While having a coffee with a colleague on a military base a few years ago, I was asked this simple question.
He had found OpenAI’s LLM, started tinkering, and liked what he saw, but had forgotten its name. He had comically replaced GPT (generative predictive transformer) with GMPG (general-purpose machine gun). Had stumbled onto something? ChatGPT (other LLMs are available) and agentic AI are belt-feeding information in ways that may be biologically implausible. Are we approaching a biological ceiling of information processing?
The general-purpose machine gun (GPMG) is an air-cooled, belt-fed automatic weapon that serves as both a light and medium machine gun. The GMPG is widely used because it is reliable, versatile, packs a punch, and can fire up to 750 rounds per minute. The GPMG’s belt-fed mechanism enables sustained fire, a significant advantage in warfighting.
In other life domains, less is sometimes more. “Sustained fire” and belt feeding of information, ideas, and work tasks can result in significant fatigue, cynicism, demotivation, and underperformance, particularly if there is a sense of overload and overwhelm combined with insufficient time for rest and recovery.
This problem is based on a biological ceiling effect. Humans cannot process information in the same way that machines do. For AI, more information can improve the output because of the training model. With humans, more information can undermine performance. We have an inherent bottleneck in the way we process information. When we belt-feed information, without pause, the system is overloaded, and performance can decrease.
Here we have a problem and an opportunity. Agentic AI can be used to offset cognition. We can offload tasks to agents so that we don’t have to think about them. No more booking padel courts in your lunch hour, just ask your virtual personal assistant to book them for you! The problem emerges when the constant flow of outputs from those agents occupies our attention and prevents us from resting, particularly when there is a lack of trust in the agent. Therefore, keeping the human ‘in the loop’ (i.e., checking all outputs and decisions) can become overwhelming.
The problem isn’t a lack of information; very few successful performers suffer from a lack of information. What to do with that information, how to process the volume, and a lack of certainty are much greater problems. AI undoubtedly helps generate volumes of information, some of which would be virtually impossible to generate without AI, but when information is not the problem, AI tools might not be the answer.
In my experience, this is particularly true of people who suffer from high degrees of perfectionistic concern. They worry about not being good enough, they cannot find satisfaction in their achievements, and the opportunity to stay connected and switched on is too great to ignore. The perfectionist has a voice in their head whispering that people expect them to do more, that they need to be working, and that the AI agent can help them increase productivity. The capability, opportunity, and motivation to keep working converge to produce very unhealthy behaviour – they never stop.
Cognitive load is a term used by psychologists to reflect the magnitude of mental effort that tasks require in any given moment. The level of cognitive load can affect how we think, make decisions, and behave. Whether a task (or a string of tasks) feels manageable, effortful, or overwhelming) could be caused by the level of cognitive load allocated to the task(s). Cognitive Load Theory provides a research basis for understanding what happens during cognitive overload. From a cognitive load theory perspective, breakdowns in performance appear when working memory, the relatively short-term store of information that reflects the conscious process of thinking, is challenged beyond its capacity. Compared with long-term memory, working memory is very short-term, lasting around 20 seconds (for most untrained people) or approximately 7 chunks of information. For example, if asked to remember a phone number, we cannot simply remember the string of numbers in one go. Instead, we use techniques to increase the amount we can remember. For example, we chunk the numbers into groups, or we try to repeat them continuously in our heads, until someone asks us a question or distracts us, and we forget. Working memory is also limited by concurrent processing potential. We cannot remember the phone number while being asked for directions or where we put the phone charger. When chunking, we are limited to around four chunks, or four groups of information. For example, the phone number 07775 533262 can be broken down into 4 groups of digits: 0 (we don’t need to remember that, since all phone numbers in the UK start with 0!), 777, 55, 33, 262. More than four chunks, and we approach processing capacity, and working memory is stretched to the breaking point.
What is Cognitive Load?
When the rubber hits the road, stretched working memory results in impaired learning and an inability to process and, therefore, leverage several streams of information. As I have said, the problem is not a lack of information; the constraint is encoded in our biology. Our brains have evolved to process limited amounts (or chunks) of information, not the belt-feeding of information from AI and her agents.
AI agents can temporarily offset cognitive load and, in doing so, increase working memory capacity. However, this effect is potentially temporary if we need to interpret the outputs. We appear to be kicking a can down the road.
The load associated with constant connection (emails, WhatsApp, Slack, etc) compete for attention and fills our cognitive buckets to the point of overflow. We offload the problem to our digital assistant, but at some point, in some form, the AI system generates outputs at scale that require our attention (either reading or signing off). The number of outputs produced by AI agents is often at a scale that would have been impossible a few years ago, let alone a few hundred thousand years ago when the neural “hardware” evolved. Because reports, summaries, and analyses can now be produced in seconds, we may be simply belt-feeding more information without truly understanding the consequences of reading, interpreting, judging, challenging, and questioning AI-created outputs. Each process incurs a cognitive cost and diminishes working memory, especially when AI outputs arrive all at once.
