Opaque, widely available tools can trigger permanent human skill loss by driving users into a low-competence equilibrium; initial skill and tool transparency set thresholds, so identical present access can produce opposite long-run outcomes depending on practice history.

Main Finding

Krakauer models a human agent, a cognitive tool, and a task as a coupled dynamical system and shows that tool use can be either complementary (augments persistent competence) or competitive (erodes competence and creates lasting dependence). The coupled competence–reliance dynamics are nonlinear and generically bistable: above a critical tool availability a+ the competent attractor disappears and competence collapses (hysteretically) to a low dependent floor; lowering availability only restores competence below a much lower fold a−. Whether an individual (or population) ends up competent or dependent depends on prior learning and the tool’s transparency (τ), and in high goal-uncertainty settings a tool whose policy/model is too large to internalize can permanently capture agency itself.

Key Points

• Minimal state variables:
  • Competence ι ∈ [0,1]: fidelity of the user’s internal model — performance retained when the tool is withdrawn.
  • Reliance x ∈ [0,1]: propensity to offload control steps to the tool.
  • Tool availability a ∈ [0,1] and tool transparency τ = ιmax (the fraction of tool function a user can reverse-engineer).
• Core mean-field dynamics (qualitative form):
  • Competence grows with self-executed steps and decays with offloaded steps: ˙ι = r(1 − x)(ιmax − ι) − µ x ι
  • Reliance increases with availability and low competence, and decays when the tool is unavailable: ˙x = β a σ(ιc − ι) (1 − x) − [α(1 − a) + ϵ] x
  • σ is a steep logistic that creates thresholded offloading when competence falls below ιc.
• Bistability and hysteresis:
  • For intermediate ranges of a there are two stable equilibria (competent and dependent) separated by an unstable separatrix. Crossing a+ (increasing availability) can irreversibly destroy competence unless availability is later reduced below a−.
  • Prior competence ι0 raises the a that a person can tolerate before collapse: a*(ι0, τ) is non-decreasing in ι0.
• Role of transparency:
  • High τ (transparent tools like maps, abacus) allow internalization of method; complements human competence.
  • Low τ (opaque tools like GPS or black-box models used uncritically) cap attainable competence and push towards dependence.
• Scale / internalizability gate:
  • If a tool’s model size m* exceeds an agent’s acquisition budget B (or is otherwise non-internalizable), an effective ceiling falls and even a transparent-seeming tool cannot be internalized; large models can therefore be non-transferable.
• Agentic reversal:
  • Under substantial goal uncertainty and when the tool has superior capacity (memory, operations, policy capacity), the tool can become the policy-holder (agent) and humans become instruments. Reclaiming agency requires internalizing a model that may be economically or physically infeasible.
• Population dynamics and lock-in:
  • Similar dynamics at the population scale yield two absorbing states (near-universal outsourcing vs near-universal competence) with a tipping point; early history and adoption costs create path dependence and possible collective lock-in.
• Empirical alignment:
  • Evidence reviewed includes skill decay meta-analysis (189 datapoints, 53 studies), GPS vs map navigation, abacus mental calculation transfer, aviation dependence patterns, and early observations on language models (LMs). Some model elements are strongly supported (decay, offloading effects), others remain predictions to test (within-subject hysteresis experiments, explicit thresholds).

Data & Methods

• Modeling approach:
  • Agent, tool, and task represented as coupled open transducers with internal registers: memory M, operations O, and policy Π (policy held initially by the human agent).
  • Microscopic behavior: each control step is a Bernoulli choice to offload with probability x; mean-field ODEs above derived as limit of many steps.
  • Parameters: r (acquisition rate), µ (atrophy/decay), β (gain to offloading), α, ϵ (abandonment/base disengagement), ιc (competence threshold), τ = ιmax (transparency ceiling), a (availability), and a logistic σ with slope k.
  • Scale gate for internalizability: ceiling multiplier g(m/B) ≈ 1/(1 + m/B), making large models effectively non-transferable.
• Empirical evidence:
  • Meta-analysis of skill decay finds cognitive skills decay faster than motor skills (half-lives ~48 vs ~106 days).
  • Navigation literature: map users retain spatial memory; GPS users show poorer spatial memory.
  • Abacus and mental-calculation literature: evidence of recoding and persistent competence when technique is internalized.
  • Aviation and cognitive offloading studies suggest dependence cycles and offloading carry-over.
  • Language models: dual modes—when outputs are reconstructed/verified they can augment; when accepted uncritically they act like opaque outsourcing and can diminish competence or displace agency.
• Predictions left to be tested:
  • Experimental demonstration of hysteresis by raising and then lowering availability within subjects.
  • Quantitative estimates of τ, m*, B for contemporary LLMs and how these map to internalizability thresholds.

Implications for AI Economics

• Design incentives and product architecture:
  • Favor transparency and verifiable outputs to make tools complementary (raise τ). Provide interfaces that support reconstruction, explanation, and stepwise verification to increase retained competence.
  • Build tooling that supports human-in-the-loop verification workflows to turn LLMs into coaches/augmenting artifacts rather than opaque authorities.
• Training and human capital policy:
  • Invest in “analog education” (prior competence) before broad rollout of high-availability tools. Subsidize or mandate training windows to shift populations into the competent basin and avoid hysteretic collapse.
  • Include periodic disconnection/verification drills to prevent atrophy and maintain retrieval strength.
• Market structure, pricing, and adoption:
  • Recognize path dependence and first-mover effects: early diffusion can create lock-in to dependent equilibria. Subsidies/taxes on tools or complementary training can steer equilibria.
  • Consider acquisition cost (B) and model scale (m*) when evaluating social value: very large models that are non-internalizable create ongoing externalities (dependence, agency loss) not captured by private valuations.
• Labor and regulation:
  • Policies that simply reduce task cost via automation may lower human competence and create long-run fragility; assess long-term resilience costs when valuing automation.
  • Regulatory standards could require minimum transparency, auditability, or verifiability for tools deployed in safety-critical or public tasks to avoid irreversible competence loss or agency transfer.
• Competition and antitrust:
  • Concentration of large, opaque models increases systemic dependence risk; competition policy should consider the social value of multiple independent competence-preserving tools and interoperability standards.
• Research & metrics:
  • Economists and policymakers should measure τ, a, and ι empirically for critical tools (e.g., LLMs in education/medicine), and run controlled experiments on hysteresis and recovery costs to quantify social welfare trade-offs.
  • Incorporate the cost of competence restoration (hysteresis) and risk of agency capture in benefit–cost analyses for AI deployments.

Overall, the paper reframes tool/AI evaluation from one-off performance gains to a dynamical, path-dependent account of competence, reliance, and agency — with clear economic trade-offs around transparency, training, adoption timing, and irreversibility that should inform AI design, deployment, and policy.