‘If you like, I can….?’ Why GenAI needs to come with a health warning

One notable thing about GenAI tools is that whatever question you pose, they rarely, if ever, give you an answer and leave you to your own devices. At the bottom of each output is a query. “Would you like me to…?”, “If helpful, I can now…?”,  “If you like, I can…” and, more recently, “If you like, I can outline…it’s actually really interesting…” Unless the prompter closes the exchange, such tools actively try to sustain the conversation. If encouraged, a never-ending cycle of suggestions or questions will follow.

This dynamic affects higher education because the use of GenAI tools is ubiquitous, with 88 per cent of students now using ChatGPT, Copilot, Gemini and the like to aid them with their assessments, and 92 per cent using AI in some way for their studies, according to a 2025 study from the UK’s Higher Education Policy Institute (Hepi). 

In many respects, the continuance of the conversation (irrespective of its validity or truthfulness) is a good thing. Education is at its heart dialogic (think Socratic dialogue, mentoring and notions of dialogue as central to liberatory education, as expressed by Brazilian educator Paulo Freire and US philosopher Martha Nussbaum). Exchanges open new avenues of enquiry and intellectual stimulation. More pragmatically, if the AI’s question is: “Would you like to convert this text to a table?”, it can save time (51 per cent of students in the Hepi study suggested that time was a key motivation).

• Show students what thoughtful engagement with GenAI looks like
• GenAI can join the dots, so teach students to draw new lines in empty space
• ‘Your bot of choice is not a filing cabinet’

So far so good, but what if the tool’s apparent helpfulness or attempts to pique further interest obscure a more sinister agenda or consequence, namely the inculcation of a cycle of cognitive and epistemic dependency or even addiction?

Much has been made about the extent to which the algorithms behind social media are intentionally designed to promote compulsive engagement. In March, Meta and YouTube were ordered to pay damages of $6 million (£4.5 million) total for deliberately designing addictive products. Social media differs from GenAI in key ways, of course, as dependency on the latter is more likely to resemble cognitive outsourcing, with the result being increased cognitive productivity and efficiency rather than behavioural addiction and social validation. And there is a fine line between dependency and addiction. 

However, like their social media counterparts, GenAI algorithms are designed to shape behaviour through persuasion, attracting curiosity and feedback loops. Through these engagement-maximising features, this can create interaction patterns that ultimately supplant self-efficacy. In terms of learning, this seems to be in stark opposition to, or even “incompatible” with, as Geoffrey M. Cox from Stanford University has written, our supposed ambition of creating independent, self-regulating, future-fit learners. 

An increasing body of convincing and rigorous research already shows how GenAI tools create epistemic deskilling and erode critical thinking skills. As Ohio State University literature professor Michael Clune said in the Guardian, many students have been left “incapable of reading and analyzing, synthesizing data, all kinds of skills”, and as WonkHE’s Trained to Stop Learning? report highlights, 47 per cent of students worry that their grades don’t fully reflect what they know, and 38 per cent admit they sometimes submit work they can’t fully explain because they used GenAI tools. Even more tellingly, 65 per cent of students questioned for the 2026 Brookings report A New Direction for Students in an AI World: Prosper, Prepare, Protect, worried that AI would undermine their cognitive development and 14 per cent were concerned that it would increase dependence. As one student in the WonkHE report put it: “I’m afraid of it changing the way my brain works.”

Yet institutional policy often seems to be wanting to accelerate what Clune, in a recent essay, has described as an eagerness to “self-lobotomize”, and often in stark opposition to what both students and indeed academics at the coalface want.

In the wake of a US jury deciding that social media and its algorithms are indeed addictive, we might wonder if, in the future, institutions will rue the day they rushed headlong to embed GenAI pedagogies as a form of “co-intelligence”. Critical GenAI literacies are obviously essential, but not without careful pedagogy. 

So, what is to be done?

It is often difficult to speak truth to power, so the immediate challenge is how to make students recognise the dependency-forming dynamics of GenAI for themselves. Simply warning them about cognitive offloading or algorithmic capture is unlikely to change behaviour. Instead, we need to design activities that expose these mechanisms in practice – for example, equipping students with the critical and analytical skills and awareness to recognise the “If you like, I can…” prompts as behavioural nudges. Such approaches make visible the feedback loops through which these systems sustain engagement and encourage reliance. By building small points of friction into learning – moments where students must pause, justify continued AI use or switch to unaided thinking – we can begin to counteract automatic dependence. 

Understanding how GenAI cultivates habitual use should therefore be treated as a core academic literacy. Ethically, this is becoming a necessity. Pedagogically, it may be one of the few effective responses to systems designed to keep students engaged.

Adrian J. Wallbank is an associate professor of academic development at Oxford Brookes University.

If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.