Cognitive Tribology: Friction, Wear and the Long-Term Effects of AI Assistance on Human Cognitive Capability

Version 1.3—August 2026 This version incorporates additions published after the original submission date. Section 5.1 adds Strömberg et al. (2026) reporting a difference-in-differences study of 26,811 Chinese students over 30 months in which AI adoption raised homework scores by 18% while lowering closed-book exam scores by 20% within six months and high-stakes entrance examination scores by 18–24% after two years, the largest causal evidence yet for the atrophy mechanism's temporal signature and the shifting-the-burden archetype quantified at educational scale. Section 5.2 adds: Anders and Dux Speltz (2026) identifying three threshold concepts students must cross before writing productively with AI—experimentation, expertise, and agency—with the central finding that productive AI use requires more metacognitive effort rather than less. The authors independently coin the "fluency trap" for the phenomenon cognitive tribology names phantom friction in the writing domain, and identify the shift from outsourcing to orchestrating as the reversibility standard stated as a writing pedagogy principle. Brcic and Frljic (2026) synthesizing causal evidence that an unguarded AI helper left students 17% worse on unaided exams while the same model rebuilt to withhold answers erased the harm, demonstrating that outcomes flip entirely on design. Their Six Moves framework independently arrives at the friction-before-output principle as a classroom access-timing gate and names the reversibility standard as the AI-free test point, while coining "illusion of learning" for the phantom friction construct in educational contexts. Wei et al. (2026) demonstrating that learning via short videos produces significantly lower memory accuracy, higher forgetting rates, and reduced neural synchrony in regions supporting episodic memory and cognitive control compared to long-form video, with frictionless fragmented content enhancing bottom-up attentional capture at the expense of the top-down processes critical for deep learning. The finding complements Sakai (2026) already in the paper to establish a cross-format neural evidence base for the productive friction principle: frictionless content design consistently reduces the neural engagement that durable learning requires. Bastani et al. (2025) whose large-scale field experiment demonstrates that unguarded AI tutoring improves practice performance while causing underperformance when AI is removed, with the harm eliminated by replacing complete answers with teacher-designed hints. The finding provides the strongest peer-reviewed causal evidence that friction before output is the decisive design variable in AI-assisted learning, confirming cognitive tribology's first design principle at experimental scale. A corresponding sentence is added to Section IV, Principle 1. Section 5.5 and Section VIII add an update from Tao (2026) and corroborating coverage (Padavic-Callaghan, 2026) in which he described a crisis in the foundations of mathematical values and practices and introduced an "AI Capability Conjecture." This is framed as extending the paper's supplement/substitute distinction, now applied at the level of field governance rather than individual practice. Section 5.7 adds: Evans et al. (2026) demonstrating that the improbability of new combinations predicts up to half the likelihood of outsized scientific impact and that most breakthroughs occur when problems in one field are solved by researchers from a distant other, providing the strongest quantitative evidence yet that AI systems converging toward statistically probable outputs systematically reduce the improbability that drives breakthrough discovery. Boussioux et al. (2026) synthesizing four studies across creative domains demonstrating that AI assistance improves individual output quality but consistently reduces collective diversity, with the greatest homogenization occurring when AI enters at the idea generation stage and variety preserved when humans lead early ideation. The finding empirically confirms both the serendipity loss argument and the friction-before-output principle. A corresponding sentence is added to Section IV, Principle 1, noting that keeping humans in charge of early ideation preserves significantly more collective creative diversity than AI-first workflows across four creative domains, providing empirical confirmation that friction before output is a measurable determinant of collective creative outcome rather than merely a theoretical design preference. Section 5.9 adds Fields and Levin (2026) who argue using the Free Energy Principle and quantum information theory that cognitive offloading is not a human-specific tendency but a universal property of all persistent information-processing systems—a theoretical grounding that makes cognitive tribology's design question more urgent rather than less: if offloading is physically inevitable, the