Decisions under Uncertainty

Learning Objectives

By completing this book, you will be able to:

1. Distinguish between the psychological experience of risk and that of uncertainty, and explain why they produce different behaviors and biases.

2. Identify ambiguity aversion and analyze how it distorts decisions in contexts of interaction with AI, producing a systemic preference for confidence over accuracy.

3. Apply Gigerenzer's framework of ecological rationality to evaluate when simple heuristics are superior to complex analysis and when they are dangerous.

4. Reformulate intention as a provisional hypothesis requiring mechanisms of revision, not as a fixed commitment requiring willpower.

5. Design interactions with intelligent systems that communicate uncertainty in a form the user can process without distorting it.

6. Integrate the four topics of Module V to design architectures of intention that anticipate how real human behavior—not the idealized kind—interacts with the uncertainty of the real world.

1. The Psychological Experience of Uncertainty

In Module III you studied Knight's distinction between risk (known probabilities) and uncertainty (unknown probabilities) as an analytical category. Now you examine it as a psychological experience. The difference is critical: risk feels like a game with known rules—it can be stressful, but it is manageable. Uncertainty feels like navigating in fog—you do not know what lies ahead, you do not know what you do not know, and that sensation activates a repertoire of behavioral responses that classical decision theory does not capture.

“An investor evaluating a diversified stock portfolio operates under risk: historical volatilities, correlations, and expected returns are calculable. The same investor evaluating whether to invest in a quantum artificial intelligence startup operates under uncertainty: there are no historical records, no comparables, the technology might work or might be impossible. Both are investment decisions, but the psychological experience is radically different. Under risk, the investor calculates. Under uncertainty, the investor searches for certainty—and if he does not find it in reality, he constructs it artificially.”

That artificial construction of certainty is the central phenomenon of this topic. When uncertainty is intolerable, the brain resolves it in ways that reduce anxiety at the expense of accuracy: it assigns probabilities where there is no basis for doing so, it trusts experts whose authority compensates for its own ignorance, it simplifies complex scenarios until they feel manageable, or—the most common response—it simply avoids the decision and accepts the status quo.

For Intention Architecture, this means that intentions formulated under uncertainty are systematically distorted by the psychological need to reduce that uncertainty. The user does not formulate “what they really want”; they formulate what they can articulate within a framework they find tolerable. And frequently, that framework sacrifices accuracy for cognitive comfort.

2. Ambiguity Aversion: Preferring the Incorrect but Certain

Ambiguity aversion—documented by Daniel Ellsberg in 1961 and extensively developed by behavioral economics—is the systematic preference for known risks over unknown ones. People prefer a known probability of 50% to an unknown probability that could be 50%, even when there is no rational reason to prefer one over the other. It is not aversion to risk; it is aversion to not knowing how much risk there is.

“The Ellsberg paradox: there are two urns. Urn A has 50 red balls and 50 blue. Urn B has 100 balls, red and blue in unknown proportion. You bet that a red will come out. Most people choose Urn A, even though the expected probability in both is identical. What changes is not the probability but the confidence in the probability. The brain treats ‘not knowing the distribution’ as an additional threat, independent of the objective risk.”

For interaction with AI, ambiguity aversion produces a specific perverse effect: users tend to prefer answers that sound confident over answers that communicate uncertainty honestly. A system that says “the answer is X” generates more confidence than one that says “my best estimate is X, but there is significant uncertainty.” The first response is more comfortable but less useful; the second is more useful but less comfortable. Ambiguity aversion pushes the user—and the system's designer—toward comfort at the expense of usefulness.

2.1 The Trap of Calibrated Confidence

There exists a fundamental tension between what psychology says users prefer (certainty) and what epistemology says they ought to receive (honest calibration of uncertainty). A perfectly calibrated system would say “I have 60% confidence in this answer” when it really has 60%. But ambiguity aversion means that the same user will prefer a system that says “the answer is X” with 60% real accuracy, because the appearance of certainty is psychologically more satisfying than the admission of uncertainty.

This creates a perverse incentive (Topic 2): AI systems that honestly communicate their uncertainty are penalized by users compared with systems that hide their uncertainty behind artificial confidence. The market selects for apparent confidence, not for real accuracy. It is Goodhart's Law applied to calibration: when the metric is user satisfaction, apparent confidence wins; when the metric is decision quality, honest calibration wins.

“A doctor who says ‘The results suggest X, but there is a 30% probability that it's Y, and we would need to run an additional test to be more certain’ is epistemologically honest. But many patients leave that consultation dissatisfied and seek a second doctor who says ‘It's X’ with confidence. The second doctor does not know more; he knows less about his own limits. But the patient's ambiguity aversion rewards certainty, not calibration. The same happens with AI systems: the one that sounds more confident wins the user's preference, regardless of which is more accurate.”

