I don’t think ‘cognitive outsourcing’ is a good concept. It often shuts down more than it opens up but it remains indispensable because it’s so widely understood. It also names a real risk, summarised here by Hadeel Naeem in this paper on pg 269:
While education aims to teach students intellectual skills, relying on generative AI is deskilling them (Sternberg 2024; Krook 2025; Ahmad et al. 2023; Zhai et al. 2024; Shukla et al. 2025; Lee et al. 2025; Laak et al. 2024; Cassinadri 2024; Kasneci et al. 2023). Students outsource essential cognitive tasks such as writing, brainstorming, critical thinking, or problem-solving and therefore miss opportunities to develop their cognitive abilities and acquire new skills.
To the extent someone uses AI for core cognitive tasks, it deprives them of the opportunity to practice existing skills or develop new ones. This is a problem for universities if student use of AI means they are not developing the capabilities which the degree implicitly or explicitly claims for them. The promissory structure of the credential breaks down to the extent this growing mismatch is identified (rightly or wrongly) in ways that could in principle lead to a collapse of trust in credentials that would be existential to the current financial model of the university. These are resource intensive institutions which undertake a time consuming mode of preparation because it’s claimed this is a necessary condition for developing graduates with certain faculties. If we are seen to be failing in developing these faculties and/or technology is understood to provide a much quicker route towards developing them then what is the university for?
The problem is that cognitive outsourcing is a verdict. To describe something as cognitive outsourcing assumes that the use of the technology necessarily forecloses the development of cognitive skills. This treats user-model interaction as a black box in which the use of the technology inevitably leads to certain kinds of outcomes. Once we question this assumption we are left with the more unsettling possibility that some kinds of user-model interaction involve cognitive outsourcing and others do not. This then presents us with the challenge of distinguishing between the two and supporting our students in using technology in a manner which precludes this outsource. I talk about ‘technology’ here because this is not a problem limited to AI. As Naeem puts it on pg 271:
Pritchard (2014, 2016) discusses whether technology is in tension with the goals of education or is a fundamental scaffold for learning. Pritchard (2014) elaborates how technology can scaffold an existing ability in two ways: the agent can go on to exhibit the ability even when the scaffold has been removed, and at other times, the scaffolded technology becomes an essential component of the skill itself. Abacus, for instance, is a scaffold that helps teach an ability, but once the abacus is gone, the student retains the ability. We teach computer skills to students and require a computer to teach them, but you can’t take away the computer. The computer is an essential part of the skill.
This complicates our understanding of ‘cognitive outsourcing’. If a student presents an AI tool with a proposition and then asks it to critically interrogate them, they are practicing skills in a way that can be sustained without the presence of the technology. They get better at reflecting on challenges, formulating responses and adapting their position in light of criticism. In contrast, if a student asks an AI tool to structure an essay for them in response to an essay question, they are failing to practice the skill themselves. It’s plausibly something they will never learn, with ramifications for a whole range of interlocking ideational competencies, if they rely on the AI tool for this throughout their degree. In both cases the student is thinking-with AI but only in the latter case are they doing it in a way which leaves them reliant on the tool for their future performance of the cognitive task.
In this sense it seems clear to me that AI tools can contribute to the development of intellectual skills. Indeed they might be able to do so in a way which mass higher education generally cannot due to operational constraints of the delivery model. The tool can enable more personalised, extensive and flexible forms of interaction which are appropriate to the development of the skill. Naeem writes that with “appropriate instruction and guided practice, students can learn not merely to ask questions, but to ask better ones: more focused, more insightful, and more conducive to eliciting information, knowledge, and understanding” (pg 276).
The hill I will die on is that many of our students are already doing this at least some of the time. The creative uses of AI tools for development of intellectual skills is exactly what gets rendered opaque by an overly expansive concept of cognitive outsourcing. To ensure that more of them are doing this in reliable and effective ways, it’s necessary to provide “appropriate instruction and guided practice”. We should be actively supporting our students in using the technology in this way, rather than just leaving them to it.
The problem is that cognitive outsourcing remains a real risk. Many of our students are not doing this because they are outsourcing tasks to AI tools before they have learned to do them and/or are using AI tools in ways which develop competencies that leave them reliant on the tool. If you can’t do something without the tool then have you really learned it? This isn’t intrinsically a bad thing. Contra Plato I think the affordances of literacy were worth dispensing with our capacity to recite epic poetry from memory. But the problem is that we cannot assume continued, reliable, equitable access to these models. Even if we could, I think it would still be a problem, just less of one. But we might be setting our students up for a lifetime of cognitive dependence upon unreliable commercial services which we cannot assume they will have access to in the future. That is the nuance which I think gets lost in an overly expansive and often moralised conception of cognitive outsourcing.
Rather than think in terms of ‘cognitive outsourcing’ I suggest we think in terms of ‘cognitive dependence’ which is acceptable vs unacceptable at the level of both infrastructure and development. We accept cognitive dependence on computers, phones, books and writing. There is stable, more or less civic infrastructure underpinning to these things: there’s not a single point of dependence for any of them because no one firms own computing or the alphabet. In contrast we need to avoid teaching our students to be dependent on unreliable, commercial, possibly-withdrawn services. At the level of development we accept cognitive dependence on these things because they open up as many developmental pathways as they shutdown, even if debates over ‘screentime’ show how deeply contested these issues still are. In terms of AI tools we should accept cognitive dependence when it opens up developmental pathways, developing capabilities more deeply and/or with a broader scope, rather than shuts them down. This is what I think the ‘cognitive outsourcing’ concept, when used judiciously (which it often isn’t), is really tracking and we need to be clearer about what is at stake here.
