brought into one system. Its claim: GenAI will not by itself make students lazy, and it will not by itself make them better learners. It enters the motivational systems schools have always shaped. The system runs from what a person brings, to how they read the situation, to the goal that follows, to the strategy chosen, to consequences that feed the next situation. Each strategy is rational given the goal, which is why the model treats a student's use as information about the situation, not a verdict on the student. A system can be mapped, measured, and changed.

What we want from students has not changed: the science of learning and development named it before AI existed — young people who can pose questions, seek relevant resources and tools, evaluate information from multiple sources and perspectives, use metacognitive skills to manage their learning, and work toward goals with agency and purpose (Darling-Hammond et al., 2020). A student who can do those things has many of the skills needed to be an effective user of GenAI. The field does not need to build this science from scratch: it can draw on decades of research on motivation, learning, identity, and the design of equitable learning environments.

The full model — the constructs it generates, and what current AI literacy frameworks leave unspecified — is the subject of a commentary in preparation; findings from its first measurement sweeps are below.

Given the breadth of possible measures the model implies, the first pass is a scoping sweep: single items and small blocks fielded across three studies — a national U.S. high-school panel (~790 students, May 2026) and two college studies (232 students, Sept 2025; the second in the field now) — to identify the promising candidates for development effort. Below I provide glimpses into some of the categories of measures we have developed and are testing. Updates will be provided as we get the data cleaned and analyzed.

Being a GenAI user is closely linked, in the public mind, to being a computer-science or tech person, so the computer-science belief system may spill over to who can become an effective user of GenAI. If it does, that shapes how we talk about effective use, how we train teachers and managers, and how we build programs for students and employees. And there are beliefs with no prior version at all: whether the capacity to resist offloading is something a person can develop, and whether teaching careful use is worth attempting.

The model's appraisal categories are the established ones — can I do this, is it worth it, is it safe, does my judgment matter — but GenAI gives each a form the existing measures don't have. Some build on known stereotypes and known inequalities in new disguise: being suspected of cheating lands differently depending on who is already presumed to cut corners.

The sweep also covers, with full item sets to be reported with the studies:

• Goals, strategies, and styles of use. Five self-rated portraits, each a first-person description a student rates as sounding like them: using AI to improve their thinking or work; switching with the situation (thinking things through sometimes, just getting the work done when stressed, behind, or tired); wanting something fast that helps them finish; choosing not to use AI; and not yet having learned how. Items about handing an entire assignment over, asking the reason: didn't feel like it, didn't see the point of the assignment, didn't believe I could do it well. And a paired vignette — two students, one outsourcing everything to an AI chatbot, one using it to find the gaps in their understanding — asking which one students believe will be more successful.
• The experiences a climate produces. Past-month event counts: accused of AI use without having used it, stress about being accused, avoiding AI for learning out of fear — and the reverse, an instructor recognizing good use. Plus open-ended episodes of real use, rated for the situation.
• The context itself. What each of a student's two courses signals about assignments, AI policy, and instructor beliefs — because the same student sits in different climates.

• How a student uses AI carries real signal: adolescents whose self-rated portrait mix leans toward thinking with the tool report more mastery behaviors, stronger mastery goals, and higher STEM aspirations; a mix leaning toward finishing fast predicts the opposite on every one.
• Students who use AI chatbots more often report more of both kinds of use at once: more using it to check their reasoning, and more handing the work over to it. So a count of how often a student uses AI moves without saying which of the two happened.
• In the college data, once we account for how a student uses AI, how often they use it adds nothing to the prediction of discerning use; and how they use it adds prediction beyond how often for AI interest and confidence.
• Confidence in the tool climbs steadily with more frequent use; the ability to catch the tool's errors does not move at all.
• Use is situational, not a stable property of the student: half of college students reported a purpose for using AI in one of their courses that they did not report in their other course — same student, same month.
• Non-use is two different things. Students who choose not to use AI look adaptive: more mastery behaviors, better grades, fewer barriers. Students who simply have not learned how show the opposite pattern: fixed beliefs, elevated barriers, worry about being suspected. An instrument that does not separate the two reports the average of a decision and a deficit.
• Computing stereotypes travel to AI as a belief system: students' fixed-ability beliefs about STEM and their fixed-ability beliefs about AI correlate about 0.5, in adolescents and college students alike.
• AI-specific beliefs earn their place next to general ones: believing that your ability with AI is something you cannot change predicts handing work over under stress, beyond what the parallel STEM belief predicts — and that replicates in the adolescent and college samples.
• The public narrative has reached students: about one in three college students sees little point in teaching students to use AI carefully, and nearly half believe that people who use chatbots heavily will end up letting the chatbot do most of the thinking, no matter how good their intentions.
• Adoption rankings mislead. When we group adolescents by their patterns of use and belief and rank the groups by apparent adoption, an agree-with-everything group comes out first and the genuinely engaged thinkers come out fourth — the two healthiest groups sit at opposite ends of the adoption scale.