ep. 68. Cognitive Load Theory as a Guide for AI Design
5 min read
How can cognitive psychology inform AI tools that support learning, reasoning, and decision-making?
Today’s episode explores this question, grounded in content I shared yesterday in my EPIC People tutorial, Designing AI to Think with Us, Not for Us: A Guide to Cognitive Offloading.
Cognitive Load: A Primer
Cognitive load refers to the amount of mental effort being used in working memory, the brain’s limited-capacity system for temporarily holding and processing information. Think of working memory as a mental workspace, and cognitive load as how full that workspace is. When a task is too complex, poorly structured, or surrounded by distraction, it can overwhelm this system - leading to cognitive overload and impairing learning and decision-making.
Cognitive Load Theory breaks cognitive load down into three types:
Intrinsic Load: The inherent difficulty of a task. For example, solving a calculus problem carries more intrinsic load than basic arithmetic. Designers can reduce intrinsic load by chunking information, sequencing steps, or adapting to the user’s level of expertise.
Extraneous load: Load added by how information is presented. A cluttered dashboard with unlabeled metrics increases extraneous load compared to a clean, focused visualization of key engagement trends. Good design reduces this type of load through clarity, hierarchy, and relevance.
Germane load: Effort spent constructing meaning - connecting ideas, reflecting, and integrating knowledge. For instance, paraphrasing a complex idea in your own words increases germane load and supports deeper learning. Unlike the intrinsic and extraneous load, we want to preserve or even amplify germane load.
Reducing intrinsic and extraneous cognitive loads allows you to allocate more cognitive “bandwidth” to learning-relevant processes, enabling more meaningful engagement with the task at hand.
Cognitive Load x AI
AI systems like ChatGPT and Claude can be powerful tools for reducing intrinsic load by breaking down complex tasks - such as helping break a challenging math problem into smaller, more manageable parts.
They can also help minimize extraneous load by presenting information in more digestible formats, like turning a technical document into a bullet-point outline or visual summary. Google’s NotebookLM’s Mind Maps is a great example of this, automatically organizing dense content into interactive visual structures, helping users see relationships between concepts. Together, these capabilities make AI a compelling aid for learning and iteration.
However, when we accept AI outputs at face value, we risk offloading germane load. That means skipping the mental steps needed to reflect, evaluate, or contextualize. We’re inclined to do this because of automation bias, our tendency to favor information from automated systems.
This offloading of germane load helps explain the growing body of research suggesting that frequent AI use can erode critical thinking. These negative effects of "over-offloading" are most pronounced in users who exhibit high confidence and trust in generative AI output, as well as younger users.
How might we design AI tools that reduce intrinsic and extraneous load, without offloading the cognitive effort that supports learning, reasoning, and decision-making?
AI as Sparring Partner, Not Answer Machine
One promising approach is to position AI as a sparring partner - a collaborator that challenges and supports users in cognitive effort, rather than bypassing it. This framing prompts critical engagement and increases germane load.
For example, Microsoft Research has explored this in Excel, where AI suggests actions like “consider adding trendlines”, inviting users to reflect more critically on their data. Similarly, Lex, an AI writing tool, allows you to run a “check” for your document (e.g., for grammar, readability, clichés) and accept or reject suggestions. This encourages users to actively engage with AI’s output, rather than passively accept them.
These examples introduce productive friction - a small but intentional nudge toward deeper thinking. Rather than making tasks harder, they make reflection feel natural and rewarding, while reinforcing user autonomy and keeping them in the loop.
Takeaways
Use Cognitive Load Theory as a lens for AI design. Reduce complexity (intrinsic load) and clutter (extraneous load), but preserve the mental effort that drives understanding (germane load).
Design AI as a Sparring Partner. Move beyond polished outputs and build in interactions that support exploration, comparison, and iteration.
Build friction that earns engagement. Encouraging reflection in fast-paced workflows is a challenge. Make these touchpoints lightweight and contextual. The goal is to deliver value and empower users, not create obstacles.
Sendfull in the Wild
Human-AI Musical Collaboration at Adobe Boards Launch
I had the chance to perform at the launch party for Adobe Firefly Boards, a generative AI-first approach to exploring concepts. I improvised on electric violin over AI-generated tracks. This technology was developed by Nick Bryan in Adobe Research, with tracks created by Ross McKegney, Head of Firefly Boards.
This performance was followed by a collaboration with multidisciplinary artist Ada Cyborg, who incorporates Boards into her creative practice. I played violin live to her original music video, Ada’s Lullaby - a piece she created in collaboration with generative AI tools.
May 21: UpScale Conf Panel
I’ll be joining AI x Design leaders Brooke Hopper and David Montero for a discussion on how you design the future when the future doesn’t exist yet. Get your tickets here: www.upscaleconf.com/#section-get-tickets
Human-Computer Interaction News
Waymo safety data: Compared to human benchmarks over 56.7 million miles, Waymo autonomous vehicles had safer interactions with vulnerable road users such as pedestrians, cyclists, and motorcyclists, and 96% fewer injury-involving intersection crashes.
Making AI models more trustworthy for high-stakes settings: Researchers at MIT developed a technique to improve the trustworthiness of machine learning models, which could enhance the accuracy and reliability of AI predictions in high-stakes settings such as health care.
CHI Best Paper Awards: Check out the Best Paper and Honorable Mention winners from the leading HCI conference, CHI, held April 26 - May 1 in Yokohama, Japan. One of my favorites? AMUSE: Human-AI Collaborative Songwriting with Multimodal Inspirations.
Designing emerging technology products? Sendfull can help you find product-market fit. Reach out at hello@sendfull.com
That’s a wrap 🌯 . More human-computer interaction news from Sendfull in two weeks!