Framing the Narrative: Ideological Mimicry in Large Language Models

Olivia Macmillan-Scott; Michael Jacobs; Nils Metternich; Mirco Musolesi

cs.CLSep 29, 2026

Framing the Narrative: Ideological Mimicry in Large Language Models

Authors: Olivia Macmillan-Scott, Michael Jacobs, Nils Metternich, Mirco Musolesi

Organizations: Centre for AI, Department of Computer Science, University College London · Public Policy Group, ETH Zürich · Department of Political Science, University College London · Department of Computer Science and Engineering, University of Bologna

Abstract

Large language models (LLMs) are increasingly used to answer questions about politically contentious issues, yet evaluations typically treat a model's stance as a relatively stable property. Real users, however, communicate political signals through their terminology, assumptions, and personal context. We investigate whether such signals produce ideological mimicry: systematic shifts in the political stance expressed by an LLM toward the position conveyed by the interaction. If LLMs adapt their responses to these signals, they risk creating personalised political information environments in which users with opposing views receive systematically different accounts of the same issue, potentially reinforcing existing divisions. We build the Poli-SHIFT dataset and evaluation framework and assess seven open-weight LLMs across ten contentious political topics in the United States, United Kingdom, and Australia, systematically manipulating contested terminology, politically valenced premises, and user information, and eliciting responses in both multiple-choice and open-text formats. Across models, we find robust evidence that prompt framing shapes the political stance of LLM outputs. Changing terminology alone reverses which side of an issue a model supports in 16.9% of matched comparisons. Stated political ideology also systematically shifts responses toward the user's position. These findings show that political stance is not a fixed property of LLMs; the views expressed are conditional on the interaction with the user. As LLMs become increasingly personalised sources of information, such interaction-dependent adaptation could contribute to political information environments that reinforce users' existing perspectives.

Figures & tables

Figure 1: Overview of the Poli-SHIFT experimental pipeline. The treatment conditions are presented along with potential underlying mechanisms that motivated each condition.

Topic	AU	UK	US
Climate Change Policy	✓	✓	✓
Immigration & Asylum Policy	✓	✓	✓
LGBT+ Rights	✓	✓	✓
Healthcare & Public Services	✓	✓	✓
Israel-Palestine War	✓	✓	✓
Russia-Ukraine War	✓	✓	✓

Table 1: Topics covered by country.

Figure 2: Mean political stance under pro-side and anti-side contested-terminology wording, shown separately for MCQ (left) and open-text (right) conditions, with 95% confidence intervals. The separation between the two dots in each row represents the directional terminology-framing effect.

Figure 3: Mean deviation from the no-biography survey baseline by stated user ideology and model, with 95% confidence intervals. Liberal biographies shift responses toward the designated pro-side position, while conservative biographies shift responses toward the anti-side position.

Appendix figures & tables36 assets

Supplementary material from the paper’s appendix.

Appendix

Pro-side	The UK should provide military assistance to the Ukrainian government in response to Russia’s unprovoked invasion .
Anti-side	The UK should provide military assistance to the Kiev regime in response to Russia’s special military operation .

Appendix

Table 2: Matched contested-terminology pair, Russia-Ukraine topic (United Kingdom), MCQ format. The pro-side and anti-side versions ask the identical substantive question; only the referring terms differ ( underlined ).

Pro-side	Immigrants contribute more in taxes than they receive in public benefits in the U.S.
Anti-side	Immigration depresses wages and takes jobs from American workers.

Appendix

Table 3: Example premise-framing items, immigration topic (United States), MCQ format. The two items are independent formulations, not a lexical substitution pair.

Originally specified	Collected
Gemma 2	Gemma 2 27B & Gemma 3 27B
Llama 3	Llama 3.1 8B & Llama 4 Scout
Falcon 3	Falcon 3 10B
DeepSeek Qwen	DeepSeek V4 Flash & Qwen 3.6 27B
Command R+	Command A
Mistral	Mistral Large

Appendix

Table 4: Mapping from the pre-registration’s originally specified model list to the models actually collected. All nine models were collected; seven (excluding Gemma 2 27B and Llama 3.1 8B) form the main-analysis set used throughout the paper, with the two excluded models retained for the generational comparison in Appendix Q .

