Catching up on some GenAI reading

I know the idea that I can catch up on reading is flawed, to say it nicely, but still I try from time to time… Here are some quick summaries of GenAI articles that lived in open tabs in my browser because they seemed too relevant to not read.

AI knowledge alone does not prevent over-reliance on AI

Wu et al. (2026) investigate “AI anxiety and AI dependence among undergraduates: a moderated mediation model of AI self-efficacy and AI literacy” in 400 full-time undergrad students in China. They find that AI anxiety and AI dependence are positively related, with AI dependence being “a maladaptive coping response” to AI anxiety. They explain that with the “core idea in CBT that affective distress can shift individuals toward short term coping patterns, which reduce perceived uncertainty and effort, especially when tasks are cognitively demanding and outcomes matter for self-evaluation“. The more anxious students are, the less confident they are in their own capabilities, the more they rely on AI to the point of dependency.

Generally, “[s]tudents who believe they can use AI competently are more likely to keep themselves cognitively engaged, treat AI as assistance rather than substitution, and maintain verification habits that prevent blind reliance“. However, the more AI literate the students, the stronger the effect of anxiety eroding their confidence, which means that AI literacy is not a protective buffer against AI dependence, but “high literacy may intensify the erosion of self-confidence during anxious states“.

The authors conclude “The results suggest that undergraduates’ AI dependence is not only a matter of convenience or habit. It is partly embedded in anxiety-driven coping and efficacy related self-regulation, which are increasingly salient as generative AI becomes a routine component of higher education.” This means that students don’t only need technical training, but technical training combined with support of emotional regulation; for example fostering self-efficacy (through skills such as prompting, verifying, error detection, …) and resilience to protect self-efficacy under stress, as well as skills to not even become super stressed in the first place, like time management. Students also need to know what exactly they are supposed to do and not do, and be encouraged to experiment with AI in environments where AI use isn’t judged as misconduct or incompetence.

Students usually don’t just read AI summaries, but integrate AI tools into reading practices to overcome barriers

One prejudice we often encounter is that students don’t read original texts any more but rely on GenAI summaries instead. Corbin et al. (2026), in “Hybrid reading practices: how GenAI summarisers can support rather than replace reading“, find that it is not as simple as that. They investigate how 24 philosophy students really work with GenAI when reading, and while some use it as shortcut (and all are aware that that is not a replacement for actual reading), most do not. Instead, they “used GenAI to overcome barriers that would otherwise prevent them from meaningfully engaging with readings: for example, to overcome material constraints such as overwhelming reading loads, time pressures from work and caring responsibilities, and epistemic challenges such as dense disciplinary language, unfamiliar theoretical frameworks, and historical prose.” In addition to translating and clarification, one interesting strategy is to use GenAI for orientation, or pre-reading framing; reading a GenAI summary to get an overview before engaging more closely with a text. Another one is to cross-check understanding, to ask for alternative perspectives or for how one might interpret something differently than the student’s interpretation. Overall, they really create “hybrid reading practices” where they move between different representations of the same material: The original text and GenAI outputs.

When learning to write, use GenAI to draft

That’s the counterintuitive result that Lai et al. (2026) find in “What task to offload to GenAI in the writing feedback process?–Effects of task-offloading approaches on EFL learners’ writing skill development“. They study 101 university English as a foreign language students in China over 6 weeks under three conditions:

1. No AI: The learner drafts, evaluates, and revises — this is the control condition
2. Learner drafts, GenAI evaluates, and learner revises — students in this condition learn more than in the control
3. GenAI drafts, but learner evaluates and revises — students in this condition learn the most

The authors explain this by which tasks are offloaded to GenAI and which tasks can thereby receive more attention: When students do not need to focus on grammar and vocabulary, they can focus on writing itself. The authors recommend also that writing tasks should be personally meaningful so that GenAI is used for learning rather than as a crutch, to set group tasks to engage with GenAI feedback, and to train GenAI skills.