This helps answer the challenge which Duncan Pritchard is raising in this paper. From pg 2:
For doesn’t this promote a reliance on technology at the expense of the development of the student’s own cognitive abilities and intellectual virtues (such as the intellectual virtue of intellectual autonomy)? Indeed, it seems that as each year goes by, our dependence on technology in education grows, and thereby the need to develop a student’s innate cognitive abilities subsides accordingly. Does it follow that according to the virtue-theoretic view of the epistemology of education, educators should eschew such a reliance on technology, and thereby attempt to mitigate the influence of the modern world on educational practices?
To integrate AI tools into higher education, we need to be able to account for this problem of dependence in ways that are consistent with other technologies that have long been integrated into university education. Every single one of my students has had a laptop or tablet in the room for every class that I have ever taught in my current job. They are dependent on computers but that dependence doesn’t trouble me. If dependence was intrinsically problematic we would presumably want to expunge technology from education. Where would it stop? Removal of pen and paper so that we all engage in a never ending viva? However unless we have a sense of better-or-wose (more or less acceptable) forms of technological dependence we can’t answer his challenge on pg 7:
If one thought that the acquisition of knowledge is all that really matters, then if that knowledge is more easily available via technological means than by using one’s own on-board cognitive resources⎯if it’s easier to solve arithmetical puzzles with a calculator, rather than working them out oneself, for example⎯then one should simply switch from developing the on-board cognitive resources to equipping students with the know-how to employ technology.
In his language we need to develop understanding as well as knowledge in our students. It’s not just the capacity to do something, it’s the capacity to understand what you are doing and to care about the knowledge invested in that capacity. In this sense technological dependence is developmentally acceptable to the extent it is combined with the development of understanding and a care for knowledge. It is problematic when it does not involve that understanding or a care for knowledge. I struggled with statistics as a PGT student and scraped a pass by teaching myself to use SPSS from online tutorials. That’s a problematic dependence. I began to love writing as a PGR student and expressed that passion through a regular blogging habit which developed my fluency, voice and love of ideas. That’s an unproblematic dependence. If we see the problem as located in the technology itself, we lose the nuance of what is actually at stake developmentally about ‘cognitive outsourcing’. It’s not the role of technology in cognition that’s the problem, it’s the consequences of that role: delineating between better or worse consequences demands that we be much more precise about the relationship itself in a way that blanket statements about ‘cognitive outsourcing’ more often than not preclude. It highlights the necessity that students, as Pritchard puts it, take “cognitive ownership” (pg 13) of the task at hand. We urgently need to learn how to support them in that.
I discovered from this paper that there’s actually a notion of ‘cognitive ownership’ in the assessment literature, even though i think it’s distinct from Pritchard’s view above. Wheres Pritchard offers a developmental concept, this is about the threshold in an activity itself: who is doing the thinking? From pg 3:
Cognitive ownership is defined as the point at which substantive thinking is required by the learner, to address an underlying construct in an assessment, ownership of this leaning is compromised when core cognitive steps are outsourced to GenAI. Metacognitive laziness (Fan et al. 2025) is where there is reduced planning,monitoring and evaluation when AI outputs are readily available. The tensions, of short-term performance gains versus long-term capability development, and scaffolding versus substitution, leave unresolved gaps for assessment design.
I don’t think ‘substantive thinking’ is a useful concept but the scaffolding/substitution dichotomy is because it suggests two modes through which we can conceive of a student relating to the technology in the task. If the AI tool is used as scaffolding it creates the conditions in which understanding and care for knowledge can be developed. If the AI tool is used for substitution those tools aren’t in place. So in effect we can think of three interlocking questions:
- Is the dependence a matter of scaffolding or substitution? [teaching and learning activities]
- What capabilities are rendered possible or foreclosed by this configuration? [developmental implications]
- What infrastructure can students access and how does it constrain or enable (a) and (b)? [infrastructural implications]
We could then think of cognitive ownership at a task-level and cognitive ownership as a longitudinal feature of the student’s learning journey. The problem at the degree level is constituted through the aggregate learning activities which the student engages in, including informal forms of learning. This licenses my intuition that we shouldn’t let the perfect be the enemy of the good: if we can successfully tip (a) towards scaffolding in another instances then it will have implications for (b), not least of all because how a student uses the tool in one setting potentially has influence over how they use the tool in other settings. The problem at the moment is that we have effectively vacated the terrain by not offering adequate guidance and failing to meet students with regards to their own actually existing use of AI tools in informal learning as well as formal assessment.
We could see enterprise AI systems for universities as an attempt to go straight to (c) while (b) is still in a disorganised state. Simply making the infrastructure accessible on an equitable basis doesn’t do anything to ensure (a): this is the difference in a nutshell between diffusion and integration. Provision of the (right) infrastructure is the necessary condition for the integration because privately accessed consumer-facing subscriptions are an inadequate basis for the difficult educational work of (b). But it’s an insufficient condition. AI integration is fundamentally a pedagogical agenda built upon a technological apparatus provided by the institution. The risk is that provision arrives before the pedagogical and developmental work is in place so the infrastructure lands in the vacuum and gets used by default in substitution. It’s well motivated but without (b) it can actually make the problem it is trying to solve worse.
Mark Carrigan 