# Kim, Wang, Janssen and Anderies: How Large Language Models Misrepresent American Climate Opinions

[Canonical reading](https://escapemeridian.com/readings/kim-wang-janssen-and-anderies-how-large-language-models-ff5afa2a)

Sola Kim, Jieshu Wang, Marco A. Janssen, and John M. Anderies argue that six large language models (LLMs) misrepresent U.S. climate opinion when used as “silicon samples”—simulated stand-ins for survey respondents. They compared model responses with the answers of 978 respondents to a nationally representative U.S. climate opinion survey across 20 questions. The authors report that the models pull predictions toward the population mean: less-concerned groups are predicted as more concerned and more-concerned groups as less, while predicted response variance is 29% lower on average than human response variance. They also report an intersectional pattern: models predict women as more pro-climate than men in every racial group, but the direction of the human pattern differs for Black Americans. The authors say an audit examining gender alone would miss that reversal. Their central concern is representation as well as accuracy: which groups’ views a synthetic sample can make legible. The authors qualify the direction, not the size, of the Black American gender difference because it rests on small cells at the ideological extreme. The paper’s v3 revision corrects normalization and revises the variance-compression figure from 27.9% to 29%.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

## Original source

arXiv cs.AI · Article

Original source date: Dec 29, 2025 (2025-12-29).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-20T07:33:49.984Z. First observed: 2026-11-20T05:04:44.315Z. Retrieved: 2026-11-20T07:32:57.668Z.

- [How Large Language Models Misrepresent American Climate Opinions](https://arxiv.org/abs/2512.23889)

## Limits

The abstract reports results for six models, 978 survey respondents, and 20 questions; it does not establish how the findings generalize beyond this study. The authors state that the direction, but not the size, of the gender difference for Black Americans is claimed because it rests on small cells at the ideological extreme. The evidence provided is the paper’s abstract and publisher page, not its full methods or results.


## Open question

How would the reported race-specific gender pattern change with larger samples in the ideological-extreme groups, where the authors say the result rests on small cells?


## Source-stated

The authors compared six large language models’ answers with those of 978 respondents across 20 questions in a nationally representative U.S. climate opinion survey.

> We prompted six large language models with the profiles of 978 respondents to a nationally representative U.S. climate opinion survey and compared the generated responses with those respondents' own answers across 20 questions.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors report that model predictions move toward the population mean, with predicted response variance averaging 29% below human response variance.

> the variance of predicted responses falls, on average, 29% below that of human responses.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors report that the models predict women as more pro-climate than men in every racial group, while the direction of the human pattern differs for Black Americans.

> The models apply one gender assumption, predicting women as more pro-climate than men in every racial group. The assumption matches the direction of the human pattern for White, Hispanic, and Other Americans and runs the wrong way for Black Americans.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors argue that an audit considering gender alone would miss the change in direction across racial groups.

> An audit of gender alone would report that the models understate the gender gap, and would miss that the error changes direction by race, which is what single-variable audits cannot detect.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors qualify the direction, but not the size, of the gender difference for Black Americans because the finding rests on small cells at the ideological extreme.

> We claim the direction of that difference rather than its size, since it rests on small cells at the ideological extreme.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors frame the issue as representation as well as accuracy, asking whose views a synthetic sample renders legible.

> What is at stake is not accuracy alone but representation: whose views a synthetic sample renders legible.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The v3 revision changed the reported variance compression from 27.9% to 29% after a normalization correction.

> a normalization correction scales items by instrument range rather than observed range, revising the reported variance compression from 27.9% to 29%

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source and contributor context

Kim, Sola, Wang, Jieshu, Janssen, Marco A., Anderies, John M. · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Figure

Figure develops humanoid robots and publishes its own tests of how they perform physical tasks.

- [Explore this context](/entities/figure)
- [Reference](https://www.figure.ai/news)
- [Reference](https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization)

## Editorial provenance

Version 1. Publication ID: 689242ad-e1a0-43f4-8728-f64a0951a2e8.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-20T07:33:49.984Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: cfae11e87f0388cea3d905bb9d0d47427356b0b41181e0cc402ed807f5ad820c. Source text hash: 3ea0c1d602327fe3d566717b8737e649fe97c9f89feac9c8342f381e7884a23d.


## Revision and source-check history

### Version 1 — 2026-11-20T07:33:49.984Z

Sola Kim, Jieshu Wang, Marco A. Janssen, and John M. Anderies argue that six large language models (LLMs) misrepresent U.S. climate opinion when used as “silicon samples”—simulated stand-ins for survey respondents. They compared model responses with the answers of 978 respondents to a nationally representative U.S. climate opinion survey across 20 questions. The authors report that the models pull predictions toward the population mean: less-concerned groups are predicted as more concerned and more-concerned groups as less, while predicted response variance is 29% lower on average than human response variance. They also report an intersectional pattern: models predict women as more pro-climate than men in every racial group, but the direction of the human pattern differs for Black Americans. The authors say an audit examining gender alone would miss that reversal. Their central concern is representation as well as accuracy: which groups’ views a synthetic sample can make legible. The authors qualify the direction, not the size, of the Black American gender difference because it rests on small cells at the ideological extreme. The paper’s v3 revision corrects normalization and revises the variance-compression figure from 27.9% to 29%.

## Original source

arXiv cs.AI · Article

Original source date: Dec 29, 2025 (2025-12-29).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-20T07:33:49.984Z. First observed: 2026-11-20T05:04:44.315Z. Retrieved: 2026-11-20T07:32:57.668Z.

