Framing

Learning Objectives

By completing this book, you will be able to:

1. Distinguish the three types of behavioral framing—attribute, goal, and risky choice—and explain the specific psychological mechanism by which each alters decisions.

2. Apply attribute framing to analyze how the description of the same fact produces radically different evaluations according to the words chosen.

3. Design behavioral interventions that use goal framing (gain vs. loss) to motivate specific behaviors, anticipating the conditions under which each frame is more effective.

4. Analyze how Thaler and Sunstein's libertarian paternalism proposes to resolve the tension between influencing decisions and respecting freedom of choice.

5. Identify how AI systems frame their responses and how that framing shapes users' decisions, frequently without either party being aware of it.

6. Apply ethical criteria to distinguish between transparent framing that empowers the user and opaque framing that manipulates them.

1. Three Types of Behavioral Framing

Framing in behavioral economics is not a single concept but a family of related yet distinct phenomena. Irwin Levin, Sandra Schneider, and Gary Gaeth proposed in 1998 a typology distinguishing three fundamental types, each with a different psychological mechanism, a different effect on decisions, and a different implication for design. Confusing the three is one of the most frequent errors in the applied literature.

This typology will give you precision where until now you had a general intuition. In Module II you learned that frames are inevitable and decisive. Now you will learn to classify them, to predict their effects, and to design them deliberately.

1.1 Attribute Framing

Attribute framing occurs when a specific characteristic of an object or event is described in positive or negative terms. The objective information is identical; the valence of the description changes. The systematic effect: the positive description produces more favorable evaluations than the negative one. Always. Regardless of whether the person ‘knows’ that both descriptions are equivalent.

“A package of meat labeled ‘95% lean’ is evaluated as healthier, of better quality, and better tasting than the same package labeled ‘5% fat.’ Buyers can calculate that these are the same thing. It does not matter. System 1 processes the valence before System 2 has any opportunity to intervene. By the time reason arrives, the evaluation is already formed.”

The underlying mechanism is associative encoding in memory: ‘lean’ activates positive associations (health, slimness, care); ‘fat’ activates negative ones (illness, excess, neglect). These associations operate automatically—they are the affect heuristic you studied in Module I. The word does not merely describe the attribute; it evokes a complete emotional field that colors the entire evaluation.

“An AI system that presents a risk analysis by saying ‘this strategy has a 70% probability of success’ produces a more favorable evaluation than if it says ‘this strategy has a 30% probability of failure.’ The user knows these are the same thing. But their System 1 has already processed the valence, and the evaluation is biased before the rational analysis begins. The intention architect must decide: what attribute framing does your system use, and who benefits from that choice?”

1.2 Goal Framing

Goal framing occurs when a persuasive message emphasizes the positive consequences of performing an action (gain framing) or the negative consequences of not performing it (loss framing). Unlike attribute framing, which describes an object, goal framing seeks to motivate a behavior. And here the relationship is not simple: the positive frame does not always win.

Research shows a pattern consistent with Prospect Theory. For prevention behaviors—actions that reduce risk, such as using sunscreen, having medical check-ups, or backing up data—loss framing is more effective: “if you don't have the check-up, you might not detect the problem in time.” For promotion behaviors—actions that generate a benefit, such as adopting a new habit, trying a tool, or changing strategy—gain framing works better: “with this tool you'll be able to analyze data in half the time.”

“An AI platform wants users to activate fact-checking on their queries. Gain framing: ‘Activate verification to get more accurate answers.’ Loss framing: ‘Without verification, you could base important decisions on incorrect information.’ Since verification is a prevention behavior (it reduces the risk of error), loss aversion predicts that the second message will be more effective. And the empirical evidence confirms it.”

1.3 Risky Choice Framing

The third type is the one Kahneman and Tversky demonstrated in their famous ‘Asian disease’ experiment. When the same options are presented in terms of gains (lives saved), people choose the safe option. When they are presented in terms of losses (deaths), they choose the risky one. The expected values are mathematically identical; the decisions are opposite.

