In this blog, former teacher and Senior Professional Development Manager at The Assessment Network, James Beadle explores cognitive offloading and how it can act as both a positive and negative force in the context of educational assessment.
What is cognitive offloading?
How many phone numbers can you remember?
Do you even know your own?
In a YouGov survey carried out in 2023, 73% of respondents had memorised just two or fewer, with 32% saying they had memorised none at all.
Thirty years ago, prior to the mass uptake of mobile phones, the figure would have likely been significantly higher; as a young teenager, I distinctly remember on several occasions having to call my parents (and reverse the charges) from a phone box when it looked like they had forgotten to pick me up.
What has occurred since then is a notable example of cognitive offloading.
With all of us carrying around mobile phones in our pockets, there is little need to remember numbers when our devices do so for us.
Through technological adoption, we have ‘offloaded’ the cognitive task of memorising numbers.
Simply put, cognitive offloading is the process of using tools or other external resources to reduce the mental demands of carrying out a particular task.
Is it just memory we offload?
Whilst the term ‘cognitive offloading’ is often associated with delegating memory to an external source, other cognitive skills can also be offloaded.
From the abacus to the electronic calculator, as humans, we can offload the skill of carrying out mathematical calculations to an external device.
In everyday life, some high-end cars now come with ‘parking assist’ as a feature, allowing drivers to offload the task of parallel parking – likely a great source of relief for some individuals.
Perhaps most pertinently, the ongoing growing usage of generative AI is intrinsically linked with cognitive offloading.
With AI models now able to undertake a variety of tasks, such as generating imagery, responding to emails, or even controversially completing academic assignments, we, and our students and candidates, are able to offload an extensive range of cognitive activities across many different domains.
Is cognitive offloading a threat, or an opportunity, for assessment?
Many of the initial concerns around the use of AI in assessment revolved around the ‘inappropriate’ use of AI by students to carry out tasks we would have previously expected them to do themselves. This offloading is often viewed as a threat to the validity of the assessment.
If on an essay-based assessment designed to measure the ability to engage in critical thinking, a student uses a language model to generate their response, then clearly that assessment is no longer valid.
In such conditions, we may wish to either ensure that the assessment takes place in controlled conditions that do not permit the use of AI or redesign the assessment in such a way that students cannot offload the critical thinking element, such as including a live question and answer component.
However, taking a step back, it is important to recognise that many forms of cognitive offloading are already permitted or will be permitted in future high-stakes assessments.
Examples include:
The importance of scaffolding in assessment
In the examples above, the cognitive offloading acts as a form of scaffolding.
This is the process by which support is provided to learners, enabling them to carry out more complex tasks. Other examples might be breaking tasks down into small parts, giving a worked example for students to follow, or helping them develop an initial outline for an essay. In many cases, scaffolding is intended to be temporary and removed once learners have demonstrated mastery of the more complex task.
In assessments, by providing additional support to candidates and reducing the cognitive load, we can enable them to carry out more complex activities.
In the case of an English exam, shifting the focus from memorising excerpts of texts word-for-word will allow students to instead focus on exploring the underlying themes of the text – something we are likely much more interested in assessing.
AI potentially provides similar opportunities for scaffolding. Examples include:
- Emerging tools in research can support literature reviews, quickly highlighting connections between academic papers and themes, allowing learners to explore areas they may not have otherwise identified as connected to their field of study.
- Language models can be prompted to act as critical friends and coaches, challenging students and helping them further expand on their initial thinking.
Should AI be used in all assessments?
So, when is it appropriate to incorporate AI usage into assessment?
Ultimately, it all comes down to a single key question: what are we trying to measure?
If it is a cognitive skill or a set of knowledge that can be fully offloaded to AI then such usage is likely inappropriate.
If on the other hand, AI usage may support learners in better showing what they know and can do, in the associated domain, by offloading more routine or trivial tasks, then such usage should likely be allowed, and even encouraged.
Of course, if a cognitive task can be fully offloaded to AI, it can pose the question ‘Should we be assessing this at all?’. In some cases, this is worth considering.
History has some strong examples of such changes driven by technology:
- With the advent of electronic calculators, the emphasis on long division in many curriculums has been reduced.
- In the 1970s and 1980s, it was possible to study an O Level in Geometrical and Mechanical Drawing – a set of technical skills that are now done by Computer-Aided Design (CAD) software, leading to the qualification becoming redundant.
It might be that soon there is another shift in stakeholder consensus about some skills or activities.
This could result in some topics no longer being assessed or a change in the relationship and measurement of how the students work with AI to achieve those desired outcomes.
However, there are two things we should keep in mind: firstly, complex skills require mastery of simple skills, and secondly, difficulty is often desirable.
Complex skills require mastery of simple skills
Whilst teaching, I often reflected on how simple skills lead to complex outputs. For instance:
- Students need confidence in their ability to spell before writing fluently: if you’re constantly questioning how to spell every other word you write, it’s difficult to think about the next sentence you’re going to write.
- Multi-digit multiplication, division and algebraic manipulation all depend on mastery of multiplication tables.