“Resistance to GenAI reflects pedagogical care rather than technological reluctance”

In “Cognitive offloading, critical pedagogy, and connectivist learning shape academic responses to generative AI in higher education: transforming resistance into pedagogical care“, Crimmins et al. (2026) report on interviews with 30 Australian university teachers about GenAI adoption, which, they find, “was shaped by peer norms and institutional signals but consistently mediated by ethical reasoning and commitments to student-centred learning“.

In the background they give, they describe the “Google effect” (if you can easily find something online, you aren’t likely to memorise the information) and the parasitic relationships with smartphones, in which we live in “near-symbiotic” relationships that might initially be beneficial but over time might not be. And concerns are the same for GenAI: If not embedded into education in a sensible way, it might become a crutch that ultimately hinders learning. In addition to the cognitive concerns, there is then of course also the problem that GenAI “can reproduce and amplify existing power relations, often privileging dominant cultural narratives while marginalising others“, as well as reproducing sexist, ableist, and other equally problematic content. Differential access to GenAI as well as mentorship and knowledge in the network of how to work with it can also deepen the existing digital divide. But they also highlight opportunities for education when GenAI is implemented well. But what do teachers think?

The 30 university teachers interviewed by Crimmins et al. (2026) describe that learning to work with GenAI requires time, effort, and institutional support. They have more ethical concerns about students using GenAI in assessment than them themselves using it in teaching, and struggle with how to talk to students about GenAI. They are also concerned about uncritical adoption of GenAI and its effects on an emerging “class divide” between students who work full-time and those that “can play with AI prompts for fun“, and about the inherent biases. Some also raised concerns about GenAI’s energy consumption. Overall, “Participants were rarely opposed to GenAI in principle. Rather, their hesitation was conditional and context-sensitive, shaped by concerns about cognition, equity, and epistemic integrity“. Thus, “resistance functions as pedagogical care” and “is not a barrier to innovation but a resource for designing more educative forms of technological integration“.

This is a nice article and I like the positive framing!

It’s a valid concern: GenAI disclosure might lead to lower grades

Grading is a subjective value judgement, but there are three big trends in higher education: grade inflation, ungrading, and criteria-based assessment as a quality assurance measure. And now, there is GenAI. Luo & Dawson (2026) ask an important question: “Exploring value judgements in grading: will teachers mark down student work assisted by GenAI, and should they?“. For this, they did scenario-based interviews with 33 teachers in China. The three controversial scenarios ended in a multiple choice question on how a teacher would react (for example to student work suddenly being much better under disclosed heavy AI use) and teachers explaining why they chose their response.

Luo & Dawson (2026) find four clusters of value judgements that are not just about the submitted work, but about who the student is (e.g. trying to cut corners), what they can do (e.g. independently vs only with help of GenAI), their relationship (e.g. do they know and trust a student), and “whether the grading decision leads to good outcomes” (e.g. fairness towards other students, and whether reporting does more harm than good).

The discussion is super interesting, because how all these value judgements play out depend on a lot of other factors. For example, teachers “who know knowledge as fixed and objective” tend to think that the same output is worth more when created without GenAI, whereas teachers further along in their development might see a more nuanced picture. But also disciplinary traditions are hugely important, for example in the humanities developing an own voice is very important, whereas in STEM I have also heard from many teachers that they’d rather read a good GenAI generated text than a bad one that students wrote themselves… And then there are also lots of implicit, unconscious, unreflected value judgements related to what counts as authorship, or even inspiration.

Unsurprisingly coming from the authors of “validity matters more than cheating“, Luo & Dawson (2026) recommend focusing on diagnosing learning more accurately, being transparent of what GenAI use is permitted, but also, importantly, to reflect on values, value judgements, and how they might creep into grading practices undetected and unchecked. This was a really interesting read!

