Using Cognitive Load Theory to Explore Impacts on Learning of Free Access to AI

Cognitive Load Theory (CLT), developed by John Sweller in the late 1980s, provides a conceptual framework for understanding how the human brain processes, stores, and retrieves information—i.e., information processing. I turn to it here as a heuristic, not as settled scientific knowledge. CTL emphasizes the limitations of working memory (WM), calling it the bottleneck in learning, using a model of WM from the 1980s which has been refined considerably, and offers strategies to optimize learning by managing cognitive load through instructional design, strategies about which I am not sanguine. Nonetheless, CTL as a heuristic that I can repurpose and reshape is very useful.

With the advent of artificial intelligence (AI), particularly large language models (LLMs) like ChatGPT, the strengths and weaknesses of information processing as the heart of learning have been exposed. This essay uses CLT's principles as a lens to examine how free access to AI tools might interact with these principles of learning not just within a transmission model of teaching, but within an improvisational model of teaching (including but not limited to direct instruction), reshaping opportunities and challenges in reading/writing and learning environments.

Cognitive Load Theory: An Expanded Framework for Learning

CLT as it is commonly seen today is not an appropriate framework for thinking ontologically about agentic reading nor writing enactments. Indeed, it places constraints on literacy instruction in arbitrary ways. Moreover, it assumes a direct model of instruction ceding considerable control of student thinking to teachers. It was built from the perspective of traditional best practice instruction where templates and prescriptions trump responsive and improvisational teaching. Teachers supply information; learners think (cognize) about this information to store it long term.

That said, CTL is grounded, rightfully so, in the understanding that working memory has a limited capacity, typically able to hold only 5-9 chunks of information at a time. Incoming information flows into a tiny space for massaging and flows out as more information comes in. Absolutely, WM is a bit of an evolutionary tragedy. Suppose humans had evolved a capacity to hold 500-900 chunks of information in WM at one time? Obviously, if the teacher’s role is to keep information moving through the brain, the pacing of a lesson place must estimate how much quality time is available to work WM at capacity before it breaks down. Stuff too much in WM and you have cognitive overload.

Long-term memory (LTM), by contrast, has virtually unlimited capacity and stores information in organized structures called schemas. During the 1990s theorists speculated that WM really has two stages, short term and working, and that LTM long-term memory has two stages—short long term memory and long long term memory. Regardless, the trick is to get information that evaporates from working memory in the time it takes to make a saccadic jump when reading safely and securely into long term memory in a retrievable form.

Classic current-traditional CLT identifies three types of cognitive load:

1. Intrinsic Cognitive Load: This refers to the inherent complexity of the material being learned. Tasks with high element interactivity—where multiple components must be processed simultaneously—impose greater intrinsic load. For example, solving a multi-step algebra problem requires understanding multiple interrelated concepts. Much of what I’ve read about classic CTL discusses the model in the context of STEM or of reading comprehension, particularly current ideological approaches to the reading comprehension specifying what knowledge everyone should learn across the grade levels. The concern is with reliable memory for fact-based frames of information or flexible algorithms.

2. Extraneous Cognitive Load: This is caused by factors unrelated to the task itself, such as poorly designed instructional materials or environmental distractions. These issues syphon off cognitive energy to deal with off-topic business, reducing the energy available to process what is supposed to be memorialized in long term memory. For instance, cluttered slides or unclear instructions increase extraneous load unnecessarily. Some extraneous seductions are beyond the control of the instructor, to be sure. As an issue of direct instruction where the world competes with the teacher for control of the learner’s mind, controllable extraneous load must be recognized and fixed during activity design. The following image from a website proffering help with reducing extraneous cognitive load reveals the sort of micromanagement of student thinking characteristic of implementation of the model:

Strategies for limiting the cognitive overload of your students

The Figure 2 illustration includes 7 vertical boxes to show how a 50-minute lecture could be organized into chunks to limit cognitive overload: Start with an advanced organizer, plan three 10-12 minute lecture segments with a 3-4 minute activity between each lecture segments, and end with a summary segment.

