É«ºÐÖ±²¥

When I ask my marketing students how they check artificial intelligence output before using it, the answers are always about accuracy. They verify citations, test claims, watch for hallucinations. Not one has ever mentioned checking what the answer assumes. Yet an AI response can be factually impeccable and still carry assumptions about who the default professional is, whose circumstances the advice fits, and which options are never worth mentioning. Accuracy checking will not surface any of that, because we have taught students to ask whether AI output is true, when the more important question is whether it is neutral.

The difficulty is that these assumptions hide behind the very qualities that make AI advice useful. My ongoing research into how people take advice from large language models (LLMs) suggests what I call the empowerment paradox: AI advice may make people more efficient decision-makers while also increasing reliance on the system and reducing their sense of decision agency. Because the answers arrive fluent, confident and instant, users stop interrogating them, and that efficiency masks two losses at once. Students stop noticing that they have handed over the deciding, and they stop noticing what the answer assumes.

What the answers assume is well documented. A Unesco of the models behind popular AI platforms found that they consistently associated women with home, family and children, and men with business, careers and salaries, with one model describing women in domestic roles four times as often as men. In June, UN Women that about 20 per cent of gendered sentence completions displayed openly sexist attitudes. A student who leans on these tools for careers guidance, essay framing or decision support is absorbing more than answers. They are absorbing a worldview, delivered too smoothly to attract notice.

• Your students use GenAI. Let’s redesign learning so thinking still happens
• The best way to teach students to think critically about AI? Make them argue with it
• Spotlight guide: Boosting AI literacy across your institution

So, the task for those of us in the classroom is not to police AI use or to perfect detection. It is to teach students to see what fluency conceals: the assumptions in the answer, and their own drift from deciding to deferring. Four practices do this reliably, and none requires new technology, budget or institutional sign-off.

Audit AI outputs for bias

The first practice is a bias audit run in class. Give students an identical prompt and vary only the person in it: ask an LLM for career advice for Emma, then for James; request a reference letter for a female candidate, then a male one with the same accomplishments; ask for financial guidance for a single mother, then for a young professional man. Have students work in pairs, compare outputs line by line, and list every difference in tone, ambition, assumed constraints and options offered. 

What they uncover is a pattern researchers have already measured: an experiment reported in PNAS Nexus leading models assigning systematically different scores to candidates according to gender and racial identities signalled by names, even when other résumé characteristics were controlled, and a study in The Lancet Digital Health GPT-4 liable to perpetuate gender and racial biases in clinical recommendations. 

The exercise takes 20 minutes, and it teaches more about algorithmic bias than any lecture because the students generate the evidence themselves. The debrief question matters most: if you had only seen your own version, would you have noticed anything wrong?

Introduce interrogation prompts

Give students interrogation prompts, not just verification prompts. They already know to ask whether AI output is true. Add a short question bank they apply to any AI advice before acting on it. Who does this answer assume I am? What would change if I were older, poorer, a woman, disabled or writing from another country? What options did it not offer, and why might that be? Whose interests does this recommendation serve? These questions convert bias from an abstract topic into a repeatable habit of mind. Print them on a slide, put them in the module handbook, and require them in any assessment where AI use is permitted.

Signpost the shift away from decision-making

The third practice makes the drift from deciding to deferring visible for students. My research suggests these disempowering effects operate through dependence and frustrated autonomy, and both are largely invisible to the user in the moment. Consumer researchers the same mechanism as agency transference, the gradual handing over of agency from humans to algorithms, which may contribute to de-skilling over time. Colleagues writing on this platform have described the endpoint as cognitive surrender, where users defer entirely to generative AI without engaging their own judgement. 

So, catch the drift early. Ask students to keep a short decision log for one assignment: each time they consulted an AI tool, did they defer to it, adapt it or override it, and why? A paragraph of reflection on the pattern can be built into the assessment itself. Students are routinely startled by their own logs. The point is not to shame AI use but to restore the noticing, because a student who can see when they have stopped deciding is a student who can start again.

Ask AI to argue against itself

The final exercise is to run a counter-answer experiment. Give the class an AI recommendation on a topic from your module, then have students prompt the model to argue against its own advice and evaluate the two positions, asking whose interests each version serves and whose each leaves out. It works as a seminar debate or in pairs, requires no change to assessment and no institutional decision about AI policy, and it lands exactly what the first three practices build towards: treating AI output as one confident voice in a debate rather than a verdict. Where AI use is permitted in assessment, the same design can also be graded, because the assessed work becomes the student’s judgement of the two positions, which cannot be outsourced.

None of these practices requires taking a position on whether AI belongs in higher education. Students are already taking the advice. The only question is if they can see what it assumes, and their ability to tell when they have stopped deciding. Teaching students to interrogate a fluent, instantly available answer may be the most transferable critical thinking skill we can offer them, because they will be practising it for the rest of their working lives. Efficiency that costs agency is not empowerment, and efficiency that distributes its costs unevenly is not progress. The classroom is still the best place students will ever have to learn the difference.

Kamila Miller is a lecturer at Henley Business School at the University of Reading. Her research examines how advice from large language models affects consumer decision efficiency and agency. She is the author of Data-Driven Marketing Strategy (Kogan Page, 2026).

If you would like advice and insight from academics and university staff delivered direct to your inbox each week, .