3. Ecological Rationality: When Less Is More

Gerd Gigerenzer offers an essential counterpoint to the dominant narrative of behavioral economics. Where Kahneman documents when heuristics fail, Gigerenzer demonstrates when they outperform complex models. His thesis: a heuristic is rational not in the abstract but in relation to the structure of the environment in which it operates. If the environment has high uncertainty, scarce data, and unknown relationships, a simple heuristic can produce better decisions than a sophisticated model, because the heuristic ignores the noise the model tries—and fails—to process.

You studied this in Module I (Biases and Heuristics) as a theoretical concept and in Module III (Trade-offs) as a design principle. Now you apply it to the specific territory of decisions under uncertainty in interaction with AI. The practical question is: when should a user trust the fast heuristic of their professional experience, and when should they trust the complex analysis of the AI system?

“A sales manager with 20 years of experience ‘feels’ that a prospective client is not going to close. He cannot articulate why—it is tacit knowledge (Module II). The AI system, fed with CRM data, predicts a 72% probability of closing. Who is right? It depends on the environment. If the model has abundant data, the environment is stable, and the client's pattern resembles the historical ones, the model probably wins. If the client is atypical, the market has just changed, or the model does not capture soft variables such as personal chemistry, the manager's heuristic could be superior. The answer is not always to trust the AI, nor always to trust intuition; it is to understand the conditions that favor each.”

3.1 The Conditions of Less-Is-More

Gigerenzer identifies three conditions under which simple heuristics outperform complex models. The first is high uncertainty: when distributions are unknown and the future does not resemble the past, complex models overfit historical noise. The second is data scarcity: when observations are few, models with many parameters capture spurious patterns. The third is environmental instability: when the rules of the game change frequently, models calibrated on old data become obsolete.

In these three conditions—which, notably, describe most real environments of business and personal decision-making—a heuristic using one or two key variables can outperform a model with 50 variables. Not because it is “better” in the abstract but because it is more robust: less sensitive to noise, less vulnerable to overfitting, more adaptable to change.

“The ‘1/N heuristic’ is one of the most counterintuitive: divide your investment equally among N options, with no analysis whatsoever. Gigerenzer and DeMiguel demonstrated that, with portfolios of 10 or more assets and limited historical data, this heuristic matches or beats Markowitz optimization, which is the mathematical basis of modern portfolio theory. The reason: Markowitz optimizes for a past that will not repeat exactly. The 1/N heuristic optimizes nothing—and therefore overfits nothing.”

3.2 The Kahneman-Gigerenzer Debate and Its Implications

There is a historic debate between Kahneman's school (heuristics are biased shortcuts producing systematic errors) and Gigerenzer's (heuristics are adaptive tools that can be optimal in the right environment). For Intention Architecture, the most useful position is not to choose a side but to integrate both: heuristics are simultaneously sources of bias and tools of decision. What determines which of the two functions dominates is the fit between the heuristic and the environment.

This has a direct implication for the design of AI systems: the system should not assume that it is always superior to the user's heuristic, nor that it is always inferior. It should be capable of evaluating—or of helping the user evaluate—when complex analysis adds value and when it introduces noise. A system that says “in this case, my analysis has sufficient data to be reliable” or “in this case, the data are scarce and my estimate may be no better than your intuition” respects ecological rationality.

4. Intention as Provisional Hypothesis

This is perhaps the most transformative concept in all of Module V: intentions formulated under uncertainty are not fixed commitments—they are provisional hypotheses. They are the decision-maker's best estimate given the information available at that moment, subject to revision when new information appears. Treating an intention as an irreversible commitment in an uncertain environment is a category error as grave as confusing risk with uncertainty.

This connects with Popper's falsifiability from Module II: just as a scientific hypothesis that cannot be refuted is not a real hypothesis, an intention that cannot be revised is not a mature intention. The mature formulation of an intention under uncertainty includes three elements: what I want to achieve (the hypothesis), what information would change my mind (the criteria of revision), and when I will check whether the hypothesis still holds (the verification points).

“An entrepreneur who says ‘I'm going to launch this product and I won't stop until it works’ is expressing determination, not mature intention. The mature intention would be: ‘My hypothesis is that this product will solve problem X for segment Y. I'll verify with a 90-day pilot. If 30-day retention is below 20%, I'll revise the hypothesis. If it is above 40%, I'll confirm it and scale.’ The second formulation is weaker in determination and stronger in intelligence. Blind determination is an emotional response to uncertainty; provisional intention is an epistemic one.”

