Why Uncertainty Changes Decision-Making - Sustainable Catalyst | Ethical Systems for Sustainable Strategy

Last Updated June 4, 2026

Uncertainty changes decision-making because it disrupts the assumptions of predictability, stable probabilities, and fully specified outcomes on which classical models of rational choice depend. When the future cannot be known in advance, when relevant variables are only partly understood, and when system behavior is nonlinear or path-dependent, decision-making becomes less a matter of precise optimization and more a matter of structured judgment, adaptation, and resilience.

This article is part of the Decision Science knowledge series.

In idealized decision problems, alternatives are known, consequences are defined, probabilities can be estimated, and preferences can be applied consistently across options. Under such conditions, tools from decision theory—including expected value, expected utility, and probabilistic optimization—provide a powerful framework for rational choice. In many real-world environments, however, these conditions do not hold. Decision-makers often face incomplete information, shifting institutions, contested models, and outcomes shaped by complex interactions rather than isolated variables.

Uncertainty therefore changes more than the difficulty of decision-making. It changes the structure of the problem itself. Under uncertainty, decision-makers must often determine which variables matter, which models are credible, what counts as acceptable risk, how to compare irreversible consequences, and when to preserve flexibility rather than commit to a single forecast. Frank Knight’s classic distinction between measurable risk and unmeasurable uncertainty remains foundational for precisely this reason: some decisions can be modeled probabilistically, while others cannot be reduced to known distributions at all. Later work on bounded rationality, heuristics and biases, ambiguity, and robust decision-making has extended this insight by showing that uncertainty affects both how we think and how we should design decisions.

Painterly editorial illustration of a thoughtful figure facing branching paths, fog, storm clouds, risk markers, weighted choices, feedback loops, and scattered evidence under uncertain conditions. — Uncertainty changes decision-making by reshaping risk perception, confidence, timing, tradeoffs, and the need for adaptive judgment.

Why uncertainty matters for decision-making

Uncertainty matters because it changes what it means to decide well. In a predictable environment, a “good decision” can often be defined as the choice that maximizes expected value or expected utility given reliable probabilities and clearly specified outcomes. Under uncertainty, however, the decision-maker may not know whether the probabilities are correct, whether the most important outcomes have been included, or whether the model itself is stable enough to justify optimization.

This means decision quality cannot be judged only by the apparent precision of a forecast. In uncertain settings, the question shifts from “What is the best option under our current estimate?” to “Which strategy is defensible if our estimate is wrong?” That shift is one of the core moves in decision science. It redirects attention from narrow precision toward resilience, flexibility, reversibility, and institutional learning.

This is especially important in domains such as infrastructure, finance, public policy, climate adaptation, health systems, and geopolitical strategy, where consequences are large, feedback effects are strong, and the future is only partially knowable. In such domains, decision-making becomes a matter of disciplined reasoning under incomplete knowledge rather than of perfect prediction. Knight’s distinction between risk and uncertainty, Simon’s theory of bounded rationality, and RAND’s later work on robust decision-making each reflect this broader transformation in how rationality is understood.

Risk vs. uncertainty

A foundational distinction in decision science is the difference between risk and uncertainty. Under risk, probabilities can be estimated with some confidence, allowing outcomes to be evaluated using probabilistic methods. Under uncertainty, either probabilities are unknown, contested, or unstable, or the relevant outcome space is itself only partly understood.

Frank Knight’s Risk, Uncertainty, and Profit remains the classic statement of this distinction. Knight argued that many consequential economic decisions involve uncertainty rather than measurable risk. This matters because expected value calculations presuppose a structure that may not actually exist. If the distribution is unknown or the scenario space is incomplete, probabilistic optimization can create an illusion of rigor without genuine epistemic security.

In practice, most important decisions involve mixtures of both. Some inputs can be modeled statistically, while others remain ambiguous or structurally indeterminate. Decision science therefore cannot rely only on methods suited to tidy risk problems. It also needs frameworks for incomplete models, qualitative uncertainty, scenario comparison, and decisions that must remain viable even when the analyst cannot specify the future in a fully probabilistic way.

