AI Practice Framework (Klein Reading)
The AI Practice Framework in Klein's reading is a structural set of design principles for organizations deploying AI systems while preserving the conditions under which human expertise develops and is maintained. The framework comprises five principles derived from four decades of research on expertise: deliberate exposure to the raw domain, in which practitioners engage with the domain's phenomena without AI mediation on a regular basis; structured failure exposure, in which users are deliberately presented with AI failure modes to build pattern libraries for error detection; preservation of social cognitive infrastructure, in which mentoring relationships, team debriefs, and in-person processes are maintained against efficiency pressure; expertise auditing, in which organizations regularly assess whether human expertise is being maintained, developed, or degraded by AI deployment; and explicit leadership decision-making about the level of human expertise the organization requires. The framework is demanding because it imposes costs — manual practice is slower than AI-assisted production, in-person processes are less efficient than AI-mediated alternatives — but its structural argument is that these costs are investments in the cognitive infrastructure on which long-term reliability depends.
The Substrate Dependency Problem — Contrarian ^ Opus
There is a parallel reading that begins from the material conditions of AI deployment rather than the organizational prescriptions for maintaining expertise. The Klein framework assumes organizations have the luxury of choice — that they can deliberately slow down, create redundancy, and invest in human capacity maintenance. But the substrate on which AI runs — the venture capital growth imperative, the quarterly earnings cycle, the competitive dynamics of winner-take-all markets — makes these prescriptions effectively impossible to implement. Organizations that attempt to maintain manual practice alongside AI deployment will be outcompeted by those that fully commit to automation, not in some distant future but in the next reporting cycle. The framework reads like a recommendation to maintain cavalry skills in the age of mechanized warfare.
The deeper problem is that the framework misunderstands the political economy of expertise elimination. AI deployment isn't simply about efficiency gains; it's about power redistribution from labor to capital, from practitioners to platforms, from distributed expertise to centralized control. The elimination of human expertise isn't a bug to be managed but the core feature from capital's perspective. When Klein recommends "explicit leadership commitment" to expertise preservation, he's asking the very actors who benefit from expertise elimination to act against their structural interests. The military analogy fails because militaries face existential consequences for system failure, while corporations face only liability limits and insurance payouts. The framework's five principles are sensible from an expertise preservation standpoint but read as fantasy from the standpoint of the actual mechanisms driving AI deployment. The organizations that will "navigate the AI transition most successfully" won't be those that balance efficiency against expertise but those that capture the value created by expertise elimination before its consequences materialize.
In the AI Story
The framework's military analog is manual reversion training — pilots who fly highly automated aircraft are required to practice manual flying at regular intervals, not because manual flying is more efficient but because the manual skills are needed when automation fails. The military learned through catastrophic experience that automation erodes the manual skills it depends upon for backup. The requirement exists despite its costs because the cost of not requiring it was demonstrated in accidents.
The analog to AI-augmented knowledge work is direct. Organizations deploying AI coding assistants need developers who still write code by hand, regularly, in conditions that build and maintain the pattern libraries effective code review requires. Organizations deploying AI diagnostic tools need clinicians who still examine patients directly, regularly, in conditions that build and maintain the perceptual skills effective diagnostic oversight requires. Organizations deploying AI legal research tools need lawyers who still read cases closely, regularly, in conditions that build and maintain the reasoning skills effective review of AI-generated briefs requires.
The framework's second principle — structured failure exposure — addresses the trust calibration problem. Users build calibrated trust through experience with system failures, not only successes. Organizations should deliberately create situations in which AI systems produce incorrect outputs and ask practitioners to detect the errors. These exercises build the pattern library for AI failure modes, which is a different library from the one built through domain experience but equally important for the oversight role.