This is the framing effect you already studied in Prospect Theory (Topic 1): the asymmetric value function produces risk aversion in the domain of gains and risk seeking in the domain of losses. What this topic adds is the design implication: whoever presents the options determines whether the decision-maker finds themselves in the territory of gains or of losses, and therefore determines their attitude toward risk. This is not a technical detail—it is enormous power.

“A consultant presents two strategic options to an executive committee. If he frames them as ‘Both options protect at least 60% of current market share’ (gain framing), the committee will choose the conservative option. If he frames them as ‘Both options involve losing between 20% and 40% of share’ (loss framing), the committee will choose the risky one. The consultant did not lie; he described the same reality from two equally valid frames. But the frame he chose determined the company's strategic decision.”

2. Framing as a Tool of Design

If framing is inevitable—and it is, because every presentation of information has a frame—then the question is not whether to frame but how to do it responsibly. Thaler and Sunstein proposed libertarian paternalism as a philosophical framework: use framing to guide people toward better decisions without eliminating their freedom of choice. It is ‘paternalist’ because the designer has an opinion about what is better; it is ‘libertarian’ because the user can always choose otherwise.

Framing as a design tool has three dimensions: what information is presented, how it is presented, and what is omitted. Each dimension is a decision of the architect that shapes the decision of the user. And every omission is an act of framing as powerful as every inclusion—perhaps more so, because what is omitted is invisible.

“An AI system that generates three strategic options and presents the first with three paragraphs of detail, the second with two, and the third with one, is framing without words. Length is a signal of importance. The user, operating with System 1, will interpret the first option as the most complete and probably the most recommendable. The system never said ‘choose the first one’; it communicated it through the architecture of the presentation.”

2.1 Defaults as Structural Framing

In Topic 1 of this module you studied the power of defaults from the perspective of status quo bias. Now you examine them from the perspective of framing: the default is the most powerful framing there is, because it does not frame the information—it frames the action itself. The default says: “this is the normal, expected, recommended option.” Every deviation from the default requires a deliberate act which, beyond the cognitive effort, communicates a social message: “I am not like the majority.”

Eric Johnson and Daniel Goldstein's research on organ donation is the paradigmatic case. Countries with opt-out (you are a donor unless you exclude yourself) have rates of 85–99%. Countries with opt-in (you must actively register) have rates of 4–27%. The preferences of the populations are similar; the default is the variable that explains the difference. It is no exaggeration to say that the choice of default in this case has determined millions of lives saved or lost.

“When an AI system offers a default level of detail (for example, ‘executive summary’ as the default and ‘detailed analysis’ as an option), most users accept the default. If the default is the summary, most will receive superficial analyses. If the default were the detailed analysis, most would receive better answers. The default is not neutral—it is the frame that determines the depth of thinking the user will apply to their problem.”

2.2 Simplification and Translation

The second mechanism of framing as a design tool is simplification. Complex information is not processed; it is ignored. If you want a user to make a good decision, you must translate the complexity into a format their System 1 can process without System 2 needing to activate. This is framing in its most fundamental sense: choosing which aspects of the complete reality to present and which to simplify or exclude.

Gigerenzer, from the tradition of ecological rationality, adds a crucial nuance: simplification is not always a loss of information. In contexts with high uncertainty and scarce data, simple representations can produce better decisions than complex ones. A chart showing a trend is, in many contexts, more useful than a table with 50 columns of data. Not because the table is incorrect, but because the human brain cannot process it.

“A doctor who says ‘you have a 1 in 10 chance of side effects’ communicates better than one who says ‘you have a 10% probability of side effects.’ Gigerenzer demonstrated that natural frequencies (1 in 10) produce better comprehension than percentages, especially in medical contexts. Framing is not only what information to present but in what format—and the right format depends on how the human brain processes information, not on how the system generates it.”

3. Framing in Interaction with AI

AI systems are, by nature, framing machines. Every response a language model generates is an act of selection, emphasis, and omission—the three constitutive functions of the frame (Module II). And unlike a human frame, which is frequently conscious or at least contestable, the AI's framing operates with an appearance of neutrality that makes it more dangerous: the user tends to perceive the response as ‘the information,’ not as one of many possible ways of presenting it.