Ambiguity and the limits of probability

Uncertainty becomes especially important when it takes the form of ambiguity rather than ordinary risk. Under ambiguity, the decision-maker is not merely unsure which outcome will occur. They are unsure how to represent the uncertainty itself. The famous Ellsberg paradox showed that people often prefer known risks to unknown probabilities, suggesting that ambiguity is psychologically and behaviorally distinct from standard probabilistic uncertainty.

This finding has major implications for decision science. It suggests that uncertainty is not always reducible to a single probability distribution, and that decision-makers may rationally or behaviorally resist options whose structure is too opaque. Ambiguity therefore complicates classical decision theory in at least two ways. First, it raises formal questions about whether a single probability measure is adequate. Second, it introduces behavioral and institutional questions about trust, transparency, and model legitimacy.

In real decision environments, ambiguity often arises when data are sparse, causal structure is poorly understood, expert views diverge, or novel conditions make historical frequencies unreliable. These are common features of strategic and policy problems. Under such conditions, uncertainty alters not just outcomes but the epistemic ground on which choices are made.

Uncertainty and the limits of optimization

Classical decision models often assume that the best choice can be identified by maximizing expected utility or minimizing expected loss. These models are powerful when the underlying structure is adequately specified. But when uncertainty is deep, optimization becomes fragile.

There are several reasons for this. Small errors in probability estimates can produce large shifts in expected outcomes. Important variables may be omitted from the model. Historical data may fail to represent future states. Feedback effects and regime changes may make apparently precise estimates highly unstable. Under such conditions, the “optimal” strategy may be optimal only inside a narrow and uncertain model world.

This is why uncertainty often shifts decision-making from optimization toward robustness. Instead of asking which option is best under one forecast, decision-makers ask which options remain acceptable across many plausible futures. RAND’s robust decision-making framework explicitly addresses this problem by focusing on strategies that perform well over a wide range of scenarios rather than those that optimize under a single best estimate.

Uncertainty and cognitive constraints

Uncertainty also interacts with the limits of human cognition. When information is incomplete, ambiguous, or rapidly changing, people rely more heavily on heuristics—mental shortcuts that simplify difficult judgments. These heuristics can be useful, but they also generate systematic distortions.

The work of Amos Tversky and Daniel Kahneman showed that judgment under uncertainty is shaped by heuristics such as availability, representativeness, and anchoring. These are not random mistakes. They are patterned tendencies in how human beings interpret incomplete information. Their 1974 Science paper remains one of the most influential demonstrations that uncertainty amplifies bias rather than merely adding noise to otherwise rational judgment.

Herbert Simon’s concept of bounded rationality provides a complementary perspective. Simon argued that people do not optimize across all possible alternatives because they lack the time, information, and computational capacity required to do so. Instead, they satisfice: they search until they find an option that is good enough relative to their constraints. Under uncertainty, this is not merely a psychological weakness. It is often a necessary and realistic strategy for action. Simon’s Nobel lecture explicitly framed bounded rationality as a response to the inadequacy of classical decision models in real organizational settings.

This is one reason structured decision processes matter. By making assumptions explicit, externalizing reasoning, and comparing scenarios systematically, decision science helps reduce overreliance on intuition alone. It does not eliminate judgment, but it can improve the conditions under which judgment is exercised.

Uncertainty in complex systems

In many important settings, uncertainty arises not merely from missing data but from the behavior of complex systems. Systems marked by feedback loops, delays, nonlinearity, adaptation, and interdependence produce outcomes that are difficult to forecast with confidence even when large amounts of information are available.

A decision in such a system can have delayed and indirect consequences. A policy may produce short-term gains while generating long-term vulnerabilities. A financial intervention may stabilize one variable while increasing fragility elsewhere. A technological system may appear reliable until interacting components create unexpected failure pathways. In these contexts, uncertainty is often structural rather than temporary.

This is why decision science increasingly overlaps with systems modeling, scenario planning, and resilience analysis. In dynamic systems, good decisions cannot be defined solely by immediate expected returns. They must also be evaluated by how they interact with evolving structures over time. Uncertainty in such settings is often irreducible. Better data can help, but it cannot fully eliminate model incompleteness or future novelty.