3. Germane Cognitive Load: Germane load is golden. This load takes the difference between intrinsic cognitive load - extrinsic cognitive load to represent the raw material that the brain has to adapt and adopt (germanity) to fit with other processed material (schemas) stored in long term memory. In a computational form, the equation is: gCL(iCL - eCL) = LTM profit. This represents the mental effort devoted to constructing and automating schemas. Activities like self-explanation or drawing connections between concepts enhance germane load, facilitating meaningful learning.

To optimize learning, educators aim to manage intrinsic load based on learners' expertise, minimize extraneous load, and maximize germane load through effective instructional design.

The Role of AI in Learning: Redefining Cognitive Load

The integration of AI tools into education has introduced new dynamics to CLT's framework. Free access to LLMs like ChatGPT significantly changes all three types of cognitive load.

Intrinsic Cognitive Load

AI tools can reduce intrinsic cognitive load by simplifying complex tasks. Read that sentence again. AI can reduce cognitive load by simplifying complex tasks. Remember that, prior to AI, the only way to reduce intrinsic task demands was by instructional design. AI creates at minimum a problem for instructional designers using a CTL model. Worst case scenario AI replaces the teacher.

Remember that levels of iCL available to the learner are determined by inherent complexity in the task and by how much prior knowledge a learner already has in LTM. Teachers are at a disadvantage on this point. AI can assess a learner’s level of prior knowledge and adjust the lesson instantly without reducing intrinsic demands. LLMs provide instant explanations and examples tailored to learners' needs, making high-element interactivity tasks more manageable. By breaking down problems into smaller steps or adjusting task difficulty based on user input, AI helps learners progress at their own pace. BF Skinner would be proud.

There could be a downside. Overreliance on AI may hinder long-term skill development. If students consistently offload complex reasoning tasks to AI without engaging deeply, they risk failing to build robust metacognitive schemas necessary for independent problem-solving. They may become AI-Dependent. Moreover, making and retrieving memories in and from longterm memory isn’t a simple cognitive pathway that can be strengthened.

As I understand it, memory formation (consolidation) involves distributed networks of neurons rather than single pathways. Information is stored across multiple brain regions in parallel. The metaphor of laboriously "carrying material" from WM to LTM along a pathway, one I’ve come across in blog discussions, isn't quite right. The metaphors of “friction,” what one might get from using a cart across a floor, or “stickiness” likewise risk oversimplification. Memory consolidation involves synaptic strengthening (long-term potentiation), protein synthesis, network reactivation during rest and sleep, and reorganization of neural circuits over time.

The idea of "schemas" in the brain is more complex than non-neurologists like me ever realized. While we do organize information in related networks, these aren't fixed "locations" where memories are “kept.” Memory retrieval involves reconstructing information from distributed networks, not simply following a pre-made "pathway." Each retrieval can actually modify the memory trace.

Research supports this understanding by showing that schemas are instantiated across multiple brain regions and play distinct roles in encoding and retrieval. For example, studies have identified the medial prefrontal cortex (mPFC) as critical for schema-based encoding, while other regions, such as the posterior medial cortex (PMC) and angular gyrus (AG), are more active during schema-guided memory retrieval. These findings highlight the distributed nature of memory storage and the dynamic interplay between regions during recall.

Moreover, schemas facilitate memory by scaffolding new information that aligns with existing knowledge structures. This process accelerates consolidation and enhances recall for schema-consistent details while potentially suppressing irrelevant or incongruent information. Retrieval itself is a reconstructive process that engages mechanisms like the medial prefrontal cortex to integrate and differentiate memory traces, allowing for updates to stored knowledge.In sum, schemas serve as flexible frameworks for organizing and retrieving information. Their distributed nature across neural networks underscores the complexity of memory processes, where retrieval not only accesses but also reshapes stored knowledge.

Note well: The CTL model is applicable in the context of direct instruction aimed at building organized funds of academic knowledge in long-term memory. Other sorts of learning—such as experiential, inquiry-based, or social-emotional learning—may not align as neatly with the cognitive load framework. These approaches often emphasize skills like creativity, collaboration, or problem-solving, which involve dynamic and open-ended tasks that are less focused on schema construction.