- [How Large Language Models Misrepresent American Climate Opinions](https://arxiv.org/abs/2512.23889)

## Limits

The abstract reports results for six models, 978 survey respondents, and 20 questions; it does not establish how the findings generalize beyond this study. The authors state that the direction, but not the size, of the gender difference for Black Americans is claimed because it rests on small cells at the ideological extreme. The evidence provided is the paper’s abstract and publisher page, not its full methods or results.


## Open question

How would the reported race-specific gender pattern change with larger samples in the ideological-extreme groups, where the authors say the result rests on small cells?


## Source-stated

The authors compared six large language models’ answers with those of 978 respondents across 20 questions in a nationally representative U.S. climate opinion survey.

> We prompted six large language models with the profiles of 978 respondents to a nationally representative U.S. climate opinion survey and compared the generated responses with those respondents' own answers across 20 questions.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors report that model predictions move toward the population mean, with predicted response variance averaging 29% below human response variance.

> the variance of predicted responses falls, on average, 29% below that of human responses.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors report that the models predict women as more pro-climate than men in every racial group, while the direction of the human pattern differs for Black Americans.

> The models apply one gender assumption, predicting women as more pro-climate than men in every racial group. The assumption matches the direction of the human pattern for White, Hispanic, and Other Americans and runs the wrong way for Black Americans.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors argue that an audit considering gender alone would miss the change in direction across racial groups.

> An audit of gender alone would report that the models understate the gender gap, and would miss that the error changes direction by race, which is what single-variable audits cannot detect.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors qualify the direction, but not the size, of the gender difference for Black Americans because the finding rests on small cells at the ideological extreme.

> We claim the direction of that difference rather than its size, since it rests on small cells at the ideological extreme.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The authors frame the issue as representation as well as accuracy, asking whose views a synthetic sample renders legible.

> What is at stake is not accuracy alone but representation: whose views a synthetic sample renders legible.

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source-stated

The v3 revision changed the reported variance compression from 27.9% to 29% after a normalization correction.

> a normalization correction scales items by instrument range rather than observed range, revising the reported variance compression from 27.9% to 29%

- [Supporting source](https://arxiv.org/abs/2512.23889)

## Source and contributor context

Kim, Sola, Wang, Jieshu, Janssen, Marco A., Anderies, John M. · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.


## Source media

No suitable media was found by the supported page extraction.


## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Large language model

A language model with many learned numerical settings, used to process and generate language.

- [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model)

## arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

- [Explore this context](/entities/arxiv)
- [Reference](https://arxiv.org/)
- [Reference](https://info.arxiv.org/help/api/index.html)

## Figure

Figure develops humanoid robots and publishes its own tests of how they perform physical tasks.

- [Explore this context](/entities/figure)
- [Reference](https://www.figure.ai/news)
- [Reference](https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization)

## Editorial provenance

Version 1. Publication ID: 689242ad-e1a0-43f4-8728-f64a0951a2e8.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-20T07:33:49.984Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material\_development.

Observation: cfae11e87f0388cea3d905bb9d0d47427356b0b41181e0cc402ed807f5ad820c. Source text hash: 3ea0c1d602327fe3d566717b8737e649fe97c9f89feac9c8342f381e7884a23d.


### uncertain — 2026-11-20T07:33:49.984Z — anthropic/claude-opus-5.5

Publication: 689242ad-e1a0-43f4-8728-f64a0951a2e8

attribution — pass: The reading correctly credits Sola Kim, Jieshu Wang, Marco A. Janssen and John M. Anderies. It frames the findings as the authors' reports and arguments, not as established fact.

> Authors:Sola Kim, Jieshu Wang, Marco A. Janssen, John M. Anderies

summary — uncertain: The summary is mostly faithful and readable. It explains silicon samples, compression toward the mean and the intersectional reversal, and keeps the authors' representation framing. The question follows from the stated small-cell caveat. However, the passage 'The authors qualify the direction, not the size' can read as hedging the direction, which inverts the source. The authors claim the direction and decline to claim the size. Smallest repair: 'The authors claim only the direction, not the size, of the Black American gender difference.' The term 'intersectional' may also benefit from a tooltip.

> We claim the direction of that difference rather than its size, since it rests on small cells at the ideological extreme.

date — pass: originalPublished 2025-12-29 matches the v1 submission date. It is not the v3 revision date of 19 Nov 2026 or the later feed date.

> \[Submitted on 29 Dec 2025 (v1), last revised 19 Nov 2026 (this version, v3)\]

support — uncertain: Most claims are verbatim-supported, including the six models, 978 respondents, 20 questions, 29% variance compression, the gender pattern by race, the limits of single-variable audits, the representation framing and the v3 correction from 27.9% to 29%. The claim that the authors 'qualify the direction, but not the size' is ambiguous and risks reversing the source's scope. Rewording it to 'claim the direction, but not the size' fixes this. The limitations section states it correctly. No unsupported benefits or costs are implied.

> revising the reported variance compression from 27.9% to 29% and all coefficient units

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.

## Recorded references

Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter.

No public references recorded here yet.

[Reference data (JSON)](/readings/ff5afa2a43bdd9ca6cd3f6f080faa9fd878b795e53755e918baa68b37ad52d4c/references.json)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/kim-wang-janssen-and-anderies-how-large-language-models-ff5afa2a.json)