Model	Format	Treatment	Effect	95% CI	$t$	$p$
Qwen 3.6 27B	MCQ	premise	+1.26	[+1.10, +1.43]	16.17	$<$ 0.001*
Falcon 3 10B	MCQ	premise	+1.06	[+0.86, +1.25]	11.04	$<$ 0.001*
Gemma 3 27B	MCQ	premise	+0.95	[+0.68, +1.22]	7.30	$<$ 0.001*
Llama 4 Scout	MCQ	premise	+0.63	[+0.41, +0.86]	5.92	$<$ 0.001*
Mistral Large	MCQ	premise	+0.59	[+0.29, +0.89]	4.04	$<$ 0.001*
DeepSeek V4 Flash	MCQ	premise	+0.46	[+0.23, +0.69]	4.20	$<$ 0.001*

Appendix

Table 5: Directional framing effects across all seven models, response formats, and framing treatments. Effect is the mean pro-side minus anti-side stance difference on the five-point scale; positive values indicate movement in the direction predicted by the political signal. 95% confidence intervals and paired

t

-tests are reported; * indicates

p<0.05

Figure 4: Mean political stance under pro-side and anti-side premise framing for each model in the MCQ condition.

Figure 5: Mean open-text political stance under pro-side and anti-side framing for contested terminology (left) and premise framing (right), by model. Larger separation between conditions indicates greater sensitivity to the corresponding political signal.

Figure 6: Stance-reversal rate for matched contested-terminology items, by model and response format. A reversal occurs when changing only the contested terminology moves the model’s response across the neutral midpoint to the opposing political side.

Figure 7: Mean deviation in political stance relative to the survey baseline, by model, framing treatment, and treatment direction. Positive values indicate movement toward the designated pro-side position; negative values indicate movement toward the anti-side position.

Figure 8: Consistency between MCQ and open-text stance measurements for items collected in both formats. Left: Pearson correlation between matched stance scores by model. Right: proportion of matched responses falling on the same side of the neutral midpoint.

Model	Attribute	Partial $\eta^{2}$	df	$F$	$p$	$N$	$R^{2}$
Command A	Ideology	0.151	2	25634.8	$<$ 0.001*	288,341	0.24
	Gender	0.019	2	2840.4	$<$ 0.001*
	Race	0.010	9	311.6	$<$ 0.001*
	Education	0.004	4	301.2	$<$ 0.001*
	Age	0.001	5	76.1	$<$ 0.001*
DeepSeek V4 Flash	Ideology	0.124	2	22107.8	$<$ 0.001*	312,304	0.18

Appendix

Table 6: Relative contribution of user-biography attributes to expressed political stance, by model. Partial

\eta^{2}

values are obtained from model-specific OLS regressions controlling for topic and country and are ranked within each model; larger values indicate greater explanatory contribution.

Model	Conservative	Liberal
Command A	-0.24* [-0.25, -0.23]	+0.59* [+0.58, +0.59]
DeepSeek V4 Flash	-0.63* [-0.64, -0.62]	+0.52* [+0.51, +0.53]
Gemma 3 27B	-0.57* [-0.58, -0.57]	+0.90* [+0.89, +0.90]
Llama 4 Scout	-0.61* [-0.61, -0.60]	+0.51* [+0.50, +0.52]
Mistral Large	-0.26* [-0.27, -0.26]	+0.71* [+0.71, +0.72]
Qwen 3.6 27B	-0.60* [-0.61, -0.60]	+0.87* [+0.86, +0.88]