From optimization to robustness

One of the most important consequences of uncertainty is the shift from optimization to robustness. Robustness does not mean ignoring analysis. It means changing the criterion of success. Instead of choosing the strategy that is best under one assumed future, decision-makers seek strategies that remain acceptable, resilient, or adaptable across many plausible futures.

Robust decision-making has been especially important in environmental planning, infrastructure, water systems, security policy, and other domains where decisions are long-term and uncertainty is deep. RAND describes robust decision-making as a method for making good decisions without first requiring confidence in a single prediction, emphasizing vulnerability analysis, scenario exploration, and adaptive planning.

This shift is intellectually significant because it broadens the meaning of rationality. Rationality under deep uncertainty is not simply the maximization of expected utility under fixed assumptions. It may instead require flexibility, diversification, option preservation, staged commitments, and policies designed to adjust as new information appears.

Uncertainty and trade-offs

Uncertainty also complicates the evaluation of trade-offs. When outcomes are uncertain, alternatives cannot be compared only by their average expected consequences. Decision-makers must consider distributions, downside exposure, catastrophic tails, reversibility, and the value of waiting or learning before committing.

This is especially important in decisions involving irreversible consequences or asymmetric harms. A strategy with a high expected payoff may still be undesirable if it exposes the system to rare but catastrophic losses. Conversely, a more modest strategy may be preferable if it preserves flexibility and avoids ruin. Under uncertainty, decision-making often becomes less about maximizing upside and more about managing exposure, preserving resilience, and clarifying what losses are unacceptable.

These issues connect uncertainty to ethics and institutional judgment. Decisions under uncertainty are not purely technical problems. They also involve judgments about whose risks count, which harms are tolerable, how precaution should be weighed against innovation, and what kind of future the decision-maker is trying to preserve.

Implications for decision-making practice

The presence of uncertainty has several important implications for practice:

Explicit assumptions: assumptions should be visible, challengeable, and revisable rather than hidden inside technical models.
Scenario thinking: multiple plausible futures should be explored instead of relying on a single baseline forecast.
Sensitivity analysis: conclusions should be tested against variation in core parameters, structures, and modeling choices.
Robust strategies: preference should often be given to options that remain viable across changing conditions.
Iterative learning: decisions should be revisited as new information arrives, especially when commitments are staged or partially reversible.
Institutional humility: model outputs should not be mistaken for certainty when the underlying world is only partly understood.

Together, these principles reflect a more mature conception of rationality. Under uncertainty, rational decision-making is not about pretending the future is known. It is about reasoning clearly in the presence of incomplete knowledge and designing choices that can survive error.

Mathematical Lens: expected utility, ambiguity, and minimax regret

Under ordinary risk, an action \(a\) can be evaluated using expected utility:

\[
EU(a) = \sum_{s \in S} p(s)\,u(x(a,s))
\]

where \(p(s)\) is the probability of state \(s\), \(x(a,s)\) is the outcome of taking action \(a\) in state \(s\), and \(u(\cdot)\) is the utility function.

This framework assumes that the state space is known and that a usable probability distribution can be assigned. Under ambiguity, however, the decision-maker may face multiple plausible probability distributions or may resist treating uncertainty as fully specifiable. One stylized way to represent ambiguity aversion is to apply a penalty to options whose probability structure is less well understood:

\[
V(a) = \sum_{s \in S} p(s)\,u(x(a,s)) – \lambda A(a)
\]

where \(A(a)\) is an ambiguity penalty and \(\lambda > 0\) captures the decision-maker’s sensitivity to ambiguity.

Under deep uncertainty, a robust approach may instead use regret. Let \(V(a,s)\) denote the value of action \(a\) in scenario \(s\). Regret is:

\[
R(a,s) = \max_{a’ \in A} V(a’,s) – V(a,s)
\]

and a minimax-regret strategy is:

\[
a^* = \arg\min_{a \in A} \max_{s \in S} R(a,s)
\]

These formulations illustrate a central point in decision science: uncertainty changes not only the inputs to a model, but often the criterion by which alternatives should be judged.

