Amerika

The West has a learning curve dilemma, and the name is damning: single-loop learning masquerading as double-loop learning.

Chris Argyris and Donald Schön defined “learning” as the detection and correction of error. Single-loop learning corrects the error within the existing governing variables, like a thermostat that learns when it’s too hot or too cold and turns the heat on or off.

Double-loop learning questions the governing variables themselves, modifying the organization’s underlying norms, policies, and objectives. It is metacognition about the variables, or a study of how we know how to think.

The argument is that datafication (digitalization of subject matter) and pidgin language have produced institutions that look like they are learning at extraordinary speed while being structurally incapable of the only learning that matters.

Datafication is, almost by definition, a single-loop technology. It optimizes within given parameters. It does not ask whether the parameters are correct. This is the failure mode. Single-loop learning is present when goals, values, frameworks, and to a significant extent strategies are taken for granted, with emphasis on techniques and making techniques more efficient.

Universities datafied their disciplines and thereby made every discipline spectacularly efficient at technique (only). Finance optimizes flows. Medicine optimizes biomarkers. Engineering optimizes throughput. None of them interrogate the objective function.

Argyris found that just about all participants in his studies operated from Model I theories-in-use, shaped by an implicit disposition to winning and to avoid embarrassment. In other words, means-over-ends thinking in the typical model of committees, unions, and social groups: accept everyone, avoid controversy.

The UN and modern governments are Model I institutions at civilizational scale. Datafication doesn’t break Model I; it arms it, because a dashboard lets an institution appear to be generating valid information while (actually) generating only confirmatory information.

Hofstede’s actual point is precise and citable: discourse in trade languages limits communication to those issues for which these simplified languages have words, and it misses the idiosyncrasies of local languages, which are the very essence of culture.

That is not a culture-war observation. It is a capability constraint the West suffers from (Germany is declaring war on Russia because of pidgin language). If an institution’s operational language cannot express a concept, the institution cannot act on it. Sanctions policy conducted in the vocabulary of “rules-based order” literally lacks the lexical resources to represent the second-order consequences for a mid-tier economy’s fertilizer imports (for example).

And Hofstede’s sharper line: “having English, the world trade language, as one’s mother tongue is a liability, not an asset, for truly communicating with other cultures.”

For international institutions, this is a structural handicap baked into the working-language regime.

The Derridean move discussed with regard to pidgin language, is worth consideration, because it converges with Argyris from an entirely different direction; “Derrida demonstrated how all assumptions on human life and behavior are grounded in our use of language.”

Argyris says the same thing operationally: our governing variables are tacit, and we cannot interrogate what we cannot articulate. Pidgin guarantees they stay tacit.

A 2026 study of nine intelligence failures (Pearl Harbor, Yom Kippur, 9/11, the early Russia-Ukraine phase, the 2023 Hamas deception) found that digital-intelligent technical factors do not directly cause cognitive biases but amplify them through specific mechanisms: cognitive resource scarcity, cognitive anchor solidification, increased time pressure, and reduced cognitive processing depth.

Read that again. Rapid information dissemination increases decision-making time pressure, and analysts unable to conduct in-depth analysis and verification can only make decisions based on intuitive judgment and past experience.

That is the thesis, empirically demonstrated. The 24-hour news cycle doesn’t inform decision-makers; it degrades the depth at which they can process, then rewards them for deciding anyway (Trump).

And the algorithm-dependence finding is the datafication half of the learning dilemma equation: algorithms generate initial results that become cognitive anchors, and because of trust in algorithms, analysts are often unwilling to question and revise those initial results, leading to solidification of cognitive anchors.

In the 2023 Hamas case, Israeli intelligence over-relied on AI system analysis and technical reconnaissance data while ignoring human intelligence, because these technical results were considered to have “objective authority.”

“Intentionally ignoring the real meaning of the data” is not a moral failure. It is a documented cognitive mechanism. The data arrives pre-interpreted, and pre-interpretation is invisible to the next manager or next generation.