“A user asks: ‘Should I invest in cryptocurrencies?’ The AI can frame its response in multiple ways. Opportunity framing: ‘Cryptocurrencies have generated significant returns for early investors and offer diversification…’ Risk framing: ‘Cryptocurrencies are highly volatile assets with a significant risk of total loss…’ Balanced framing: ‘Cryptocurrencies present both significant opportunities and risks; the decision depends on your risk tolerance and time horizon.’ Each framing activates a different evaluation in the user. And the one the system chooses—or the one its training produces by default—is not neutral.”

3.1 The Invisible Framing of Training

Language models inherit the framings of their training data. If most of the texts about entrepreneurship in the training corpus frame the activity as ‘opportunity’ rather than ‘risk,’ the model will reproduce that framing bias without anyone having explicitly programmed it. It is a systemic attribute framing: the valence inherited from the corpus becomes the default valence of the responses.

This phenomenon connects directly with the algorithmic biases you will study in Module VI: framing bias in training data is amplified at scale when millions of users interact with the model. What was the prevalent framing in a corpus of texts becomes the framing that shapes the decisions of millions of people. Scale transforms a statistical bias into a systemic one.

“If an AI model has been trained on more texts framing productivity as ‘doing more’ than as ‘doing better,’ it will tend to respond to productivity requests with suggestions of volume (more tasks, faster) instead of suggestions of quality (fewer tasks, better thought out). The user who does not detect this inherited framing will incorporate the bias into their own formulation of intentions without realizing it. The conceptual metaphor of Module IV operates here: PRODUCTIVITY IS QUANTITY is the invisible frame determining which solutions can be imagined.”

3.2 The Framing of the Options Presented

When an AI system presents multiple options—strategies, solutions, approaches—the order, the length, and the structure of the presentation constitute a framing that shapes the choice. Four presentation biases operate simultaneously. Primacy bias: the first option receives more attention and a more favorable evaluation. Length bias: the most detailed option is perceived as more complete and trustworthy. Numbering bias: numbered options imply a ranking, even when there is none. Confidence bias: the option presented with more certainty is perceived as better, regardless of its real quality.

“A user asks an AI for three approaches to reorganizing their team. The AI presents: ‘1. Reorganization into squads… (three paragraphs). 2. Matrix structure… (two paragraphs). 3. Functional teams… (one paragraph).’ The user, operating with System 1 under time pressure, chooses squads—not because it is the best option but because primacy, length, numbering, and the confidence implicit in the presentation conspired to push them in that direction. The AI did not recommend squads; its framing did.”

4. Framing and Language: The Connection with Module IV

Behavioral framing does not operate in a vacuum: it materializes through language. Each of the tools of Module IV is a framing mechanism. Grice's implicature frames by communicating without saying. Presupposition frames by taking for granted. Lakoff's conceptual metaphor frames by structuring one domain in terms of another. Searle's indirect speech acts frame by doing one thing with words that say another. Behavioral framing and linguistic framing are, ultimately, the same phenomenon seen from two disciplines.

“An AI system that responds ‘You might consider…’ frames its suggestion as optional and exploratory. The same content presented as ‘The best strategy is…’ frames the suggestion as directive and confident. Linguistically, the first case is an indirect speech act (a suggestion disguised as a possibility); the second is a strong assertive. Behaviorally, the first activates evaluation by the user; the second activates acceptance. The choice of words is a choice of framing.”

Conceptual metaphor deserves special attention because it is the deepest and most invisible framing mechanism. When Lakoff demonstrated that we speak of TIME IS MONEY (‘I spend time,’ ‘I invest hours,’ ‘I have no time left’), he revealed that metaphor is not a rhetorical ornament but a cognitive structure that determines how we think about time. In Intention Architecture, detecting the metaphor a user employs to describe their problem is detecting the deep framing that determines which solutions they can conceive.