Again, it’s a valid concern: GenAI disclosure might lead to lower grades

Tan et al. (2025) investigate “To disclose or not to disclose: Exploring the risk of being transparent about GenAI use in second language writing” and find that “teachers tend to score the essays lower when they are informed that GenAI tools are involved in the writing process“. For the online survey part of the study, 89 teachers saw two student essays (written without GenAI use) but under two conditions each: One with the information that the essays were written without access to GenAI, and one where they were told that students used GenAI to develop ideas and revise sentences. Seven of the teachers agreed to a later interview. And it turns out that perceived GenAI use is penalised (and especially in an essay in an argumentative genre rather than the narrative one). This is mostly due to assumptions that the teachers make about the student’s writing ability and how the process and end product were influenced by GenAI use. Errors and unconventional choices were interpreted as genuine effort, even when GenAI use was disclosed, but I guess that will only work until teachers realise that those can also be generated by GenAI…

And again, it’s a valid concern: GenAI disclosure might lead to lower grades

In “The Hidden Cost of Disclosure: A Multi-institutional Study on Undergraduate Students’ Generative AI Usage and Faculty Accusations“, Vetter et al. (2025) investigate trust and distrust relationships between teachers and students, and how they interact with course policies, frequency, kind, and disclosure of GenAI use, and being wrongly accused or being caught. They surveyed 560 undergraduate student participants, including 110 non-GenAI users and find that students are worried about unclear expectations and policies and therefore are strategic with how transparent they are about their GenAI use for fear of punishment. Especially users who use GenAI frequently do tend to not disclose it, overall 66% of students never or rarely disclose (and the more vague or punitive an AI policy is, the more students tend to not disclose), with “fear of academic penalty” being the main reason for that decision. And they are right: “students who “always” disclosed their AI use were over three times more likely to be accused of improper use“!! But “clear policy is not enough as it relates to the somewhat paradoxical relationship between student disclosure and faculty accusation. Just as students need guidance on when and how to use or not use GenAI in a given course context, faculty need to understand that disclosure by students should never be followed by accusations“, which would only damage the trust relationship. So the authors’ advice is that “Faculty need to create AI policies that clearly articulate how, why, and for what students can use AI policies. Given students’ general GenAI usage, faculty risk disconnecting their pedagogy and content from many students’ workflows. Faculty need AI policies to ensure students can connect their learning processes with their courses. Otherwise, faculty further risk students not disclosing their AI usage or worse, deliberately trying to obscure it.” They also advise that “Faculty training should emphasize equipping instructors with the skills in engaging in trust-building conversations with students rather than relying solely on AI detection tools and policy enforcement.” Additionally, they recommend to co-create policies, always good advice!

And now I need a little break from papers about GenAI!

Corbin, T., Bearman, M., Fawns, T., Flenady, G., Formosa, P., Liang, Y., … & Walton, J. (2026). Hybrid reading practices: how GenAI summarisers can support rather than replace reading. Studies in Higher Education, 1-16.

Crimmins, G., Sanders O’Leary, E., & Strempel, E. (2026). Cognitive offloading, critical pedagogy, and connectivist learning shape academic responses to generative AI in higher education: transforming resistance into pedagogical care. Higher Education Research & Development, 1-16.

Lai, C., Pan, M., Guo, K., & Cui, Y. (2026). What task to offload to GenAI in the writing feedback process?–Effects of task-offloading approaches on EFL learners’ writing skill development. Computers & Education, 252, 105675.

Luo, J. (Jess), & Dawson, P. (2026). Exploring value judgements in grading: will teachers mark down student work assisted by GenAI, and should they? Studies in Higher Education, 51(9), 1970–1984. https://doi.org/10.1080/03075079.2025.2552825

Tan, X., Wang, C., & Xu, W. (2025). To disclose or not to disclose: Exploring the risk of being transparent about GenAI use in second language writing. Applied Linguistics, amae092.

Vetter, M., Farag, I. M., Jiang, J., Lucia, B., & Silvestro, J. (2025). The Hidden Cost of Disclosure: A Multi-institutional Study on Undergraduate Students’ Generative AI Usage and Faculty Accusations. Available at SSRN 5755762.

Wu, H., Ni, H., He, J., & Wee, E. H. (2026). AI anxiety and AI dependence among undergraduates: a moderated mediation model of AI self-efficacy and AI literacy. Frontiers in Psychology, 17, 1884382.

Walking to a dip a couple of weeks ago

Always pretty here!

Love morning dips!

Pink sky, pink water

Even just looking at the pictures has a calming effect on me!

And the sun is much higher up on the way home!