This issue should be isolated because the literature is more sympathetic to its critique than is admitted by practitioners, which include investors, executive managers, and politicians. One documented criticism of scenario methods is that they may lack political credibility

More tellingly: use of particular foresight tools can have predictable effects on strategy making, providing positive changes in mental models but, more negatively, can also serve to narrow and shape managers’ anticipations of the future.

Scenario planning was designed to expand the space of thinkable futures. Institutionalized, it collapses that space into a menu of pre-framed options, each already morally legible. Once every scenario is pre-sorted into “moral high ground” versus “not,” you have eliminated the possibility of double-loop learning by construction, because the framing itself is no longer in play.

The framing is the governing variable. And it is precisely what protects pidgin from scrutiny. This reaches critical stages in international politics.

Groupthink describes the phenomenon by which existing information relevant to a decision is not processed by group members in a rational and systematic way but is instead biased in the direction of a preferred decision alternative within the group. Large-scale, high-stakes groups such as the UN are explicitly cited as exemplars of groups whose decisions have widespread consequences and where these dynamics apply.

Now add the incentive layer. An unreasonable incentive mechanism will encourage analysts to cater to superiors’ intentions or pursue short-term interests, leading to aggravation of cognitive biases. In foreign ministries, the incentive is to not be the person who delayed the response. So the institution is optimized for speed of alignment, not depth of understanding.

And the acceleration is now structural, not cultural. Governments are confronting more decisions to make in less time, and social acceleration has led to the disintegration of both individual personality and political organization, making it difficult to orient around time-resistant priorities.

Meanwhile, across agencies, AI, automation, and data platforms are compressing decision cycles from days to minutes, even seconds, yet decision quality is not advancing at the same rate as decision speed. Speed without structure, is not progress; it is risk at scale.

To put a finer emphasis on the different learning speeds of computers and humans, as follows: The conclusion is that the real learning curve depends on human ability to learn, measured in decades rather than months. This direction of thought can be sharpened, because the framing risks a defensive reading: “computers are fast, humans are slow, therefore go slow.” The better formulation is Argyris’s own: the dilemma is structural, not temporal.

The underlying theory is that the reasoning processes employed by individuals in organizations inhibit the exchange of relevant information in ways that make double-loop learning difficult, and all but impossible in situations in which much is at stake.

The authors add: this creates a dilemma as these are the very organizational situations in which double-loop learning is most needed.

So it isn’t that humans are slow. It’s that the more consequential the decision, the more the institution’s defenses against embarrassment suppress the learning required to make it well. Sanctioning a sovereign state, debanking an individual, restructuring an organization by firing thousands: these are maximum-stakes, maximum-defensiveness decisions. Which means they are the decisions least likely to be learned from.

The 2004 Senate report on prewar intelligence found that established mechanisms to challenge inherent assumptions and groupthink “were not utilised,” and managers did not counsel analysts who lost their objectivity. The mechanisms existed. They weren’t used. That’s the whole story.

Universities advanced their disciplines by datafying them. Datafication is single-loop by construction. Pidgin language is the governance technology that keeps the governing variables unexamined. The result is a class of leaders trained to be extremely fast at correcting errors inside a frame they were never taught to interrogate, in institutions whose incentive structures punish anyone who tries.

The learning curve didn’t shorten. The measurable learning curve and the actual learning curve decoupled. Everything you can dashboard is single-loop. Everything that matters is what the dashboard can’t see.

The one genuinely hopeful finding in the literature: unlike earlier experiential learning models where you had to make a mistake and reflect on it, Argyris and Schön’s framework makes it possible to learn by critically reflecting on the theory-in-action itself, without going through the entire cycle of failure first.

Double-loop learning doesn’t strictly require catastrophe. It requires only that someone be permitted to ask what the thermostat is for.

Tags: double-loop learning, management theory, pidgin language, single-loop learning