How to avoid getting lazy using these tools and losing skills to AI

AI use can indeed be a slippery slope! One study that shows that is the one by Poulidis et al. (2025), who use a 12-week home online chess training provided to 216 members of chess clubs (with at least a year of training, so already demonstrably motivated and engaged, and about half of them 18 or younger). Participants were invited to participate and reminded to go practice on that platform by their coaches, and there were financial incentives in place (10$ as base incentive and bonuses up to 150$).

On the platform, students were assigned two conditions: Either they received automated tips in “critical moments” but couldn’t ask for help, or they received the same kind of tips but also could request additional help by clicking a button. And clicking the button they did! And that had consequences on learning: Performance gain in the self-regulated group were a lot lower than in the other group. In the group where students could not ask for hints but received targeted feedback at points in the process determined by the algorithm, students had to go through productive struggle which contributed to their learning, whereas in the other group, based on survey results, “students knowingly over-relied on AI assistance, thereby diminishing their sense of accomplishment. Despite recognizing these drawbacks, they not only continued to rely on it but did so increasingly over time”. The size of that effect is moderated by student motivation (but not skill!): “motivation moderates the learning losses induced by self-regulated AI use, with more motivated students experiencing substantially smaller learning losses”. Students in the self-regulated group also played 24% fewer training games than students in the other group. When asked for a preferred training type for hypothetical future training, the largest group of self-regulated students (40%) picked “no tips”!

But Poulidis et al. (2025) also compared learning gains against students who had not been part of the study and thus did not have access to the platform at all, and overall, students learned more on the platform, no matter the experimental condition they were under. So you can learn from AI! However, “[g]iving students control over when to receive assistance can substantially hinder learning; effective AI tutors should therefore target help to moments when it best supports learning rather than providing assistance on demand“. Otherwise it is easy that students fall in the “agency trap: when highly accurate solutions are easily accessible, even students who genuinely want to learn over-rely on AI assistance.”

Poulidis et al. (2025) conclude that “students recognized that overuse would harm their learning yet still relied heavily on AI assistance—awareness alone cannot prevent misuse”. Emphasis in that quote is mine because I find it to be so important: even though participants were motivated to learn and knew that by clicking the button, they were harming their own learning, they still could not resist the temptation, and 40% therefore (or at least that’s my interpretation) would wish to not even be led into temptation in hypothetical future training by completely removing the option to get feedback. Poulidis et al. (2025) close by writing “[a]s self-regulated AI use becomes increasingly ubiquitous in education and the workplace, preventing harm to long-term learning and the atrophy of human capabilities is a central design challenge”.

This is a really interesting article with super relevant results. It seems likely that what they find — it is really difficult to resist an apparent AI shortcut, even if people know it is going to backfire on them — will work similarly in situations where AI can so clearly give a correct answer that results in clearly marked “wins” (a win does not get much clearer than winning a game of chess). The reason I am posting this whole novel in response to your question, though, is that I think being aware of such studies might help you and your students to make the decision to put measures in place to remove temptation for themselves, so as not to even step on that slippery slope at all, which is also my own solution to the problem. But I am curious to read other approaches!

Reference:

Poulidis, S., Bastani, H., and Bastani, O. (October 01, 2025). Self-Regulated AI Use Hinders Long-Term Learning. The Wharton School Research Paper, Available at SSRN: https://ssrn.com/abstract=5604932 or http://dx.doi.org/10.2139/ssrn.5604932

This is a big question — or pair of questions — and the answer is correspondingly long. It's in four parts: in the first, I ask which skills; the second are some ideas for making temptation easier to resist; the third about helping your students (pedagogical aspects); and the final a thought on the systemic context.

On the skills 'at risk'

Who's losing skills? Individuals or humanity as a species? (Or, rather the subset who have access to and make use of gen AI?)

The fear of this reduction in capability, in competence, is one I share, and one of the reasons I staunchly advocate to strongly advise students to not use generative AI (at all if they can bear to, but at least never without very careful thought). I could highlight valuable skills and knowledge like being able to write (string a sentence together), knowing what code does or being able to write computer code in a novel way, being able to do maths, especially in general form, or create figures that say what one actually wants to say, but when I came to write this, I realised that there may be an even bigger skill at risk.

Perhaps the two most valuable 'skills' that generative AI erodes at the individual level are the ability to think (which Kosmyna et al.'s 2025 pre-print provides a preliminary quantification of), and one's ability to trust oneself (which I haven't seen discussed, and which I suspect may be even more pernicious): one's trust in one's ability to write text that will be understood — despite imperfections, to do calculations, to solve problems; one's trust in one's own judgement and ability to (critically) evaluate one's own work. It 'promises' improvements; reassurance of being 'on the right lines' or have 'reached the correct answer' but what it becomes is a crutch. Using generative AI runs the risk of leaving the user an intellectual cripple and coward, unable (or at least unwilling) to do anything that is not in some way 'approved' by a large language model (with all its biases).

