Framing Effect
The tendency for decisions and judgments to change based on how information is presented (framed) rather than on the underlying substance. Logically equivalent descriptions can produce systematically different choices when one emphasizes gains vs. losses, or uses different reference points.
Origin
Tversky & Kahneman (1981), “The Framing of Decisions and the Psychology of Choice,” Science, 211(4481), 453–458. The classic demonstration:
- Gain frame: “This treatment saves 200 of 600 people.” → 72% choose it
- Loss frame: “This treatment results in 400 deaths out of 600.” → 22% choose it
Logically identical. The reference point (lives saved vs. lives lost) reverses the majority preference.
Mechanism
Rooted in Prospect Theory (Kahneman & Tversky, 1979) — people evaluate outcomes relative to a reference point rather than in absolute terms, and losses loom larger than equivalent gains (loss aversion). Framing manipulates the reference point.
Types of framing effects:
- Risky choice framing — gain vs. loss descriptions of the same gamble
- Attribute framing — describing the same attribute positively or negatively (“95% fat-free” vs. “5% fat”)
- Goal framing — positive outcomes of acting vs. negative outcomes of not acting
- Issue framing — which aspects of a complex issue are made salient
Intelligence Analysis Context
Framing is particularly pernicious in intelligence analysis because:
- Question framing: how a policymaker asks the question shapes what the analyst considers relevant
- Historical analogy framing: “is this like Munich?” vs. “is this like WWI?” activates different mental models for the same situation
- Threat framing vs. opportunity framing: the same foreign behavior read as aggressive or defensive depending on the analytic frame
The CIA Tradecraft Primer (2009) addresses framing implicitly through its emphasis on mind-sets (frames) and the explicit alternative-generation techniques.
LLM Agentic Systems Context
LLMs are extraordinarily susceptible to framing effects:
- Prompt framing dominates: the linguistic framing of a prompt is the primary determinant of LLM output; the same underlying task produces qualitatively different responses based on framing alone
- Persona framing: assigning a role (“you are a skeptical critic” vs. “you are a supportive mentor”) systematically shifts output independent of task
- Valence framing: asking “what’s wrong with X?” vs. “what’s right with X?” produces strongly asymmetric responses even when balanced analysis would give the same facts
- Reference point sensitivity: LLMs evaluate “good” and “bad” relative to implicit reference points established in the prompt
See SATs for LLM Agents for SAT-based mitigations.
SATs That Control For This Bias
- Outside-In Thinking — deliberately reframes the problem from the outside environment inward, changing the reference point
- Brainstorming — generating alternatives before settling on a frame prevents premature frame lock-in
- Alternative Futures Analysis — builds multiple scenarios from different starting frames, making the frame-dependency of any single scenario visible
- Key Assumptions Check — the current analytic “frame” is a hidden assumption; making it explicit allows it to be challenged
Key References
- Tversky, A. & Kahneman, D. (1981). “The Framing of Decisions and the Psychology of Choice.” Science, 211(4481), 453–458.
- Kahneman, D. & Tversky, A. (1979). “Prospect Theory: An Analysis of Decision under Risk.” Econometrica, 47(2), 263–291.
- Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. (Chapter 26: “Prospect Theory”)
Empirical Evidence (LLM)
| Study | Finding |
|---|---|
| Echterhoff et al. (BiasBuster, 2024) | Direct measurement of framing effects across LLMs — model responses shift with prompt framing even when underlying facts are identical. Mitigatable via self-debiasing prompts. |