Biases and Heuristics

Learning Objectives

Upon completing this book, you will be able to:

1. Explain what heuristics are, why they exist, and when they are useful versus when they produce systematic errors.

2. Identify the anchoring effect in negotiations, estimates, and intention formulations, and design mechanisms to counteract it.

3. Recognize the availability heuristic in risk and frequency assessments, and understand why it distorts user priorities.

4. Detect the representativeness heuristic and its tendency to ignore base rates, resulting in incorrect probability judgments.

5. Understand the confirmation bias as the most dangerous bias for intention architecture and design countermeasures.

6. Apply Gigerenzer's concept of ecological rationality to distinguish when heuristics serve the user and when they deceive them.

1. Introduction: Shortcuts that Work (Until They Don't)

In the previous topics, you learned that the brain is an energy-constrained organ that prioritizes speed over accuracy, operates mostly in automatic mode (System 1), and builds coherent narratives with incomplete information (WYSIATI). Heuristics are the specific mechanism by which System 1 achieves quick decisions under these constraints.

“Imagine you're in a forest and you hear a rustling in the bushes. Your brain doesn't start a statistical analysis of the probability that it's a predator versus a falling branch. In milliseconds, it activates a heuristic: 'if it sounds dangerous, act as if it's dangerous.' You run away. This heuristic will save your life 1% of the times it's actually a predator, and it will make you run unnecessarily the other 99%. But in evolutionary terms, an extra 1% of survival is huge.”

Heuristics are not thinking errors. They are adaptive strategies that evolved because they work well enough in the environments where human cognition developed. The problem arises when these same shortcuts are applied in modern, complex, or artificial contexts for which they were not designed. That's when they produce systematic deviations from rational judgment: cognitive biases.

“It's like using a machete in the jungle versus in surgery. A machete is a brilliant tool for cutting your way through vegetation. Using it to operate on someone would be disastrous. Heuristics are cognitive machetes: perfect for quick decisions in natural environments, dangerous when applied to problems that require precision.”

2. The Anchoring Effect: The First Figure Wins

The anchoring effect is one of the most robust and well-documented biases in all of decision psychology. It consists of the fact that initial exposure to a number—even a completely irrelevant one—disproportionately influences subsequent numerical estimates. The first number you encounter becomes an "anchor" around which your estimates revolve.

“In a classic experiment, Kahneman and Tversky spun a rigged wheel that only landed on 10 or 65. They then asked participants: "What percentage of African countries are part of the United Nations?" Those who saw the 10 estimated an average of 25%. Those who saw the 65 estimated 45%. A roulette number—obviously irrelevant—altered the estimate by 20 percentage points.”

2.1 Why Anchoring Works

Anchoring operates through two complementary mechanisms. The first is insufficient adjustment: when you don't know the answer, you start from the anchor and "adjust" in the direction that seems correct to you, but the adjustment is always insufficient—you stay too close to the anchor. The second is selective activation: the anchor activates information in your memory that is consistent with it, biasing the material that your brain uses to build the estimate.

“If you're asked "Was Gandhi more or less than 140 years old when he died?", you obviously know that no one lives 140 years. But your subsequent estimate will be systematically higher than if the question had been "more or less than 35 years?" The absurd anchor pulls you, even though you know it's absurd. System 1 can't ignore it; System 2 corrects it partially, but never completely.”

2.2 Anchoring in Formulating Intentions

Anchoring directly affects how users formulate requests to AI systems. The first example they see, the first figure they encounter, or the first formulation that occurs to them becomes an anchor that conditions everything that follows.

Anchoring in budgets: If a user searches for "how much does a website cost" and the first result says $50,000, that figure anchors their expectations. If they then ask an AI system to estimate costs, their reaction to the result will be conditioned by that initial anchor, regardless of whether $50,000 is a representative number for their case.

Anchoring in scope: A user who has seen an 80-slide presentation will tend to ask for long presentations. One who has seen a 10-slide presentation will ask for short presentations. The anchor is not their real need; it's their previous exposure.

“An entrepreneur asks: "Make me a business plan like the one I read from Uber." The Uber plan they read—probably a 60-page document oriented towards Silicon Valley investors—becomes the anchor. Their three-person startup that needs a 5-page operational plan remains invisible behind that anchor. The user doesn't ask for what they need; they ask for what the anchor dictates.”

Anchoring in the system's own responses: The first response an AI system produces becomes an anchor for the user. If the first version of a text has 500 words, the user will evaluate subsequent versions in relation to that length, even if the optimal length could be completely different.

3. The Availability Heuristic: What You Recall, Exists

The availability heuristic is the cognitive shortcut by which people judge the frequency or probability of an event based on how easily examples of that event come to mind. If you can recall something quickly, your System 1 concludes that it must be frequent or important. If you can't recall it easily, it must be rare or irrelevant.