I dare to be hopeful though because, in terms of ability to think, three years is a very short period of time in evolutionary terms (Homo sapiens has been around as a species for c. 200,000 years), but for undergraduate students three years is a big chunk of their lives (around 15% or even more for most). It may be useful to think of generative AI use as an addiction.

Avoiding addiction

Resisting temptation (or quitting, if the use has become habitual) may not be easy, and is basically a matter of self-discipline, but I have a some suggestions for strategies I would consider. Some of these (making access harder) I used when I spent far too much time (hours a day) playing minesweeper — which I now haven't played for months; those about deliberately making the experience unpleasant are based on the fact that I find generative AI easy to resist because I really dislike the experience.

• Make accessing the generative AI harder (this may also add a bit of inconvenience into other web browsing, but that's the price one pays for independence, and when the habit is broken, you can relax the browser history ones):
  • Do not have any generative AI web pages bookmarked;
  • If you use institutional access, log out and close the page after every single use (so you have to go through the rigmarole of logging in again next time);
  • Change your browser settings so it doesn't save (or at least doesn't suggest) your browsing history when you type into the search bar.
• Reduce your exposure to generative AI by adjusting browser and search engine settings so AI summaries are never shown.
• Deliberately make the experience of using a generative AI interface unpleasant or harder work, e.g., by having a standard 'start of 'conversation'' prompt (which may include prompting for 'slop').
• Return generative AI firmly to the roll of a tool: Decide exactly what you will and will not use generative AI for (and which model you will use) — stick the list on your computer screen.
• If you find yourself thinking 'I could put that through generative AI and see what it 'thinks'', or that you might use generative AI for something not on your list (previous point), do something else! (E.g., go for a walk round the building.)
• Proofread on paper (NB: paper uses a lot of resources) — or work offline (disconnect from the internet) — in the summer, this might well be outside.

But be patient with yourself: your goal is to not become reliant on generative AI — or use it very deliberately.

Helping students to become independent humans

So much for weaning oneself off generative AI (or making the slippery slope harder to reach and maybe even disgustingly gooey). How about helping your students? The above practical suggestions still apply, and there is some hope since it is students who feel entitled, and for whom going to university is seen more as a rite of passage, and the point is to form connections, including, sickeningly, 'finding a wife', who will use generative AI to cheat, while those for whom a university education is still a genuine educational opportunity that will enable them to do more than they would be able to do without it are more likely to simply boycott the tools (as R. Purser reported in 2025).

I'm going to draw up three topics that it may be useful to address with students:

1. general uncertainty and enabling students to self-evaluate;
2. normalising, and providing clarity of declaring AI usage (and never question what is declared);
3. making students aware (or reminding them) of the damage they may be wantonly doing to their abilities, including explaining what they will gain by putting in the effort.

1. Student uncertainty, unwillingness to do or write things unless they were confident it was what was expected, and their struggles evaluating their (and peers') work long predates generative AI. But now they have a means to obtain 'reassurance' that they didn't have before. I suspect generative AI may also be revealing issues of tacit knowledge that have previously been left implicit, or otherwise not deemed necessary to mention. And long-standing complaints (e.g., 'students can't read' — which is actually 'don't have reading stamina') can no longer be ignored.

The basic problem of developing judgement and the associated ability to trust one's own work is, I suspect one of (in)experience. Which means students need opportunities to develop an understanding of what 'quality' work is and practice evaluating work. In isolated form, you can spend some time highlighting, and having students discuss, in groups, examples of good and bad work, whether texts, graphs etc. as relevant. This can cover both in general (stylistic) (e.g., Zombie Nouns), and technical aspects ('join the dots' graphs do not show trends; they would be better plotted as discrete data points (the dots) or bar charts so the discreteness of the data points is preserved). To go one step further, developing (even with students), making available to students along with the assignment, and marking using descriptive rubrics (no vague 'good quality x' etc.) are likely to help. Andrade (2005) is a very useful introduction. A certain amount of restructuring of assignments and teaching will be needed, but making expectations clear, and placing students in a position to judge those expectations without hoping that the 'average' outputted by generative AI will be what's expected cannot harm. Although such clarity also enables the student determined to not learn to prompt generative AI to a 'better' output, without such clarity, the temptation to refer to, to check against the 'more knowledgeable' generative AI is higher.