These are clear examples where it is certainly not desirable for these key skills to be regularly cognitively offloaded. If students are constantly reaching for a dictionary or a calculator, the continual process of disengaging and reengaging with the task at hand will place a significant burden on their working memory, hindering their performance.
The issue of overreliance on AI as a student
At a deeper level, if we want students to analyse the complex relationship between economic growth, quality of life and environmental degradation, they will likely need to have a familiarity with various case studies in a range of contexts.
Cognitively, it is far easier to make comparisons and identify complexities if you already have a selection of key examples and associated facts stored in memory.
In the case of economic growth, students studying this area need to understand key relationships, such that growth can often lead to affordable healthcare but may also lead to a greater incidence of diseases such as asthma.
With AI now able to carry out higher-level thinking skills, students can potentially offload a far greater range of activities.
However, if they have yet to master these skills themselves, then this is likely to be significantly detrimental for them in the long run.
A study by MIT showed that the use of generative AI during essay writing led to significantly less cognitive activity and looked to hinder long-term learning, whilst a study by Microsoft found that generative AI use “can inhibit critical engagement with work and can potentially lead to long-term overreliance on the tool and diminished skill for independent problem-solving” (Lee et al, 2025).
In short, when considering if students should be using AI tools to carry out a particular skill, we should first ask: Is this a skill they need to master?
If the answer is yes, then AI use is likely inappropriate, unless they have already shown they have mastered the skill.
For example, AI agents are increasingly adept at programming. However, this hasn’t removed the need for those working within the computing sector to understand code.
The phrase ‘human in the loop’ is often thrown around; if humans are to add value in these environments, potentially directing, reviewing and analysing the work of AI models, then they will need a substantial knowledge base – critical thinking simply isn’t possible without it.
In short, even if AI is now great at programming, it hasn’t removed the need for students studying computer science to learn how to code; rather, once they have mastered coding, they can then look at how to best incorporate AI usage into their practice.
Difficulty can be desirable in assessment
One of the most important, and perhaps counter-intuitive, ideas in education in the last 50 years comes from the work of Bjork and Bjork (2009), who argue that some forms of difficulty whilst learning (and assessing) are often desirable.
Activities like carrying out tasks in different contexts, interleaving (mixing) practice across different topics, and having to engage in active retrieval and generating an answer (rather than looking it up or having someone show you) lead to better learning over time, despite being more challenging.
This was demonstrated recently in a study by Cambridge University Press & Assessment and Microsoft Research, which found that students who engaged in active note taking retained more information over time than those who solely used an LLM.
This highlights the importance of separating learning from performance: activities (such as looking up answers or focusing on a single area of knowledge at a time) that lead to higher initial performance do not necessarily lead to successful long-term learning.
Generative AI has the potential to make educational tasks significantly easier for students.
The brief appearance of the ‘Einstein’ agent (no longer available), which directly integrated with learning management systems and could actively read course material, complete quizzes and post replies in discussion forums, essentially allowed students to complete online courses without ever engaging in the taught content.
Whilst this is perhaps an intuitively inappropriate use of AI, even more potentially appropriate uses, such as using ChatGPT to support with essay writing (as in the MIT study), look to hinder long-term learning: this simply isn’t a desirable way of making tasks easier.
The importance of assessment literacy
So what do we do?
It is neither practical nor desirable to carry out formative assessment tasks in controlled conditions; by their very nature, these tasks work best when they are low-stakes and need to take place in a range of environments, including outside of the classroom.
Nor is this necessarily new: students have often been able to inappropriately ‘offload’ homework tasks, by either sourcing answer sheets, using the internet, or simply copying from a peer.
When I was a teacher, I never viewed this as cheating – rather they were simply wasting their time, engaging in a rather pointless activity that was unlikely to help them learn in the long run!
The solution in this case is to recognise that as students age and are more likely to engage in independent behaviour that is undesirable (such as ‘cheating’), then they need to possess a degree of pedagogical and assessment literacy.
In my experience, we need to educate students about the process of learning and the importance of mastering skills that they can then build on, even if such skills can be offloaded to an AI.
If a student asks ‘Why do I need to learn multiplication if I can always use a calculator’, the answer is not ‘Well one day you might be in a supermarket without one’, but rather ‘Because it helps you then learn other areas of mathematics, which you need to understand if you want to design video games, or be a psychologist, or a pilot, or an architect’.
In a similar manner, a student might ask why they need to be able to write a clear, persuasive email if this can instead be done by an inbuilt AI agent; in this case, a suitable answer might be ‘because one day you may be running your own business and want to implement an automated email response system. At that point, you will need to know what a good reply to an email looks like, which is best done by learning how to write one yourself.
Ultimately, in any future that involves AI, we will still want humans to be the arbiter of what a ‘good’ output looks like, which means we need to be able to produce one ourselves – even if we can’t do it as quickly as AI!
James Beadle is Senior Professional Development Manager at The Assessment Network. You can learn more about his recent work on enhancing assessment literacy in this case study.