Ingest: Metaphors We Think With: The Role of Metaphor in Reasoning

Type: types/ingest-report.md

Classification

Scientific paper (PLoS ONE, open access). Genre justified by the source's form: five controlled experiments with quantitative results, participant samples, coding methodology, and statistical analysis reported in a peer-reviewed venue.

Author: Paul H. Thibodeau and Lera Boroditsky. Credibility signal: published in PLoS ONE (2011) after peer review; funded by NSF; authors affiliated with Stanford University cognitive science community; no declared competing interests.

Summary

Thibodeau and Boroditsky (2011) report five experiments testing whether metaphorical framing affects how people reason about a complex social problem (crime) and gather information to solve it. Using a fictional city ("Addison") with escalating crime statistics, they contrasted two metaphors: crime as a virus (infecting, plaguing) versus crime as a beast (preying, lurking). The key finding is that even a single metaphorical word in the opening sentence ("Crime is a virus/beast ravaging the city of Addison") shifted participants' proposed solutions by 18-22% toward the metaphoric frame—larger than the opinion difference between Democrats and Republicans (8-9%). The effect was covert: participants consistently cited the crime statistics (identical across conditions) as influential, not the metaphor. Experiment 3 showed the effect requires an integrated metaphor within a knowledge structure, not mere lexical priming. Experiment 4 showed the effect persists when all solution options are explicitly listed, and influences information-foraging choices. Experiment 5 showed the effect depends on early presentation in the narrative—late-presented metaphors had no effect. The paper argues metaphors instantiate "frame-consistent knowledge structures" and invite "structurally-consistent inferences."

Contribution assessment (our opinion)

This paper contributes a clean empirical demonstration that minimal linguistic framing—a single word—can produce large, covert shifts in how people approach a problem and what information they seek. For the problem of understanding figurative wording's influence on reasoning, the paper provides probably the most directly relevant evidence in the literature: it isolates the manipulation (single noun), measures both solution proposals and information-seeking behavior, tests alternative mechanisms (lexical priming vs. knowledge-structure instantiation), and benchmarks effect size against real-world opinion divides. The contribution to the problem itself is the demonstration that metaphor's influence is not merely rhetorical but cognitive—it changes which knowledge structures become active and which inferences seem natural. The central claim is well-supported by the five-experiment design. The main limitation for KB purposes is that the domain is crime policy, a politically loaded topic whose framing effects may not generalize uniformly to all reasoning contexts; the paper's own discussion acknowledges this by reference to the pervasiveness of metaphor in everyday language, but the empirical scope is one domain. The paper does not test whether the effect transfers across domains or addresses counterfactual framing (what happens when a metaphor is explicitly challenged).

Quotes

In each of the five experiments, participants were presented with a survey that included a short paragraph about crime in the fictional city of Addison and some follow-up questions. The survey differed subtly between experiments, but always contrasted a crime-as-virus framing with a crime-as-beast framing. --- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Methods — Materials

Overall, participants were more likely to emphasize enforcement strategies (65%) than reform (35%), χ² = 41.85, p<.001. However, as predicted, the solutions participants proposed to the crime problem in Addison differed systematically as a function of the metaphorical frame encountered in the crime report (see Fig. 1). Participants given the crime-as-beast metaphorical framing were more likely to suggest enforcement (74%) than participants given the crime-as-virus framing (56%), χ² = 13.94, p<.001. See Table 1 for response frequencies.

--- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Experiment 1 — Results

Interestingly, when asked to identify the most influential aspect of the report, most participants ignored the metaphor. Only 15 participants (3%) identified the metaphoric frame as influential to their problem solving strategy. Removing these participants from the analysis did not affect the results (the proportion of responses that were congruent with the metaphor was not different in the two analyses, χ² = .0001, p = .991). The vast majority of the participants identified the statistics in the crime report as being most influential in their decision – namely, the final three sentences of the paragraph that state the increasing crime and murder rate. --- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Experiment 1 — Results

In Experiment 2 we tested this hypothesis by removing the relational verbs from the report. We replaced them with a single word metaphor that described crime as a “virus” or “beast” in the introductory sentence. The two conditions differed only in this one word, and otherwise included all the same information. --- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Experiment 2 — single-word manipulation

In Experiment 3 we tested whether the influence of the metaphor observed in the first two studies could have come about through simple spreading activation from lexical associates of the words “beast” and “virus.” We dissociated the words “beast” and “virus” from the story, so that they could act as non-metaphorical lexical primes. These disconnected lexical primes did not yield differences in people's crime-fighting suggestions. These results suggest that metaphors act as more than just isolated words – their power appears to come from participating in elaborated knowledge structures. --- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Experiment 3 — Discussion

However, unlike Experiment 4, there was no effect of the metaphorical frame. Participants who were presented with the crime-as-a-beast metaphor were about equally likely to gather additional information about the city's social situation (69%) as participants who were presented with the crime-as-a-virus metaphor (64%), χ² = .29, p = .59 (see Fig. 3). See Table 1 for response frequencies by condition.

--- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Experiment 5 — Results

In Experiment 5 we investigated whether when a metaphor is introduced affects the metaphor's influence. Experiment 5 repeated the design of Experiment 4, but we moved the metaphorical frame so that instead of being the first sentence in the crime report it was the last. Unlike the results of Experiment 4, this late metaphorical framing had no effect on people's crime-related information foraging. These findings suggest that metaphors can gain power by coercing further incoming information to fit with the relational structure suggested by the metaphor. --- kb/sources/.snapshots/metaphors-we-think-with.md @ sha256:d894f7412d22bf24c3084ade0d1f8f066b3d0fd964772c495b71a56d6d3fee3c — Experiment 5 — Discussion

Connections Found

The source connects to four theory notes and one existing source, with one additional source as a reverse-edge candidate:

• knowledge-storage-does-not-imply-contextual-activation — is-evidence-for: The paper provides empirical evidence that framing can instantiate "frame-consistent knowledge structures" and invite "structurally-consistent inferences," directly supporting the note's distinction between storage, read-back, and activation. The covert-effect finding (participants ignore the metaphor and cite statistics) is a clean demonstration of the "expert-witness pattern": knowledge is present but not self-triggering. The paper shows that a single word shifts reasoning by 18-22% without people recognizing the influence—a strong data point for the note's claim that "the model knows X" is often the wrong operational question.

• human-writing-structures-transfer-to-llms-because-failure-modes-overlap — is-evidence-for: The paper provides empirical evidence about content effects on reasoning—how the same information framed differently leads to systematically different conclusions. This supports the failure-mode overlap argument that LLMs share human-like reasoning vulnerabilities. The Experiment 3 dissociation (lexical prime vs. embedded metaphor) is relevant to understanding what kinds of human cognitive patterns transfer and at what mechanism depth: mere word activation is insufficient; the metaphor must participate in an elaborated knowledge structure.

• frontloading-is-partial-evaluation-not-divide-and-conquer — compares-with: The paper explicitly argues that metaphors are not "mere rhetorical flourishes" but have "profound influences on how we conceptualize and act"—directly relevant to the frontloading note's argument that the PE framing is "structurally precise, not metaphorical." The relationship is complementary: the metaphor paper shows metaphors do have real cognitive effects (so framing is never nothing), while the frontloading note clarifies that in a shared-token medium, some framings are precise substitutions rather than loose metaphors. A future ingest could note this tension: when does a framing's effect come from structural precision versus metaphorical knowledge-activation?

• prompt-ablation-converts-human-insight-to-deployable-framing — see-also: The metaphor paper is foundational empirical evidence for why framing matters. The prompt ablation methodology tests different framings to find which cognitive moves agents can reliably execute; the metaphor paper demonstrates that minimal framing changes (single word) produce significant behavioral shifts. Relevant as background evidence for why ablation studies are worth running and as a model of controlled framing experiments.

• gick-holyoak-analogical-problem-solving — compares-with (reverse edge): Both papers examine how mental representations affect problem-solving behavior. Gick & Holyoak show that memory alone doesn't guarantee analogical use (hinting is the differentiator); the metaphor paper shows that metaphors instantiate knowledge structures that guide reasoning even without awareness. Complementary evidence about the recognition/activation gap—Gick & Holyoak from the analogical-transfer side, the metaphor paper from the framing side. The metaphor paper's Experiment 3 (lexical prime vs. embedded metaphor) parallels Gick & Holyoak's finding that recall ≠ use.

The connect report also flagged human-llm-differences-are-load-bearing-for-knowledge-system-design as a see-also candidate (the covert-influence finding and political-affiliation interaction inform the naive anthropomorphism warning), and prompt-ablation as a see-also with the explore paper. These are retained as secondary connections.