2. The following youtube video (https://www.youtube.com/watch?v\=TWi1eYjkQLE, from 15-21 minutes) describes one way of really integrating generative AI use explicitly into teaching, and making its use part of the learning process. For example, when conducting a literature review, students also submit an AI evidence sheet in which they describe and document their generative AI use, from the models used and the prompts they tried, to the use they made of the output. And this is marked too. From what I understand, this is in conditions where generative AI use is compulsory, but the point is that it rewards clarify, and makes the process both visible, and this visibility part of the learning process. I'm not sure how it translates into conditions where students can decide to not use generative AI at all; but the follow up actions, e.g., performing Google or database searchers, and actually reading the articles, can still be documented. But I'm sure that any claim 'no generative AI use' should not be questioned — or at most explored, on the same basis as generative AI use in a viva (minutes 37-41). There is also what I think is an side comment distinguishing machine learning from generative AI, and the use of local models.

3. The downsides of generative AI are well known, for example as summarised by the IMPACT RISK framework. This, and other, more specific risks, e.g., to thinking ability, as Kosmyna et al.'s (2025) research (note, preprint) points to, can be brought up in discussion with students. I would suggest starting off positive (e.g., 'what can generative AI do or be used for?'), before asking them to consider whether it's all that good (e.g., 'what risks are there, both large scale, and to you as individuals from using generative AI?') and allowing students a role in setting the agenda of which specific issues are discussed (though some you will know you definitely want to bring up). You could even task students with finding out what has been said, in both research and popular venues on these topics (e.g., references here). One thing it may be worth noting is that even (one of) the first developer(s) of chatbots was far from convinced that chatbots, and their extension generative AI programmes, were any sort of a good thing — in 1976 (Tarnoff, 2023).

Challenging the system practices

A last thought: More radically, and at a systemic level, you might wish to bring up the points raised by Martha Kenney and Martha Lincoln of San Francisco State University (in Purser, 2025), and others like them, at any relevant opportunity (e.g., in meetings discussing a university's strategy on or guidelines around generative AI). Martha Kenney: "Normally when we buy a tech license, it's for software that's supposed to do something specific... but ChatGPT doesn't." Martha Lincoln: "Why would our institution buy a license for a free cheating product?" Many universities have uncritically accepted and, quite literally, bought into generative AI platforms, which effectively endorses or approves their use. The money spent on such licences (most of the platforms can be accessed for free, and even licences are no guarantee of availability) could arguably be much better used, for example, to pay salaries or ensure that staff have work spaces that allow them to work — e.g., individual offices (or at least properly secluded work spaces) for those for whom constant (potential) interruptions from nearby and passing colleagues destroy focus and all hopes of getting anything done. Or at least to ensure "coherence between their stated missions, pedagogical practices, and approaches to emerging technologies" (Taylor and LaCroix, 2026)

Repeated attention can also be drawn to the various implications for student learning (or lack thereof) highlighted above (and others).

It may also be worth pondering the following sentence from Albert Camus' 'The Rebel' (pg 183): "Work in which one can have an interest, creative work, even if it is badly paid, does not degrade life."

Sources

Andrade, H. G. (2005). Teaching With Rubrics: The Good, the Bad, and the Ugly. College Teaching, 53(1), 27–31. https://doi.org/10.3200/CTCH.53.1.27-31

Camus, A. (1971 [1951]) The Rebel. Penguin. Translated by A. Bower.

Kosmyna N., Hauptmann E., Yuan Y. T., Situ J., Liao X. H., Beresnitzky A. V., Braunstein I., & Maes P. 2025, Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv preprint arXiv:2506.08872. https://arxiv.org/abs/2506.08872

Noon, P. and Myers, T. (2026) AI in Academia: Why You Can’t Govern What You Can’t Question Youtube video https://www.youtube.com/watch?v=TWi1eYjkQLE

Purser, R. 2025, AI is Destroying the University and Learning Itself. Current Affairs. https://www.currentaffairs.org/news/ai-is-destroying-the-university-and-learning-itself ; accessed, 2026-06-20

Tarnoff, B. (2023) Weizenbaum’s nightmares: how the inventor of the first chatbot turned against AI. The Guardian, 2023-07-25. https://www.theguardian.com/technology/2023/jul/25/joseph-weizenbaum-inventor-eliza-chatbot-turned-against-artificial-intelligence-ai ; accessed 2026-06-20.

Taylor, T. B., LaCroix, T. (2026) Purpose before policy: academic integrity, generative AI, and rhetorical stance. Higher Education. https://doi.org/10.1007/s10734-026-01706-1