In such contexts, factors like intrinsic motivation, emotional engagement, and real-world application may play a more prominent role than the strict management of cognitive load. While elements of CLT, such as minimizing extraneous load or scaffolding intrinsic complexity, can still be relevant, these forms of learning often require a broader theoretical lens to fully capture their processes and outcomes.

Extraneous Cognitive Load

Teachers may have to spend time planning the details of CTL-based lessons designs. As noted earlier, intrinsic cognitive load depends on two factors: Complexity of the task and prior knowledge and experience of the learner. Task complexity may have some parameters to adjust within a lesson, but prior knowledge is inherently variable. To create truly effective CTL-based lessons, teachers would need to have high level of understanding regarding not only students general background in a subject, but their specific background regarding this particular topic.

LLMs can offer clear instructions and immediate feedback, reducing confusion caused by poorly designed materials—poorly designed because teachers are not mind-readers capable of assessing the knowledge level of every student let alone their moods, motivations, and mitigating circumstances. Students can ask specific questions and receive concise answers without sifting through irrelevant content. Yet, challenges remain. Students need to be guided in the fine art of prompting a bot, or inaccurate AI outputs can introduce new sources of extraneous load, including hallucinations. Students must develop skills to critically evaluate AI-generated content to avoid misinformation.

Germane Cognitive Load

In my opinion, AI tools have the potential to enhance germane cognitive load, but the degree of enhancement depends on the level of expertise the user has with AI and within the subject/topic area. LLMs can engage students in activities like generating hypotheses, evaluating arguments, or synthesizing information from multiple sources. By prompting learners with reflective questions or encouraging them to explain their reasoning, AI fosters deeper engagement with material.

However, germane load may decrease if students rely on AI for answers without actively processing information themselves. As discussed previously, the hard work of locking content in long term memory builds neural pathways that can be cued and retraced, making retrieval more robust. Coming upon new information, a learning can not only retrieve old schemas but bring new information in a remodel them—or erase them if they are wrong. Research suggests that excessive cognitive offloading can reduce opportunities for critical thinking and schema construction.

Balancing Opportunities and Challenges

While free access to AI tools offers transformative potential for education, it also presents risks that must be carefully managed. One of the greatest risks right now is the standoff between a traditional instructional paradigm modeled on principles of information processing and knowledge acquisition and an AI-enabled world where the differences between “school learning” and “real-world learning” have never been so clear.

Educational equity hangs in the balance. It’s already the case that affluent schools are doing more to integrate AI into the curriculum than high-poverty schools. Moreover, because affluent students tend to bring higher levels of prior knowledge and expertise with them as they move up the grades, they are likely going to have higher-quality experiences with AI and to be mentored and monitored by well-educated teachers to avoid the risks of cognitive offloading, the rabbit holes of hallucinations, and the life-long penalty of plagiarism.

AI is being lauded because it has the potential to democratize access to high-quality resources, but it will just as surely widen the gap between the haves and the have nots if students in poverty or living inside systemic prejudice and bias are not taught how to think about and use AI. Educators must integrate AI thoughtfully into curricula, ensuring it complements rather than replaces active learning strategies.

Above all, AI in the context of formal learning must not be viewed as a replacement for a human teacher. That’s like purposefully encouraging cognitive offloading, making memories in disappearing ink, ignoring the primarily sociocultural nature of learning, and dehumanizing the role of curiosity and emotions as the bedrock of learning intentions. AI gives off the illusion of teaching looked at through the lens of a transmission model of teaching, the model that informs original CTL. But through the lens of responsive and improvisational teaching, the coming wave in professional development and teacher preparation, AI has no illusions. It’s just a machine with a talkback function.

Cognitive Load Theory remains a vital heuristic for understanding how humans learn effectively. The introduction of free access to AI tools like LLMs reshapes this framework by impacting intrinsic and extraneous loads while offering new opportunities to enhance germane load. However, these benefits depend on how students and educators experience and use these tools. To maximize their potential while mitigating risks, educational systems must focus on fostering critical thinking skills, ethical awareness, and germane cognition—and on the human relationships between teachers and learners, the super power of the school.