There is an assumption running through most AI learning strategies right now that deserves more scrutiny than it tends to receive. The assumption is that giving employees smarter tools will make them smarter learners. It sounds logical, and it may even feel obvious, but we have seen this pattern play out before in other areas of life, and the outcome was not what anyone intended.
The GPS analogy is not a metaphor
In the early years of GPS navigation, the technology worked exactly as promised. People got better at getting from A to B, journey times shortened, and wrong turns decreased. By every immediate metric, the tools were delivering. Then, over time, a different picture emerged. Research published in Scientific Reports (2021) found that habitual GPS use was associated with reduced hippocampal activity, the region of the brain central to spatial memory and independent navigation. A separate study by Bohbot et al. (2020) showed that regular GPS users performed significantly worse on self-guided navigation tasks compared to people who had navigated independently over the same period.
GPS made people better navigators until it made them incapable of reading a map. Convenience came with a cognitive trade-off, and the trade-off only became visible once the capability that had been quietly atrophying was actually needed. This is not a failure of the technology itself. It is a predictable consequence of how human cognition responds when a task is routinely handed over to an external system. The question for anyone designing or buying AI learning tools is whether we are building the same dynamic into workplace learning, and whether we will notice it before it matters.
What the research shows about cognitive offloading
The science behind this effect is well established. Risko and Gilbert (2016), writing in Trends in Cognitive Sciences, documented cognitive offloading as the process by which people delegate mental tasks to external tools when those tools are available. Done selectively, this is sensible and adaptive. Done habitually for tasks that would otherwise build capability, it erodes the skills the tools were meant to support.
The more recent evidence specific to AI gives this a sharper edge. Fan et al. (2024), publishing in the British Journal of Educational Technology, identified what they called “metacognitive laziness” in learners using generative AI: a measurable reduction in willingness to monitor their own understanding, plan their own approach to learning, or persist through difficulty when an AI answer was immediately available. When the answer is always a click away, the habit of working through confusion until something clicks begins to erode, and with it, the capacity for independent thinking that most organisations are actually trying to develop.
This sits alongside the broader challenge of cognitive load in the workplace. AI that adds more noise, more choices, more content to navigate, creates the same problem from a different direction.
Why removing all friction is not the goal
One of the most consistent findings in learning science is that the conditions which feel hardest tend to produce the best long-term retention and transfer. Bjork and Bjork (2011) showed that retrieval practice, spaced repetition, and interleaved content reliably outperforms smoother, more comfortable learning conditions when it comes to what people can actually do with knowledge later. The testing effects documented by Roediger and Karpicke (2006), and Cepeda et al.’s synthesis on distributed practice (2006), point in the same direction: the conditions that require more cognitive effort tend to produce more durable learning.
An AI system that removes all friction, that always answers, always scaffolds, and always makes the learning experience feel easy, may be generating completion rates and satisfaction scores while quietly undermining the capability it was designed to create. This is not an argument against AI in learning, but it is an argument for being deliberate about where friction is removed and where it should be preserved. Good bespoke learning design has always made this distinction. AI makes it more pressing, and when approached with that awareness, it also makes it more achievable at scale.
The difference between supporting thinking and replacing it
There is a meaningful distinction between AI that supports thinking and AI that replaces it, and it has significant implications for how platforms should be evaluated and designed.
AI that supports thinking might surface a useful question at the right moment, help a learner connect a development goal to a concrete next action, or reduce the cognitive noise that makes it hard to get started. It makes effort feel possible rather than unnecessary. AI that replaces thinking provides the answer, completes the task, or removes the productive struggle that would have produced understanding. It creates the experience of learning, including the interaction, the output, and the completed module, without necessarily building the capability the learning was supposed to create. Most AI learning tools, if evaluated honestly against this distinction, have not been designed with it in mind. The more useful question to ask of any platform or tool is not whether it delivers learning, but whether it builds learners.
What self-determined learning actually requires
The goal of a genuinely healthy learning culture is not employees who are proficient at using a particular AI engine, platform or tool. It is employees who are good at learning, with or without it, and who can direct their own development as their role and context evolve.
Heutagogy, developed by Hase and Kenyon (2000) and expanded by Blaschke (2012), describes self-determined learning as the capacity to identify what you need to learn, find the resources to do it, and reflect critically on the process. It goes a step beyond self-directed learning, where a learner follows a path someone else has designed, to a learner who can design their own. That capacity does not emerge automatically, and it can be systematically undermined by platforms that consistently take the navigation out of learners’ hands. Saffron’s approach to developing self-determined learners has been grounded in this principle for years. What AI introduces is the possibility of supporting that kind of development for every person in an organisation, not just those with access to a coach or a generous development budget.
What the evidence shows when the design is right
HR Grapevine’s examination of AI coaching in practice (2026) found that outcomes varied significantly depending on whether a tool was designed to prompt learner reflection and agency, or simply to provide answers and content recommendations. The technology itself was not the determining factor. The intent behind the design was.
Our AI coaching for skills platform, AIDA shows, that 90% of participants discovered possibilities they had not previously considered. The tool opened doors rather than walking through them on the learner’s behalf. Confidence in career progression moved from 36% to 92%, and sense of control over development increased substantially. You can read more about what drove those results and how AIDA approaches the challenge of building learner agency.
To explore how AIDA is designed to build learner capability and autonomy, visit the AIDA page or speak to the Saffron team.