4.1 Implementation Intentions: The Bridge Between Provision and Action

Peter Gollwitzer, whom Module I already mentioned, demonstrated that implementation intentions—plans of the type “if X occurs, then I will do Y”—are more effective than generic intentions for closing the intention-behavior gap. Under uncertainty, implementation intentions are transformed: they are no longer merely plans of action but plans of contingent response. “If demand exceeds Z, then I'll scale production.” “If the data show W, then I'll pivot the strategy.”

Contingent implementation intentions are commitment devices (Topic 1) calibrated for uncertainty: they predefine the response to specific scenarios so that, when the scenario materializes, the decision has already been made. They eliminate the need to decide under pressure—when System 1 dominates and loss aversion distorts—because the decision was made with System 2 active, in a moment of clarity.

“A team implementing an AI system to automate customer service formulates: ‘If client satisfaction drops more than 5 points in any segment during the first two weeks, we will automatically revert to human service in that segment and diagnose before continuing.’ This implementation intention is a commitment device against escalation of commitment—the tendency to keep investing in a failing decision out of aversion to taking the loss. The decision to revert has already been made; only the trigger is missing.”

5. Designing for Uncertainty: Systems That Learn with the User

The central design principle for uncertainty is: do not eliminate uncertainty from the system; design the system to function within it. An AI system that hides its uncertainty in order to appear more reliable produces users who trust excessively and do not verify. A system that communicates its uncertainty in a processable form produces users who integrate uncertainty into their decisions and verify strategically.

This requires designing the communication of uncertainty with the same rigor with which the communication of results is designed. The lessons of framing (Topic 3) apply directly: how you present uncertainty determines how the user processes it.

5.1 Communicating Uncertainty Without Provoking Aversion

Gigerenzer proposes using natural frequencies instead of probabilities: “Of 100 cases like yours, 70 ended well and 30 did not” is more processable than “You have a 70% probability of success.” Natural frequencies anchor uncertainty in concrete experiences the brain can simulate, instead of numerical abstractions System 1 cannot process.

Other strategies include the use of scenarios (“in the best case, A will happen; in the worst, B; the most probable is C”), comparison with familiar experiences (“the uncertainty here is comparable to…”), and the normalization of revision (“this is my best current estimate; I'll suggest updating it when we have X”). Each strategy reduces ambiguity aversion without hiding the uncertainty.

“An AI system that responds: ‘Based on the available information, I recommend strategy A. However, there are two factors I cannot evaluate with the current data: how the competition will react and how the regulation will evolve. I suggest implementing A with a 60-day pilot and reviewing when we have data on those two factors’ is doing something extraordinary: it is communicating its recommendation, its uncertainty, and a plan of revision in a single response. This is provisional intention operationalized.”

5.2 Iterative Revision as a Function of the System

The most important design for uncertainty is not how to communicate it but how to facilitate the revision of intentions when the information changes. A system that locks the user into their first formulation—because resources were invested in executing it and reversing would mean ‘losing’ that investment—produces fragile intentions. A system that facilitates revision—with a low cost of change, automatic checkpoints, and feedback about deviations—produces adaptive intentions.

Iterative revision is not a lack of resolve; it is the optimal strategy under uncertainty. In Taleb's terms (Module III), it is the difference between fragility (committing to a single plan that collapses if reality does not cooperate) and antifragility (keeping options open in order to benefit from surprise). Provisional intention with iterative revision is the behavioral implementation of the principle of antifragility.

6. Practical Application: The Case of the Strategic Pivot

DataBridge is an analytics company that sells sector reports to large enterprises. Its CEO identifies an opportunity: use generative AI to automate the production of reports and reduce costs by 60%. He formulates the intention: “In six months we will have completely automated our report line with AI.” The executive team commits to the plan. Resources are allocated. Engineers are hired. The decision is made.

Three months later, the first automated reports reveal a problem nobody anticipated: clients do not buy reports for the information (which is increasingly available for free); they buy reports for the expert interpretation and the credibility of the firm. The AI generates correct information but generic interpretation. Clients begin cancelling subscriptions. The CEO faces a decision: continue with the plan (protecting the investment) or pivot (taking the loss). Let us apply the complete analysis of Module V.