The connection between behavioral framing and linguistic framing produces a powerful analytic tool for the intention architect: when a user formulates a request, their choice of words reveals the behavioral framing under which they are operating. Detecting that framing—through the pragmatic analysis of Module IV—is the first step toward being able to question it, complement it, or transform it.

5. The Ethics of Framing

Framing raises an ethical dilemma that anticipates Module VII and that must be addressed now: if the presentation of information shapes decisions, and if every presentation is a framing, then when is framing legitimate and when is it manipulation? The line is not trivial, and the course's reference authors offer different criteria.

Thaler and Sunstein argue that libertarian paternalism is ethical if it meets three conditions: it preserves freedom of choice (the user can reject the framing), it is motivated by the user's wellbeing (not the designer's), and it is transparent (the user can know how the system works if they ask). Sunstein, in The Ethics of Influence, goes deeper: the key is whether the framing would survive the test of transparency—would the designer be comfortable explaining publicly how they frame the information and why?

“An AI system that frames its responses so that the user reflects more before acting—for example, by implicitly presenting trade-offs the user had not considered—passes the transparency test: if the user asked ‘why are you showing me the trade-offs?’, the answer ‘because we want you to make an informed decision’ is legitimate. A system that frames its responses so that the user accepts without questioning—for example, by hiding uncertainty in order to appear more trustworthy—fails the test: if the user asked ‘why don't you tell me when you're not sure?’, there is no legitimate answer.”

Gigerenzer offers a complementary criterion from ecological rationality: ethical framing is that which facilitates genuine comprehension, not that which facilitates persuasion. Presenting probabilities as natural frequencies (1 in 10) instead of percentages (10%) is ethical framing because it improves the user's comprehension. Presenting only the benefits of a treatment without the risks is manipulative framing because it facilitates acceptance at the expense of comprehension.

5.1 Five Criteria for Ethical Framing

1. Transparency: The framing must be explicable openly. If you would hide how you present the information, you are probably manipulating.

2. Beneficiary: Who benefits from this framing? If it benefits the user, it is architecture. If it benefits only the designer, it is manipulation. If it benefits both, it is aligned design.

3. Reversibility: Can the user easily access the alternative framing? If the system presents ‘95% success’ and the user can ask ‘show me the failure rate too’ without friction, the framing is legitimate. If accessing the alternative framing requires disproportionate effort, it is suspect.

4. Comprehension vs. persuasion: Does the framing help the user understand the situation, or accept a specific conclusion? Gigerenzer's natural frequencies are for understanding; the selective omission of risks is for persuading.

5. Autonomy: Does the framing preserve the user's capacity to reach their own conclusion, or does it predetermine it? Framing that expands the perceived options respects autonomy; framing that reduces them violates it.

6. Practical Application: The Case of the Financial Platform

FinSmart is an automated financial advisory platform that uses AI to recommend investment portfolios to retail users. Its interface presents each recommended portfolio with a prominent ‘expected annualized return,’ a chart of projected growth, and a small asterisk leading to a page of ‘risk warnings.’ The adoption rate for its recommendations is 78%—a success by the company's metric. But a behavioral audit reveals a systemic framing problem.

6.1 Framing Analysis

Attribute framing operates at multiple levels. The return is presented in positive terms (‘8.2% expected’) without the equivalent negative framing (‘91.8% probability that the return will differ from what is expected’). The chart shows only the line of projected growth, not the band of uncertainty. The risk warnings are separated from the decision by one click—a friction that 94% of users never overcome.

Goal framing uses gain exclusively: “Put your money to work,” “Grow your wealth.” Never loss framing: “Without diversification, you could lose 40% in a market downturn.” Since investing is a risk behavior, not a prevention behavior, gain framing is consistent with Levin's effect—but does it benefit the user, or the platform that collects commissions on active portfolios?

Risky choice framing presents the investment options exclusively in the territory of gains: “Portfolio A: expected return of 6%.” “Portfolio B: expected return of 10%.” Never: “Portfolio A: maximum historical loss of 15%.” “Portfolio B: maximum historical loss of 45%.” Framing in the territory of gains produces moderate risk aversion; but the user never confronts the territory of losses where the reality of the risk would become tangible.