“What kills more people per year: shark attacks or falling coconuts? Most people answer sharks without hesitation, because movies, news, and documentaries have filled your memory with vivid images of attacks. In reality, coconuts kill around 150 people per year; sharks, around 5. Coconuts are 30 times more lethal, but invisible to your availability heuristic because nobody makes documentaries about death by coconut.”

3.1 What Makes Something “Available”

Availability does not depend on the actual frequency of an event, but on factors that make certain memories more accessible:

Vividness: Vivid, emotional, and sensorially rich events are stored more strongly. A plane crash you saw on TV weighs more in your risk assessment than thousands of safe flights you never observed.

“After watching a movie about corporate fraud, your perception that ‘companies are corrupt’ increases temporarily. You didn't learn new information; you simply activated a vivid memory that now dominates your evaluation.”

Recency: What happened recently is more available than what happened years ago, regardless of which is more representative.

“If the last time you ate sushi you felt sick, your estimate of the probability of getting sick from sushi is disproportionately high. It doesn't matter that you've eaten sushi 200 times without problems; the recent experience dominates the calculation.”

Personal relevance: Events that happened to you personally are more available than those that happened to others. A robbery you suffered weighs more than the crime statistics of the entire city.

Media coverage: The media distorts availability by overrepresenting dramatic events and underrepresenting everyday risks. Terrorism dominates the news; diabetes kills 100 times more people but is almost never reported.

3.2 Availability in Formulating Intentions

The availability heuristic distorts the user's priorities when formulating intentions for an AI system. The user doesn't ask for what they objectively need; they ask for what their mind has most “at hand”.

“A manager who just read an article about cybersecurity asks an AI system: ‘Perform a security audit of our company.’ It sounds reasonable. But if the company doesn't even have a basic backup plan, the real priority is probably not a sophisticated audit but basic infrastructure. The article they read made ‘advanced cybersecurity’ available in their mind; ‘basic backups’ wasn't.”

A system with good intention architecture doesn't simply execute the availability-anchored request. It explores the context: “What prompted this request? What security measures do they already have in place? What's the most likely threat to their type of business?” These questions allow the user to escape the availability trap.

4. The Representativeness Heuristic: It Looks Like, So It Is

The representativeness heuristic is the shortcut by which people judge the probability of something belonging to a category based on how much it resembles the prototype of that category, ignoring critical statistical information such as base rates.

“Kahneman and Tversky presented participants with the description of Tom: “Tom is introverted, orderly, methodical, and enjoys logical puzzles.” Then they asked: is it more likely that Tom studies computer science or humanities? Most said computer science, because Tom “looks like” the stereotype of a computer scientist. But the number of humanities students is enormously larger than that of computer science. Statistically, it is more likely that Tom studies humanities simply because there are many more students there. Representativeness overrode the base rate.”

4.1 The Base Rate Fallacy

The base rate is the frequency with which something occurs in the general population. Ignoring it is one of the most frequent and dangerous statistical errors of human thinking. The representativeness heuristic systematically makes us ignore base rates in favor of superficial similarities.

“A medical test has 99% accuracy. Your result is positive. What is the probability that you are actually sick? Most say 99%. The correct answer depends on the base rate: if the disease affects 1 in 10,000 people, and the test has 1% false positives, then out of 10,000 people tested, 1 will have a true positive and 100 will have false positives. Your real probability of being sick is approximately 1 in 101, or less than 1%. Representativeness (“the test was positive, it looks like being sick”) defeated the base rate.”

This error has real consequences in medicine, criminal justice, job hiring, and any area where decisions are based on probabilistic indicators.

4.2 The Conjunction Fallacy

The representativeness heuristic also produces the conjunction fallacy: the belief that the combination of two events is more probable than one of them alone. This is logically impossible—the probability of A and B together can never be greater than that of A alone—but representativeness deceives us when the combination “looks more coherent.”

“The most famous example is the Linda problem: “Linda is 31 years old, single, straightforward, and brilliant. She studied philosophy, participated in anti-nuclear and social justice demonstrations.” What is more probable: that Linda is a bank teller, or that Linda is a bank teller and a feminist activist? Most choose the second option because it fits better with the description. But mathematically, “bank teller and feminist” is always less probable than just “bank teller”—because it is a subset.”

4.3 Representativeness in Formulating Intentions

Representativeness affects the formulation of intentions when users categorize situations based on superficial similarities rather than analysis.