Extractable Value

1. Empirical evidence for framing effects on human reasoning — The paper provides controlled experimental evidence that minimal linguistic framing (single word) produces large (18-22%), covert shifts in problem-solving strategy and information-seeking behavior. This bears directly on the planned note about figurative wording and reasoning. The virus/beast contrast and the covert-effect finding are reusable evidence for any KB claim about how wording shapes what reasoning strategies come to mind. [{quick-win}]

2. Covert-influence finding as evidence for the activation gap — Participants consistently cited the crime statistics (identical across conditions), not the metaphor, as the influential aspect of the report—yet the metaphor caused the shift. This is a clean data point for the KB's "expert-witness pattern" and the knowledge-storage-does-not-imply-contextual-activation distinction: knowledge/framing can be present without being self-triggering, and the gap is often unrecognized by the reasoner. [{quick-win}]

3. Experiment 3 dissociation: lexical priming vs. knowledge-structure instantiation — The result that a disconnected lexical prime ("beast"/"virus" presented alone, then a non-metaphorical crime report) produced no framing effect, while the embedded metaphor did, distinguishes mere word activation from structural knowledge instantiation. This supports the human-writing-structures-transfer note's mechanism-depth argument: not all human cognitive patterns are equally transferable; the relevant ones operate through elaborated structures, not surface associations. [{deep-dive}]

4. Experiment 5: early vs. late framing and narrative integration — The finding that early-presented metaphors have effect but late-presented ones do not (when both are near the measure) suggests metaphors work by coercing subsequent information into a frame, not merely by fresh-memory activation. This is relevant to the frontloading note's argument about structural precision: it provides evidence that when framing is introduced matters, and that narrative position can determine whether a framing operates through assimilation or fails. [{experiment}]

5. Effect-size benchmark against real-world opinion divides — The paper reports that metaphor-induced shifts (18-22%) exceed partisan differences (8-9%) and are larger than gender differences. This is a useful quantitative anchor for claims about framing magnitude and supports the prompt-ablation note's framing-selection methodology: if a single word can shift reasoning by more than partisan identity, framing choices in prompts are not cosmetic. [{quick-win}]

6. Mismatch with occasion: no direct evidence about LLM/figurative wording interaction — The paper studies human reasoning, not LLM behavior. It informs the planned note by providing human-reader evidence about how figurative wording affects reasoning, but does not test whether the same effects hold for language models. If the note's scope includes LLM-specific claims, this source supports only the human-side mechanistic background. [{just-a-reference}]

Limitations (our opinion)

The source should not be over-generalized in these ways:

• Single domain (crime policy). The experiments use crime in a fictional city as the test domain. The paper's own discussion argues metaphor is pervasive in everyday language and social policy discourse, but the empirical scope is one domain. Whether the effect magnitude transfers to other reasoning contexts (technical problem-solving, scientific reasoning, everyday decisions) is not tested. The covert-effect and knowledge-structure claims are plausible across domains, but the specific numbers are domain-bound.

• Framing, not figurative wording broadly. The paper contrasts two specific conceptual metaphors (virus vs. beast) that invite different knowledge structures. It does not test other forms of figurative language (metonymy, irony, personification) or non-metaphorical figurative wording. Claims about "figurative wording" more broadly than conceptual metaphor framing exceed what this source establishes.

• Covert effect is demonstrated, not explained at the cognitive-architecture level. The paper shows the effect is covert (participants don't recognize it) and argues it operates through frame-consistent knowledge structures, but does not provide a detailed cognitive model of how those structures are built, maintained, or interfered with. The mechanism remains at the level of "analogical transfer" and "elaborated knowledge structures"—useful but not architecturally precise.

• Not an LLM study. The paper says nothing about language models. Any KB claim that transfers these findings to LLM behavior (e.g., "LLMs show covert framing effects like humans") would be extrapolation beyond the source, not evidence from it. The human-writing-structures-transfer note's argument about failure-mode overlap is supported by this paper only insofar as it establishes that humans have these framing vulnerabilities; the LLM-side claim requires separate evidence.

• Capture boundary: full-source. The snapshot is a full-source capture, so the limitation is not from incomplete capture. The limitation is inherent to the paper's scope.

Recommended Next Action

Promote this ingest's extractable value items 1 and 2 into the planned Commonplace note on figurative wording and reasoning as empirical evidence for framing effects and the covert-activation gap. Specifically, cite the single-word manipulation result (18-22% shift) and the statistics-vs-metaphor attribution finding as supporting evidence for the note's claims about how wording shapes reasoning strategies and why the influence is often unrecognized. File the full ingest as a source-only reference for the remaining items (Experiment 3 dissociation, Experiment 5 timing, effect-size benchmark) pending a decision on whether the note needs that level of mechanistic detail. If the planned note's scope expands to include LLM-specific claims, return to this ingest to assess whether the human-side evidence is sufficient grounding or whether an LLM-framing study is now required.