“When a FinSmart user loses 30% of their investment during a market correction, the reaction is fury and distrust. Not because they were not warned—they were warned, in small print behind a click. But because the framing of the entire experience implicitly communicated to them that investing was safe, easy, and profitable. The legal warning existed; the behavioral framing annulled it. The platform complies with the law but fails the ethics of design.”

6.2 A Redesign Based on Ethical Framing

Integrate gain and loss on the same screen. Do not separate the expected return from the possible loss. Present both with the same visual prominence: “This portfolio has generated an average return of 8.2% a year. In its worst year, it lost 28%. Would you be comfortable with a temporary loss of that magnitude?”

Use natural frequencies, not percentages. Instead of “8.2% expected return,” show: “If you invest 10,000 euros, in a typical year you would have 10,820. In a bad year, you could have 7,200.” Absolute quantities make tangible what percentages abstract away.

Show bands of uncertainty, not projection lines. Replace the linear growth chart with one that shows the range of possible outcomes. The band of uncertainty communicates visually what the legal warning tries to say in words: that the future is uncertain.

Invert the friction. Instead of the risk sitting behind a click, make accepting the recommendation require confirming that the risk was understood. The productive friction of Topic 1 applied to framing: it is not that the user cannot invest; it is that they cannot invest without having confronted the complete framing.

Offer explicit multi-framing. Allow the user to change the framing of the presentation with a button: “View as opportunity / View as risk / View both.” This makes visible the fact that every framing is a choice, and returns that choice to the user.

6.3 Derived Design Principles

Principle 1: Present gain and loss together, never separately. Separation is a framing that favors action; integration is a framing that favors reflection.

Principle 2: Use the format that maximizes comprehension, not the one that maximizes persuasion. Natural frequencies, absolute quantities, and bands of uncertainty are ethical framings; isolated percentages and projection lines are persuasive framings.

Principle 3: Apply Sunstein's transparency test to every framing decision. If you could not explain why you chose that framing without embarrassment, choose another.

Principle 4: Offer multi-framing: the possibility of seeing the same information from multiple frames. This does not weaken the decision; it strengthens it.

Principle 5: Remember that omission is the most powerful framing. What you do not show shapes the decision as much as what you do show. Audit your omissions with the same rigor with which you audit your presentations.

7. Glossary of Key Terms

8. Final Reflection and Self-Assessment

Before moving on to the last topic of this module (Decisions Under Uncertainty), verify that you can diagnose, design, and ethically evaluate framing.

9. Bibliography and Recommended Readings

Essential Readings

• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. [Chapters 25–34]

• Thaler, R. & Sunstein, C. (2008). Nudge: Improving Decisions About Health, Wealth, and Happiness. Yale University Press.

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

Complementary Readings

• Levin, I.P., Schneider, S.L. & Gaeth, G.J. (1998). “All Frames Are Not Created Equal: A Typology and Critical Analysis of Framing Effects.” Organizational Behavior and Human Decision Processes, 76(2), 149–188.

• Sunstein, C. (2016). The Ethics of Influence: Government in the Age of Behavioral Science. Cambridge University Press.

• Tversky, A. & Kahneman, D. (1981). “The Framing of Decisions and the Psychology of Choice.” Science, 211(4481), 453–458.

• Lakoff, G. (2004). Don't Think of an Elephant! Chelsea Green Publishing. [Connects with the linguistic framing of Module IV]

Next Topic: The last book of this module addresses Decisions Under Uncertainty: how people formulate intentions and make decisions when outcomes are unknown, probabilities imprecise, and information incomplete. You will study with Kahneman, Tversky, and Gigerenzer the distinction between risk and uncertainty, ambiguity aversion, the ecological rationality of fast heuristics, and the design of systems that treat intention as provisional—not as a fixed commitment—and that build mechanisms for review, adaptation, and an elegant response to surprise.