So, what happens when we are overloading our working memory? We find shortcuts. Our information-processing system will find ways to conserve working memory, so we start to skim-read, take shortcuts, and blindly accept the information presented without critical appraisal of what it is saying. This pattern of responses is normal and highly predictable in the face of cognitive overload.
So, What Can we do About Cognitive Overload?
The key is to plan AI use in line with cognitive load theory to ensure cognitive offloading doesn’t come back to bite you by overloading the system later. Within cognitive load theory, scholars have suggested that there are different types (or sources) of cognitive load: intrinsic, extraneous, and germane. The intrinsic load represents the mental effort required for the primary task, for example, learning something new. The extraneous load can be considered the “noise” or distractions that can interfere with the primary task. Finally, germane load is the background processing, the building blocks of thinking, that we use and reuse to help us use information efficiently the next time we are presented with it (psychologists sometimes label this mental shortcutting as schemas). If extraneous load or non-germane load (things that don’t build useful or relevant schemas) is too great, we compromise our working memory capacity.
It is my belief that AI, specifically the misuse/overuse of AI agents, risks adding unnecessary extraneous load because information is presented at a pace, without context, in non-germane forms (i.e., fragmented across platforms and unstructured). There is a need for humans to order, interpret, and judge AI-produced information before it can be used, which risks overloading working memory capacity. Basically, you need to piece information together, which takes a cognitive toll before meaningful thinking about the information presented can begin.
When cognitively overloaded, we cannot combine or recombine disconnected concepts. These disconnected ideas add to the noise and distraction. This is particularly true for situations characterised by volatility, uncertainty, complexity, and ambiguity (VUCA). VUCA environments, a term popularised in the military, are unstable, unpredictable, and can change quickly. Decisions need to be made in the absence of certainty, which AI agents are unlikely to help with. Pattern-forming, probability engines like AI cannot make decisions in the way humans do, at least not yet.
Moving forward, we can plan for cognitive overload by considering the sources of intrinsic, extraneous, and germane load before introducing support tools such as AI agents. We can optimise our cognitive performance by managing the intrinsic load, reducing the extraneous load, and maximising the germane load.
I don’t want to throw the baby out with the bathwater. AI is here to stay and is potentially helpful. The capacity to process large volumes of data can reduce extraneous load if used strategically. However, when the situation requires interpretation, judgment, agility, and adaptation, the volume of data generated by AI systems can be problematic.
In terms of optimising cognitive performance while offloading cognition to AI systems, we can consider a couple of tools. Firstly, if we recognise that unfiltered avalanches of information from disconnected AI systems can be the source of extraneous load, we can create a “master” agent of sorts that provides a filter so that the data isn’t dumped on you in a “oner”. The master agent (similar to what Perplexity has called the computer) aggregates and formats information into a single output. Having a master agent can reduce extraneous load. Next, consider prompting your master AI in a way that manages the intrinsic load. For example, ask for a BLUF (bottom line up front) or a succinct executive summary. These types of output are essentially “chunks” of information. As discussed earlier, we know working memory is limited to around 20 seconds, 7 pieces of data, or concurrent processing of around 4 chunks. We need to ask the AI agent to produce outputs that match our working memory capacity. AI won’t do this unless we ask it to.
If you cannot use a master agent and have multiple streams of information, consider using dual coding, which engages our different senses to process information. A few years ago, I was involved in a research project in which operators were asked to watch a screen (visual input) on which potential threats were shown. At the same time, the operators needed to control multiple systems, all displayed on computer screens. The problem of switching between screens caused cognitive overload. The answer was to use visual, auditory, and haptic “alarms” for different threats and systems, which increase cognitive capacity. Now that we know about the threats posed by different sources of cognitive load, we can request filtered outputs by using both visual and audio media. Finally, the germane load represents the ordering of information and the creation of schemas to aid future use. If we are constantly interrupted by AI agents producing more outputs, we will never reach a state of interpretation and understanding. We can prompt the agent to set up our work schedule so that AI outputs are delivered to us within a pre-specified window. The constant interruptions can prevent the formation of schemes, so we can avoid this by creating a system of disconnections. This gives us “quiet time,” or what Cal Newport calls “Deep work time.”
Conclusion
If we belt-feed information, we risk overloading the system. When we receive information continuously, there is limited time to think, question and challenge assumptions. When we lose the cognitive space to judge and interpret what we read, we are forced to make decisions more quickly, and the quality of our thinking and performance diminishes. Instead of accepting this situation, we can apply cognitive load theory to identify factors that support or hinder our cognitive performance. We can use AI in collaboration with an understanding of how we process information (and what can get in the way), so we can offload cognition efficiently and ensure that the outputs of that offloading don’t turn around and bite us later.