Advanced R Workflow: Comparing expected utility, ambiguity penalties, and robust regret

The R workflow below compares several strategies under three different evaluative logics: expected utility under risk, ambiguity-adjusted valuation, and minimax regret under deep uncertainty.

# Install packages if needed:
# install.packages(c("tidyverse"))

library(tidyverse)

# ------------------------------------------------------------
# Advanced R Workflow:
# Comparing Expected Utility, Ambiguity Penalties,
# and Robust Regret
#
# Purpose:
#   1. Define strategies and scenario payoffs
#   2. Evaluate under expected utility
#   3. Add an ambiguity penalty
#   4. Compute regret and identify robust options
# ------------------------------------------------------------

# ------------------------------------------------------------
# Step 1: Define scenario-specific payoffs
# ------------------------------------------------------------
decision_data <- tibble(
  scenario = c("Baseline", "Adverse", "Severe"),
  probability = c(0.50, 0.30, 0.20),
  Expand = c(110, 40, -90),
  Hedge = c(85, 60, 15),
  Preserve_Option = c(65, 55, 35)
)

print(decision_data)

# ------------------------------------------------------------
# Step 2: Convert to long format
# ------------------------------------------------------------
long_data <- decision_data %>%
  pivot_longer(
    cols = c(Expand, Hedge, Preserve_Option),
    names_to = "strategy",
    values_to = "payoff"
  )

# ------------------------------------------------------------
# Step 3: Define a utility function
# We shift payoffs upward to keep the square-root transform valid
# in this stylized example.
# ------------------------------------------------------------
shift_value <- 100

long_data <- long_data %>%
  mutate(
    shifted_payoff = payoff + shift_value,
    utility = sqrt(shifted_payoff)
  )

# ------------------------------------------------------------
# Step 4: Expected utility under ordinary risk
# ------------------------------------------------------------
expected_utility <- long_data %>%
  group_by(strategy) %>%
  summarise(
    expected_utility = sum(probability * utility),
    .groups = "drop"
  )

print(expected_utility)

# ------------------------------------------------------------
# Step 5: Introduce an ambiguity penalty
# Higher penalty means more concern about model uncertainty
# ------------------------------------------------------------
ambiguity_penalties <- tibble(
  strategy = c("Expand", "Hedge", "Preserve_Option"),
  ambiguity_penalty = c(0.35, 0.15, 0.05)
)

ambiguity_adjusted <- expected_utility %>%
  left_join(ambiguity_penalties, by = "strategy") %>%
  mutate(
    ambiguity_weight = 1.5,
    ambiguity_adjusted_value =
      expected_utility - ambiguity_weight * ambiguity_penalty
  )

print(ambiguity_adjusted)

# ------------------------------------------------------------
# Step 6: Compute regret by scenario
# Regret = best payoff in scenario - strategy payoff
# ------------------------------------------------------------
regret_table <- long_data %>%
  group_by(scenario) %>%
  mutate(
    best_payoff = max(payoff),
    regret = best_payoff - payoff
  ) %>%
  ungroup()

print(regret_table)

# ------------------------------------------------------------
# Step 7: Minimax regret summary
# ------------------------------------------------------------
robust_summary <- regret_table %>%
  group_by(strategy) %>%
  summarise(
    max_regret = max(regret),
    mean_regret = mean(regret),
    .groups = "drop"
  )

print(robust_summary)

# ------------------------------------------------------------
# Step 8: Combine results into one comparison table
# ------------------------------------------------------------
comparison <- ambiguity_adjusted %>%
  left_join(robust_summary, by = "strategy") %>%
  arrange(desc(ambiguity_adjusted_value))

print(comparison)

