Mirror Imaging
The tendency to assume that adversaries, foreign actors, or other agents think, value, and make decisions the same way the analyst does. In psychology: the Fundamental Attribution Error applied cross-culturally and cross-institutionally.
Origin and Names
Fundamental Attribution Error (Ross, 1977): the general cognitive bias of attributing others’ behavior to their character/nature while attributing one’s own behavior to situational factors.
Mirror Imaging is the intelligence community’s term for the specific manifestation: analysts project their own decision-making frameworks onto foreign actors, assuming rational calculation matches their own rationality norms, risk tolerances, and value hierarchies.
Named in CIA Tradecraft Primer (2009)‘s causality bias taxonomy:
“Behavior of others is attributed to some fixed nature of the person or country, while our own behavior is attributed to the situation in which we find ourselves.”
Intelligence Analysis Context
Mirror imaging produces systematic failures in adversary analysis:
- Rationality projection: assuming adversaries weigh costs/benefits as the analyst would; missing that an adversary may have different values, time horizons, or risk tolerance
- Motivation blindness: failing to consider that an action that would be irrational for “us” may be rational given the adversary’s constraints and goals
- Historical examples: analysts consistently underestimated Soviet willingness to take aggressive action because the risk seemed unacceptable by Western rationality norms (Cuban Missile Crisis is partly a mirror imaging failure in the opposite direction — Khrushchev’s belief the US would accept a fait accompli)
LLM Agentic Systems Context
LLMs exhibit a specific form of mirror imaging rooted in their training:
- Value alignment projection: LLMs trained on human feedback project the values of their RLHF raters onto all users and all modeled agents — they assume universally “reasonable” goals
- User perspective capture: agents frequently adopt the user’s framing as the correct framing without stepping back to consider whether the user’s mental model is accurate
- Persona collapse: agents asked to roleplay or model adversarial actors often sanitize or rationalize the adversary’s behavior through the agent’s own value system
- Homogeneity in multi-agent systems: multiple instances of the same model “mirror image” each other, producing echo chambers rather than genuine independent analysis
See SATs for LLM Agents for SAT-based mitigations.
SATs That Control For This Bias
- Red Team Analysis — directly and structurally requires adopting the adversary’s perspective; the entire technique exists to counter mirror imaging
- Outside-In Thinking — expands the frame beyond the analyst’s natural reference point to include external constraints and logics
- Alternative Futures Analysis — generates scenarios driven by adversary logic, not analyst preferences
- Devil’s Advocacy — builds the strongest possible case from a different set of premises
Key References
- Ross, L. (1977). “The intuitive psychologist and his shortcomings: Distortions in the attribution process.” In L. Berkowitz (Ed.), Advances in Experimental Social Psychology, Vol. 10, pp. 173–220. Academic Press.
- Heuer, R. J. Jr. (1999). The Psychology of Intelligence Analysis, Chapter 4: “Strategies for Analytical Judgment.” Richards j. heuer jr.
- Chang, W., et al. (2018). “Lessons from the Intelligence Community on Improving Forecasting.” Perspectives on Psychological Science.
- CIA Tradecraft Primer (2009) — attribution bias in causality taxonomy
Empirical Evidence (LLM)
| Study | Finding |
|---|---|
| Durmus et al. (2023) | LLM default responses align most with USA / Western European opinions (GlobalOpinionQA benchmark, derived from Pew + World Values Survey). Country-conditioning shifts opinion but introduces stereotyping. Translation alone does not shift opinion alignment. |
Implication for SATs: Red Team and Outside-In Thinking in an LLM context are operating on top of WEIRD priors. Persona prompts work imperfectly. Adversary modeling tends to produce sanitized Western-flavored adversaries by default. See H4.
Case Studies
- Iraq WMD 2003. The Kerr Report (cited in Grey Dynamics 2024) found US satellite imagery analysis “failed to acknowledge the political/cultural context” of Saddam’s actions — a direct instance of mirror imaging at the collection-strategy level. US intelligence had zero HUMINT sources in Iraq from 1998–2003, meaning analysis ran on overhead imagery interpreted through Western assumptions about a leader operating with very different incentives and information.
See Also
Cognitive Bias | Groupthink | Motivated Reasoning | Red Team Analysis