“A business owner sees that a successful competitor uses Instagram. Their business “looks like” the competitor's, so they conclude: “I need an Instagram strategy.” But perhaps the competitor is successful despite Instagram, not because of it. Or perhaps the competitor's customers are demographically different. The superficial similarity between the businesses activated the representativeness heuristic and produced an intention based on superficial correlation, not causal analysis.”

5. Confirmation Bias: Seeing What You Want to See

Confirmation bias is the tendency to seek, interpret, favor, and remember information that confirms the beliefs or hypotheses you already have, while ignoring or discrediting information that contradicts them. It is, arguably, the most dangerous and consistent bias for the architecture of intention.

“If you believe your employee Juan is lazy, you'll start to notice every time he's five minutes late, every break he takes, every task he doesn't complete on time. But you won't notice the times he stays late, helps a colleague, or solves a difficult problem. You're not lying; you genuinely “see” more evidence of laziness than productivity. Your attention is being directed by your prior belief.”

5.1 The Three Dimensions of Confirmation Bias

Biased Search: You actively seek information that confirms what you already believe. If you believe the real estate market is going to rise, you'll read articles that say it's going to rise and avoid those that say it's going to fall.

“A user who has already decided they want to start a food truck business asks an AI: “Give me reasons why food trucks are a good business.” They don't ask: “Give me a balanced analysis of pros and cons.” The request itself is biased towards confirmation.”

Biased Interpretation: The same evidence is interpreted differently based on your prior beliefs. Ambiguous data is read as supporting your position.

“Two people see the same statistic: “60% of new restaurants close within 5 years.” The optimist thinks: “40% survive, and I'll be one of them.” The pessimist thinks: “Most fail, it's too risky.” Same data, two opposing interpretations, both “confirming” the prior belief.”

Biased Memory: You remember better the information that confirms your beliefs. After a week, the optimist will remember the 40% that survive; the pessimist will remember the 60% that fail.

5.2 Confirmation Bias in Interaction with AI

Confirmation bias is particularly dangerous in interaction with AI systems because the user controls the formulation of the request. If the user already has a conclusion in mind, they will formulate their question in a way that biases the response towards that conclusion.

“Compare these two requests to the same AI system: (1) “Why is Python the best programming language?” versus (2) “Compare Python, JavaScript, and Rust for my data analysis project.” The first question already contains the conclusion; the AI will produce arguments in favor of Python because that's what it was asked. The second allows for a genuine analysis. But a user with confirmation bias naturally formulates the first version.”

An AI system that simply responds to what it's asked is acting as a confirmation mirror: it returns to the user what the user wants to hear. A robust architecture of intention must include mechanisms to present disconfirming evidence, even when the user doesn't request it.

6. Metabias: Dunning-Kruger and Ecological Rationality

Beyond specific biases, there are two concepts that frame our general relationship with our own judgment errors: the Dunning-Kruger effect, which describes when we don't know that we don't know, and Gigerenzer's ecological rationality, which rescues heuristics as legitimate tools when used in the correct environment.

6.1 The Dunning-Kruger Effect

The Dunning-Kruger effect reveals an asymmetry in metacognitive accuracy: people with low competence in an area tend to overestimate their ability, while experts tend to underestimate theirs. This occurs because evaluating one's own competence requires the same competence being evaluated—a vicious circle.

“A beginner in investment who gains profits in a bull market may be convinced they have a talent for finance. They don't recognize that anyone makes money when everything rises; their 'success' is due to the market, not their ability. Meanwhile, a professional manager with decades of experience says cautiously: 'Markets are unpredictable, and I have much to learn.' The overconfident beginner and the humble expert are expressions of the same phenomenon.”

For intention architecture, this has direct practical implications: a novice user who expresses high confidence in their problem formulation may need more guardrails than an expert who expresses uncertainty. Declared confidence does not correlate with competence; it often correlates inversely.

“A user who says 'I know exactly what I need: do a multiple regression analysis with these variables' may be asking for the wrong statistical tool for their problem. A user who says 'I'm not sure how to approach this problem, but I want to understand what factors affect my sales' will probably receive better help because their open formulation allows the system to explore options. Paradoxically, the second user's uncertainty is more informative than the first user's certainty.”

6.2 Gigerenzer's Ecological Rationality

Gerd Gigerenzer offers an essential counterpoint to the Kahneman-Tversky program. While Kahneman documented when heuristics fail, Gigerenzer demonstrates when they work better than complex methods. His concept of ecological rationality proposes that a heuristic is rational when it is adapted to the structure of its environment.

“In medical emergencies, a simple decision tree of three questions ('Chest pain?', 'Systolic pressure over 145?', 'Age over 62?') predicts cardiac risk better than a complex statistical model with dozens of variables. Why? Because in emergencies, information is limited, time is scarce, and data is noisy. The simple heuristic is more robust because it ignores the noise that the complex model tries—and fails—to process.”