# ------------------------------------------------------------
# Step 9: Plot evaluation metrics
# ------------------------------------------------------------
comparison_long <- comparison %>%
  select(strategy, expected_utility, ambiguity_adjusted_value, max_regret) %>%
  pivot_longer(
    cols = c(expected_utility, ambiguity_adjusted_value, max_regret),
    names_to = "metric",
    values_to = "value"
  )

ggplot(comparison_long, aes(x = strategy, y = value)) +
  geom_col() +
  facet_wrap(~ metric, scales = "free_y") +
  labs(
    title = "Decision Evaluation Under Risk, Ambiguity, and Deep Uncertainty",
    x = "Strategy",
    y = "Metric Value"
  ) +
  theme_minimal(base_size = 12)

# ------------------------------------------------------------
# Step 10: Export outputs
# ------------------------------------------------------------
write_csv(comparison, "uncertainty_decision_comparison.csv")
write_csv(regret_table, "uncertainty_regret_table.csv")

Advanced Python Workflow: Simulating uncertain futures, ambiguity aversion, and adaptive choice

The Python workflow below simulates repeated decisions under uncertain futures and compares three stylized agents: an expected-value chooser, an ambiguity-averse chooser, and an adaptive chooser that updates behavior after poor outcomes.

# Install packages if needed:
# pip install pandas numpy matplotlib

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# ------------------------------------------------------------
# Advanced Python Workflow:
# Simulating Uncertain Futures, Ambiguity Aversion,
# and Adaptive Choice
#
# Purpose:
#   1. Define strategies and scenario-specific payoffs
#   2. Compare different decision rules under uncertainty
#   3. Track how strategies perform over repeated trials
# ------------------------------------------------------------

np.random.seed(42)

# ------------------------------------------------------------
# Step 1: Define scenarios and probabilities
# ------------------------------------------------------------
scenarios = ["Baseline", "Adverse", "Severe"]
scenario_probs = np.array([0.50, 0.30, 0.20])

strategy_payoffs = {
    "Expand": np.array([110, 40, -90]),
    "Hedge": np.array([85, 60, 15]),
    "Preserve Option": np.array([65, 55, 35])
}

ambiguity_penalties = {
    "Expand": 18,
    "Hedge": 8,
    "Preserve Option": 3
}

# ------------------------------------------------------------
# Step 2: Decision rules
# ------------------------------------------------------------
def expected_value_choice():
    """Choose the strategy with the highest expected payoff."""
    ev_scores = {
        strategy: np.sum(payoffs * scenario_probs)
        for strategy, payoffs in strategy_payoffs.items()
    }
    return max(ev_scores, key=ev_scores.get), ev_scores

def ambiguity_averse_choice():
    """
    Choose the strategy with the highest ambiguity-adjusted score.
    This subtracts a fixed ambiguity penalty from expected payoff.
    """
    adj_scores = {
        strategy: np.sum(payoffs * scenario_probs) - ambiguity_penalties[strategy]
        for strategy, payoffs in strategy_payoffs.items()
    }
    return max(adj_scores, key=adj_scores.get), adj_scores

def adaptive_choice(previous_loss_count):
    """
    Stylized adaptive rule:
    if recent severe losses accumulate, shift toward preserving flexibility.
    """
    if previous_loss_count >= 3:
        return "Preserve Option"
    elif previous_loss_count == 2:
        return "Hedge"
    return "Expand"

# ------------------------------------------------------------
# Step 3: Determine ex-ante fixed strategies
# ------------------------------------------------------------
ev_strategy, ev_scores = expected_value_choice()
amb_strategy, amb_scores = ambiguity_averse_choice()

print("Expected value scores:", ev_scores)
print("Expected value strategy:", ev_strategy)
print("Ambiguity-adjusted scores:", amb_scores)
print("Ambiguity-averse strategy:", amb_strategy)

# ------------------------------------------------------------
# Step 4: Simulate repeated uncertain futures
# ------------------------------------------------------------
n_trials = 250
loss_count = 0
records = []

for trial in range(1, n_trials + 1):
    scenario_index = np.random.choice(len(scenarios), p=scenario_probs)
    scenario_name = scenarios[scenario_index]

    ev_payoff = strategy_payoffs[ev_strategy][scenario_index]
    amb_payoff = strategy_payoffs[amb_strategy][scenario_index]

    adaptive_strategy = adaptive_choice(loss_count)
    adaptive_payoff = strategy_payoffs[adaptive_strategy][scenario_index]

