The cycle that began with [YOU] on AI documents the winter of 2025–2026 as the moment when AI tools transformed from impressive demonstration to ambient infrastructure. Bjork is the cycle’s cognitive scientist of the cost that transformation carries—the cost that is invisible in every quarterly dashboard, every productivity metric, every user satisfaction survey, and visible only on the delayed test that almost no one administers. His framework explains a pattern the cycle identified empirically: the engineer who achieves a twenty-fold productivity multiplier with AI assistance and months later finds herself making architectural decisions with less confidence than before. The confidence erosion is not imaginary and not mysterious. The four hours of daily work that the AI tool eliminated had contained, embedded within the tedium, the desirable difficulties through which architectural intuition was built. The difficulty was the learning. When the difficulty disappeared, the learning stopped.
Bjork’s concept of storage strength versus retrieval strength reframes the human-AI relationship with surgical precision. AI tools operate as devices that maintain permanent maximal retrieval strength for any information the user might need. Every piece of knowledge is instantly accessible—and therefore never re-encoded through the effortful retrieval that builds durable storage strength. The user who can always ask the model never experiences the partial forgetting that would trigger the deep re-encoding that converts rented competence into owned capability. The result, played out over months and years of AI-assisted work, is a practitioner whose retrieval strength is maintained entirely by the tool and whose storage strength receives little or no investment. Remove the tool, and the gap becomes visible. The cycle calls this the dependency audit: the test almost no organization wants to administer, and the test that is, in Bjork’s framework, the only honest measure of what has actually been learned.
His fluency trap explains why the problem is self-concealing. The brain’s metacognitive monitoring system uses the ease of current processing as its primary proxy for learning depth. AI tools produce fluency at a scale and consistency that no previous technology approached. Each fluent AI interaction generates a strong positive metacognitive signal—the feeling of comprehension, of productivity, of mastery—that is systematically wrong. The practitioner who has used AI assistance for six months has not merely become dependent on the tool; she has become a worse judge of what she knows and does not know, because thousands of fluency-rich interactions have recalibrated her metacognitive monitoring toward overestimating independent capability. The illusion is self-sealing: the very confidence it generates prevents its detection.
The cycle’s prescription draws directly from Bjork’s research: the generate-first protocol, the deliberate introduction of spaced intervals between AI-assisted sessions, the institutional design of periodic dependency audits. These are not anti-AI positions; they are the structural interventions that Bjork’s forty years of evidence show are necessary to preserve the conditions for genuine expertise development in an environment that has made ease the default. The difficulty is not an obstacle. It is the substrate.
Born in 1939, Bjork received his doctorate in mathematical psychology at Stanford in 1966 and moved to UCLA in 1974, where he has remained. His early work focused on human memory architecture and the mechanisms of interference and forgetting; by the 1980s this had coalesced into a sustained research program on the relationship between learning conditions and long-term retention. The key early finding—that massed practice produces high immediate performance and poor delayed retention, while spaced practice produces the opposite profile—was not new; the spacing effect had been documented by Hermann Ebbinghaus in 1885. What Bjork contributed was the theoretical framework that explained all four desirable difficulties under a single account—the account that difficulty is desirable when it engages the cognitive processes of effortful retrieval and deep encoding—and the experimental rigor that accumulated a thousand-study replication base for each component.
The New Theory of Disuse, developed with Elizabeth Ligon Bjork and published in 1992, was the framework’s theoretical crown: the proposal that every item in memory possesses two independent strengths, storage and retrieval, that respond differently to learning conditions and that can be simultaneously high or low in any combination. This independence resolved the apparent paradox of desirable difficulties—how spaced practice can produce worse performance during learning while producing better performance on delayed tests—and made it possible to specify precisely what AI tools do to the human memory architecture: they maximize retrieval strength while starving storage strength of the effortful-retrieval events that would build it.
In the 1990s and 2000s, Bjork extended the framework to the metacognitive dimension: the systematic study of why learners fail to choose the most effective conditions even when given accurate information about what those conditions are. The fluency heuristic—the automatic interpretation of easy processing as evidence of good learning—turned out to be remarkably resistant to correction, operating below the level of conscious belief and producing overconfidence even in learners who could articulate the principle and identify the trap in hypothetical scenarios. This finding is perhaps the most sobering of his career: the problem is not ignorance. It is architecture.
Desirable difficulties. Four canonical conditions enhance long-term retention and transfer while degrading immediate performance: spacing (distributing practice across time so that partial forgetting forces effortful re-encoding), interleaving (mixing problem types so that categorization must precede solution), the generation effect (requiring the learner to produce an answer before receiving one), and variation (practicing under varied conditions to produce flexible encoding). Each is supported by hundreds of independent replications across domains. AI tools eliminate all four simultaneously in their default operation. Desirable difficulties are not arguments against AI; they are a specification of what must be deliberately preserved if AI-assisted work is to develop rather than erode capability.
The performance-learning dissociation. Performance and learning are not merely different; under many conditions they are inversely related. The conditions that maximize current performance—massed practice, blocked problems, immediate feedback—minimize durable learning. The conditions that maximize durable learning degrade current performance. AI tools optimize exclusively for performance. No AI product is evaluated on what the user can do next month without the tool. The entire feedback loop governing the design, deployment, and adoption of AI tools operates on the wrong metric.
The fluency trap and metacognitive illusions. The brain’s metacognitive monitoring system uses processing fluency as its primary proxy for learning depth. This heuristic was adaptive in environments where fluency correlated with familiarity and familiarity with genuine understanding. AI tools produce fluency without the repeated encounter and effortful engagement that fluency was evolved to track. The heuristic is intact; its ecological validity is destroyed. The result is the fluency trap: every AI-assisted interaction generates the metacognitive signal that learning is occurring, while the conditions that produce durable learning have been bypassed.
Storage strength vs. retrieval strength. The New Theory of Disuse proposes that memory has two independent dimensions. Storage strength reflects encoding depth and increases monotonically—it only rises, with each genuine encoding event. Retrieval strength reflects current accessibility and fluctuates constantly, rising with recent exposure and falling with time. The conditions that maximize retrieval strength actively undermine storage strength. AI tools maintain permanent maximal retrieval strength for any information the user needs; the price is that the effortful-retrieval events that would build storage strength never occur. The practitioner is left with rented competence that exists only in the presence of the tool.
The generate-first protocol and dependency audit. The structural interventions Bjork’s research implies are technically simple and commercially almost impossible. The generate-first protocol—attempt the problem before consulting the AI—preserves the generation effect while allowing subsequent AI assistance. The dependency audit—periodic assessment of independent capability without AI assistance—is the only reliable measure of whether storage strength is being built alongside retrieval strength. Both interventions reduce short-term performance in the service of long-term capability. Neither is rewarded by market metrics.