critical challenge is calibrating it so that productive cognitive engagement is preserved alongside efficiency gains. Section 5.10 adds Dong et al.’s (2026) systematic review of 152 papers across four decades of AI deployment introducing the Collaboration Convergence Framework. The paper documents a jangle fallacy in which automation complacency, algorithm appreciation, and over-reliance are repeatedly relabeled across AI generations without recognizing their common construct, precisely the unification cognitive tribology's taxonomy provides. Two novel Generation 3 phenomena identified—epistemia and motivational withdrawal—map directly onto phantom friction and the reversibility standard's failure mode respectively. Section 6.4 adds Brassey et al.’s (2026) McKinsey practitioner framework proposing five brain capital design principles for AI-era organizations, independently citing the same Kosmyna et al. and Gerlich evidence base cognitive tribology draws on. No substantive changes to the argument, framework, figures or conclusions. Version 1.2—June 2026 This version incorporates additions published after the original submission date. Section 5.1 adds clinical evidence from Budzyń et al. (2025) documenting a 6.0 percentage point decline in adenoma detection rates among experienced endoscopists performing unassisted colonoscopies following the introduction of AI assistance, described by the authors as the first real-world evidence of automation-induced deskilling linked to patient outcomes, and the most direct clinical confirmation yet of the reversibility standard's failure mode. Also, a Section 5.1 footnote adds Winthrop (2026), who independently coins the term "cognitive stunting" for the foreclosure phenomenon in children and calls for the development of metrics to track children's healthy cognitive development in relation to AI use, extending the atrophy/foreclosure distinction into policy terms. Section 5.1 includes a perspective from Ke et al. (2026) introducing never-skilling, trainees failing to develop foundational clinical reasoning through early AI reliance, and mis-skilling, internalizing AI errors as fact, as constructs that map directly onto cognitive tribology's foreclosure argument and phantom friction construct respectively. Section 5.2 adds Lodge and Loble (2026), who distinguish beneficial from detrimental cognitive offloading in education and find that the risk of foreclosure—rather than mere atrophy—is most acute for novice learners building foundational knowledge, providing policy-level confirmation of the atrophy/foreclosure distinction and the expertise gradient. Section 5.2 also adds Sakai (2026), whose fMRI evidence shows that reading on paper produces more efficient linguistic and narrative-structural integration than reading on digital devices, with digital readers requiring greater compensatory frontal activation to achieve comparable comprehension, a neurological parallel to the desirable difficulties effect. Section 5.3 adds findings from Hackenburg et al. (2026) demonstrating across four preregistered experiments with 6,923 participants that frontier AI systems reliably out-persuade expert humans, including world championship debaters, with the advantage stemming from information volume rather than rhetorical insight. The finding extends the social phantom friction argument from affirmation to active persuasion, establishing epistemic friction as a cognitive safety requirement rather than merely a design preference. Section 5.6 adds Mitchell (2026), who characterizes "jagged intelligence" as the defining architectural feature of current AI systems—an uneven capability landscape without reliable internal signals of failure—establishing that epistemic friction is structurally necessary rather than merely useful. Section 5.7 adds Moon et al. (2025), who introduce a "diversity growth rate" metric across an analysis of 2,200 admissions essays and demonstrate that AI-generated essays contribute fewer novel ideas than human-written ones, with the gap widening as more essays are included, providing empirical confirmation of collective serendipity loss through homogenization. Section 5.7 also adds Vatansever (2026) proposing mental navigation, the brain's spontaneous traversal of conceptual space through the default mode network, as the default computational mode underlying creativity and discovery, providing the neural mechanism through which frictionless AI suppresses the conditions for serendipitous connection. No substantive changes to the argument, framework, figures or conclusions. Version 1.1—May 2026 This version incorporates additions published after the original submission date. Section 5.1 adds neuroplasticity evidence from Rossi et al. (2026) establishing that passive AI reliance weakens activity-dependent brain plasticity while active co-creation sustains it, providing a neurobiological substrate for the atrophy mechanism. Section 5.5 adds empirical evidence from Chakrabarty and Dhillon (2026) docum