    # Update loss memory for adaptive agent
    if adaptive_payoff < 0:
        loss_count += 1
    else:
        loss_count = max(0, loss_count - 1)

    records.append({
        "trial": trial,
        "scenario": scenario_name,
        "ev_strategy": ev_strategy,
        "ev_payoff": ev_payoff,
        "ambiguity_strategy": amb_strategy,
        "ambiguity_payoff": amb_payoff,
        "adaptive_strategy": adaptive_strategy,
        "adaptive_payoff": adaptive_payoff,
        "loss_memory": loss_count
    })

df = pd.DataFrame(records)
print(df.head())

# ------------------------------------------------------------
# Step 5: Summarize performance
# ------------------------------------------------------------
summary = pd.DataFrame({
    "agent": ["Expected Value", "Ambiguity Averse", "Adaptive"],
    "average_payoff": [
        df["ev_payoff"].mean(),
        df["ambiguity_payoff"].mean(),
        df["adaptive_payoff"].mean()
    ],
    "minimum_payoff": [
        df["ev_payoff"].min(),
        df["ambiguity_payoff"].min(),
        df["adaptive_payoff"].min()
    ],
    "maximum_payoff": [
        df["ev_payoff"].max(),
        df["ambiguity_payoff"].max(),
        df["adaptive_payoff"].max()
    ]
})

print(summary)

# ------------------------------------------------------------
# Step 6: Cumulative payoff paths
# ------------------------------------------------------------
df["ev_cumulative"] = df["ev_payoff"].cumsum()
df["ambiguity_cumulative"] = df["ambiguity_payoff"].cumsum()
df["adaptive_cumulative"] = df["adaptive_payoff"].cumsum()

plt.figure(figsize=(10, 6))
plt.plot(df["trial"], df["ev_cumulative"], label="Expected Value")
plt.plot(df["trial"], df["ambiguity_cumulative"], label="Ambiguity Averse")
plt.plot(df["trial"], df["adaptive_cumulative"], label="Adaptive")
plt.xlabel("Trial")
plt.ylabel("Cumulative Payoff")
plt.title("Decision Performance Under Uncertain Futures")
plt.legend()
plt.tight_layout()
plt.show()

# ------------------------------------------------------------
# Step 7: Export outputs
# ------------------------------------------------------------
summary.to_csv("uncertainty_simulation_summary.csv", index=False)
df.to_csv("uncertainty_simulation_trials.csv", index=False)

Conclusion

Uncertainty changes decision-making by transforming it from a problem of clean optimization into a process of structured judgment under incomplete knowledge. It weakens the assumptions needed for precise prediction, intensifies the role of cognitive and institutional limits, and makes robustness, adaptability, and transparency more important than the appearance of exactness.

In uncertain environments, the goal is not to eliminate uncertainty altogether. It is to reason well despite it. Decision science helps by making assumptions explicit, comparing alternative futures, testing sensitivity, and designing choices that remain defensible even when forecasts fail. That is why uncertainty does not sit at the margins of the field. It is one of the central reasons decision science exists at all.

References

Ellsberg, D. (1961) ‘Risk, Ambiguity, and the Savage Axioms’, Quarterly Journal of Economics, 75(4), pp. 643–669. Stable record available at: JSTOR.
Knight, F.H. (1921) Risk, Uncertainty, and Profit. Boston, MA: Houghton Mifflin. Available at: Online Library of Liberty.
RAND Corporation (2021) Robust Decision Making for Planning Under Deep Uncertainty. Available at: RAND.
RAND Corporation (n.d.) ‘Decision Making under Deep Uncertainty’. Available at: RAND.
RAND Corporation (n.d.) ‘Robust Decision Making’. Available at: RAND.
Simon, H.A. (1978) ‘Rational Decision-Making in Business Organizations’, Nobel Prize lecture, 8 December. Available at: NobelPrize.org.
Tversky, A. and Kahneman, D. (1974) ‘Judgment under Uncertainty: Heuristics and Biases’, Science, 185(4157), pp. 1124–1131. Available at: Science.