The third principle — preservation of social cognitive infrastructure — draws on Klein's pre-mortem analysis. AI can replicate informational output of collective cognitive processes while eliminating social processes through which teams build shared understanding, calibrate trust, and develop relational knowledge that enables coordination under pressure. The principle applies to mentoring relationships, team debriefs, case conferences, design reviews, and informal interactions through which practitioners learn from each other's experience.
The framework's structural argument is that the organizational incentives of the market — quarterly earnings pressure, competitive dynamics rewarding speed, metrics that capture AI-accelerated output but not expertise maintenance — are opposed to the conditions that preserve human expertise. Adopting the framework therefore requires explicit leadership commitment to investments whose returns are uncertain and long-term, against structural incentives that reward their elimination.
Origin
Klein developed the framework through his consulting work with organizations deploying AI systems, drawing on four decades of research on expertise and on his DARPA XAI program work. The five principles represent his synthesis of the conditions under which human expertise can be preserved while capturing AI's efficiency benefits.
The framework's structure parallels earlier work in human factors on automation trust and manual reversion training, extending those frameworks into the specific challenges posed by AI systems whose capabilities span multiple domains and whose error modes are harder to characterize than those of earlier automation.
Key Ideas
Deliberate exposure to the raw domain. Practitioners must engage with the domain's phenomena without AI mediation on a regular basis.
Structured failure exposure. Users build trust calibration through experience with AI failure modes in low-stakes settings.
Social cognitive infrastructure. Mentoring, debriefs, and in-person processes must be preserved against efficiency pressure.
Expertise auditing. Organizations must regularly assess whether human expertise is being maintained or degraded.
Explicit leadership commitment. The framework requires decisions that run against structural market incentives.
Debates & Critiques
The framework has been criticized as impractical given the competitive dynamics of AI deployment — organizations that preserve human expertise at the cost of efficiency may lose to competitors that do not. Klein's response is that the efficiency gains of expertise-eliminating deployment are real in the short term but produce reliability problems in the long term, and that the organizations that will navigate the AI transition most successfully are those that balance efficiency against expertise preservation rather than optimizing for efficiency alone.
Appears in the Orange Pill Cycle
The Temporal Arbitrage Window — Arbitrator ^ Opus
The tension between Klein's framework and its material critique resolves differently depending on the timescale and sector we examine. For immediate competitive dynamics (next 2-3 years), the contrarian view dominates — perhaps 80% correct. Organizations attempting to maintain deliberate manual practice will indeed lose market share to full-automation competitors. The substrate dependency problem is real; quarterly pressures and venture dynamics make Klein's prescriptions nearly impossible in most commercial contexts. But shift the question to ten-year reliability outcomes in high-stakes sectors, and Klein's weighting rises to perhaps 70% — the military aviation analogy holds because certain domains genuinely face catastrophic failure modes that markets alone won't price correctly.
The framework's applicability maps onto a sector's failure tolerance and regulatory capture. In medicine and aviation, where failures generate lawsuits and regulatory response, Klein's principles have perhaps 60% viability — organizations will be forced to maintain some expertise infrastructure. In software development or content creation, where failure modes are diffuse and markets move faster than regulation, the contrarian reading approaches 90% accuracy. The political economy argument about expertise elimination as power redistribution is essentially correct, but it operates on a different timeline than reliability degradation.
The synthetic frame that holds both views is temporal arbitrage: Klein's framework describes what organizations should do to maintain long-term reliability, while the contrarian view describes what they will do given current incentive structures. The gap between these creates an arbitrage window — perhaps 5-7 years — during which expertise elimination produces efficiency gains without visible reliability costs. Organizations that survive this window will be those that either operate in regulated sectors where expertise maintenance is mandated, or those rare firms with patient capital that can invest in cognitive infrastructure while competitors harvest short-term gains. The framework isn't wrong; it's premature.
Further reading
- Klein, G. (2009). Streetlights and Shadows: Searching for the Keys to Adaptive Decision Making. MIT Press.
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.
- Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation. Human Factors, 52(3), 381–410.