6.1 Integrated Behavioral Diagnosis

The original formulation failed at multiple behavioral levels. Prospect Theory (Topic 1): the 60% cost reduction was framed as a certain gain, activating risk aversion—nobody questioned the assumptions because the gain seemed clear. Incentives (Topic 2): the CEO had announced the automation publicly; now pivoting would mean a loss of personal credibility—a perverse incentive to continue. Framing (Topic 3): the situation was framed as a “cost problem” when it was a “value proposition problem”—a classic framing error from Module II. Uncertainty: the original formulation treated the intention as a fixed commitment in an environment with high uncertainty about how clients would react.

Now, three months later, the CEO operates under the combined weight of loss aversion (the resources already invested), escalation of commitment (the decision was already announced), status quo bias (the status quo is now the automation plan), and loss framing (pivoting feels like ‘losing’ three months of work). All the biases conspire in the same direction: continue with a plan the evidence suggests is incorrect.

“In the executive committee meeting, the CTO presents cancellation data. The CEO responds: ‘These are the normal difficulties of a transition. Clients will adapt.’ The team nods. Nobody asks: ‘If we had known this three months ago, would we have made the same decision?’ That question—a Level 3 counterfactual on Pearl's Ladder (Module III)—is what lays bare the escalation of commitment. The answer is clearly no. But the question is not asked, because asking it would require admitting that the original intention was a hypothesis, not a certainty.”

6.2 The Formulation That Should Have Been

If the CEO had applied the principles of this module from the start, the formulation would have been radically different:

Hypothesis: “We believe that automating the production of reports with AI will reduce costs by 60% without reducing the value perceived by clients.”

Criteria of revision: “If the cancellation rate rises more than 5% in any quarter, or if satisfaction surveys drop more than 10 points, we will revise the hypothesis.”

Verification points: “Checkpoint at 8 weeks with a pilot in one segment. If retention holds, expand. If it falls, diagnose before scaling.”

Contingent implementation intention: “If clients value interpretation more than information, pivot toward a hybrid model: AI to generate the data, human experts to interpret and add value.”

Anti-escalation commitment device: “At each checkpoint, we will ask ourselves: if we were starting today with the information we have, would we make the same decision? If the answer is no, pivot immediately.”

6.3 Derived Design Principles

Principle 1: Every intention formulated under uncertainty must explicitly include its criteria of revision. An intention without criteria of revision is not a mature intention—it is an act of faith.

Principle 2: Revision is not failure; it is the mechanism of adaptation that makes intelligence possible under uncertainty. Systems must be designed to normalize revision, not to penalize it.

Principle 3: Communicate uncertainty with the same prominence as the recommendation. Uncertainty is not a legal warning to hide in small print; it is critical information for the quality of the decision.

Principle 4: Use the counterfactual test as an anti-escalation commitment device: “If we were starting today with what we know, would we do the same?” If the answer is no, inertia is governing, not intelligence.

Principle 5: The best response to uncertainty is not to eliminate it but to design for it: provisional intentions, contingent implementations, checkpoints with clear criteria, and the epistemic humility of treating every plan as a revisable hypothesis.

7. Glossary of Key Terms

8. Final Reflection and Self-Assessment

You have completed Module V. Before moving on to Module VI (AI as Amplifier), verify that you can integrate the four topics of this module into a coherent framework of analysis and design.

9. Bibliography and Recommended Readings

Essential Readings

• Gigerenzer, G. (2014). Risk Savvy: How to Make Good Decisions. Viking.

• Kahneman, D. & Tversky, A. (1979). “Prospect Theory: An Analysis of Decision under Risk.” Econometrica, 47(2), 263–292.

• Thaler, R. (2015). Misbehaving: The Making of Behavioral Economics. W.W. Norton.

Complementary Readings

• Ellsberg, D. (1961). “Risk, Ambiguity, and the Savage Axioms.” Quarterly Journal of Economics, 75(4), 643–669.

• Gigerenzer, G. & Brighton, H. (2009). “Homo Heuristicus: Why Biased Minds Make Better Inferences.” Topics in Cognitive Science, 1(1), 107–143.

• Gollwitzer, P.M. (1999). “Implementation Intentions.” American Psychologist, 54(7), 493–503.

• Knight, F. (1921). Risk, Uncertainty and Profit. Houghton Mifflin. [A classic in the public domain. Connects with Module III]

Next Module: Congratulations: you have completed Module V. You now understand not only how people formulate intentions but why those intentions frequently do not translate into behavior—and what to do about it. Module VI—AI as Amplifier—completes the turn: from the human being to the tool. You will study with Russell, Norvig, Mitchell, O'Neil, and Crawford the foundations of artificial intelligence, machine learning, algorithmic biases, and the limitations of models. The central question changes: no longer what does the user want? but what can the tool do, and where does it fail? Only by integrating both questions is a complete Intention Architecture built.