Gigerenzer calls these tools “fast and frugal heuristics.” Their power lies precisely in what they ignore: by discarding irrelevant or noisy information, they are less susceptible to overfitting and more robust in changing environments.

“Imagine you have to choose between two restaurants. One, you've heard mentioned several times. The other, you've never heard of. The recognition heuristic says: 'choose the one you recognize.' This sounds simplistic, but in many contexts, it works because recognition correlates with quality, popularity, or permanence. Gigerenzer showed that this heuristic outperforms complex algorithms in sports and financial predictions when available information is limited.”

7. Practical Application: Biases and Heuristics in Action

Let's integrate the biases we've studied into a scenario where they all operate simultaneously—as they do in real life.

7.1 Case Study: The Marketing Campaign

Roberto, the marketing director of a medium-sized company, asks an AI system: “I want a viral campaign on TikTok. I saw that a competitor made a video that had 5 million views. I need something like that. Budget: $10,000.”

Let's analyze the biases present:

Anchoring: The 5 million views of the competitor are the anchor. Roberto will evaluate the success of his campaign relative to that figure. If his video gets 500,000 views—a great result for his company size—he'll consider it a failure because it's anchored to 5 million.

Availability: “TikTok” and “viral” are available in his mind because he just saw the competitor's video. Other potentially better options for his case are not available: email marketing (with an average ROI of 36:1), Google optimization (where his customers actually search), or strategic partnerships.

Representativeness: Roberto assumes his company “looks like” the competitor and therefore the same strategy will work. He doesn't consider that the competitor has an established brand, a young audience, a visually appealing product, and a 10-person content team—invisible variables behind the visible “success”.

Confirmation Bias: Roberto has already decided that TikTok is the answer. He'll look for validation of that decision and interpret any ambiguous data as support. If the AI suggests an alternative, he'll feel resistant.

Dunning-Kruger: Roberto is not an expert in digital marketing, but his exposure to a successful case gives him the illusion of understanding. His confidence is high; his competence in the area, limited.

“A system with good intention architecture would respond: “I can help you design an effective campaign. Before defining the channel, I'd like to understand better: what's the real objective—brand visibility, leads, direct sales? Where are your current customers—are they active on TikTok? With $10,000, there are several routes: TikTok is one, but we could also consider channels where your audience is already. Do you want to see a comparison of options with return estimates?” This response unanchors (introduces options), reintroduces base rates (where the customers really are), and challenges representativeness (your case is not identical to the competitor's).”

7.2 Anti-Bias Design Principles

1. Actively unanchor. Presentation of multiple reference points instead of just one. Offer ranges instead of single figures.

2. Reintroduce base rates. When the user generalizes from a case, present the actual frequency of the phenomenon.

3. Counter availability. Presentation of options that the user probably doesn't have “at hand” but that could be superior.

4. Offer disconfirming evidence. Proactively include information that challenges the user's hypothesis, especially when the formulation suggests confirmation bias.

5. Calibrate confidence. Treat very specific requests from non-expert users with more caution than open requests from experts.

6. Respect appropriate heuristics. When the user's heuristic is backed by genuine experience in the relevant domain, support rather than question it.

8. Glossary of Key Terms

9. Final Reflection and Self-Assessment

Before moving on to the last topic of Module I (Motivation and Emotion), verify that you can apply these concepts.

10. Bibliography and Recommended Readings

Essential Readings

• Kahneman, D., Slovic, P. & Tversky, A. (1982). Judgment Under Uncertainty: Heuristics and Biases. Cambridge University Press.

• Gigerenzer, G. (2007). Gut Feelings: The Intelligence of the Unconscious. Viking. [Decisiones instintivas. Editorial Ariel]

• Ariely, D. (2008). Predictably Irrational. HarperCollins. [Las trampas del deseo. Editorial Ariel]

Complementary Readings

• Tversky, A. & Kahneman, D. (1974). “Judgment Under Uncertainty: Heuristics and Biases.” Science, 185(4157), 1124–1131.

• Kruger, J. & Dunning, D. (1999). “Unskilled and Unaware of It.” Journal of Personality and Social Psychology, 77(6), 1121–1134.

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

• Nickerson, R.S. (1998). “Confirmation Bias: A Ubiquitous Phenomenon in Many Guises.” Review of General Psychology, 2(2), 175–220.

Next Topic

The next and final book in Module I addresses Motivation and Emotion: the engine of intention. You will explore the theory of self-determination, intrinsic versus extrinsic motivation, the cognitive evaluation theory of emotions, and the affect heuristic. You will connect everything learned in the previous topics with the force that initiates and sustains human action.