Appendix

Table 7: Regression coefficients for ideology (reference category: ideologically neutral), with 95% confidence intervals, from the same OLS models reported in Table 6 . * indicates

p<0.05

Model	20	30	50	60	70
Command A	+0.06* [+0.04, +0.07]	+0.03* [+0.02, +0.04]	-0.02* [-0.03, -0.01]	-0.03* [-0.04, -0.02]	-0.02* [-0.03, -0.01]
DeepSeek V4 Flash	+0.05* [+0.03, +0.06]	+0.03* [+0.01, +0.04]	+0.00 [-0.01, +0.02]	-0.01 [-0.03, +0.00]	-0.01 [-0.02, +0.01]
Gemma 3 27B	+0.05* [+0.04, +0.06]	+0.03* [+0.02, +0.04]	-0.03* [-0.03, -0.02]	-0.04* [-0.05, -0.03]	-0.05* [-0.06, -0.04]
Llama 4 Scout	+0.03* [+0.02, +0.04]	+0.02* [+0.01, +0.03]	-0.01* [-0.02, -0.00]	-0.01 [-0.02, +0.00]	-0.03* [-0.04, -0.01]
Mistral Large	+0.03* [+0.02, +0.03]	+0.02* [+0.01, +0.02]	-0.01 [-0.02, +0.00]	-0.02* [-0.02, -0.01]	-0.02* [-0.03, -0.01]
Qwen 3.6 27B	+0.06* [+0.05, +0.07]	+0.02* [+0.01, +0.04]	-0.02* [-0.04, -0.01]	-0.05* [-0.06, -0.04]	-0.07* [-0.08, -0.06]

Appendix

Table 8: Regression coefficients for age (reference category: 40), with 95% confidence intervals, from the same OLS models reported in Table 6 . * indicates

p<0.05

Model	No Formal Education	Technical Certification	Undergraduate Degree	Postgraduate Degree
Command A	-0.00 [-0.01, +0.01]	+0.05* [+0.04, +0.06]	+0.08* [+0.07, +0.09]	+0.14* [+0.13, +0.15]
DeepSeek V4 Flash	-0.04* [-0.05, -0.02]	+0.07* [+0.06, +0.09]	+0.09* [+0.08, +0.11]	+0.13* [+0.11, +0.14]
Gemma 3 27B	-0.12* [-0.13, -0.11]	-0.02* [-0.03, -0.01]	+0.00 [-0.01, +0.01]	+0.03* [+0.02, +0.04]
Llama 4 Scout	-0.07* [-0.08, -0.06]	+0.09* [+0.08, +0.10]	+0.04* [+0.03, +0.05]	+0.08* [+0.07, +0.09]
Mistral Large	-0.03* [-0.03, -0.02]	+0.04* [+0.03, +0.05]	+0.05* [+0.04, +0.06]	+0.09* [+0.08, +0.10]
Qwen 3.6 27B	+0.01 [-0.01, +0.02]	+0.06* [+0.05, +0.07]	+0.04* [+0.03, +0.06]	+0.09* [+0.08, +0.10]

Appendix

Table 9: Regression coefficients for education (reference category: secondary education), with 95% confidence intervals, from the same OLS models reported in Table 6 . * indicates

p<0.05

Model	Woman	Non-Binary Person
Command A	+0.08* [+0.08, +0.09]	+0.28* [+0.27, +0.28]
DeepSeek V4 Flash	+0.11* [+0.10, +0.12]	+0.19* [+0.18, +0.21]
Gemma 3 27B	+0.15* [+0.15, +0.16]	+0.27* [+0.26, +0.27]
Llama 4 Scout	+0.11* [+0.10, +0.11]	+0.19* [+0.18, +0.20]
Mistral Large	+0.12* [+0.12, +0.13]	+0.24* [+0.23, +0.24]
Qwen 3.6 27B	+0.09* [+0.08, +0.10]	+0.16* [+0.16, +0.17]

Appendix

Table 10: Regression coefficients for gender (reference category: man), with 95% confidence intervals, from the same OLS models reported in Table 6 . * indicates

p<0.05

Model	Arab	Black	Central Asian	East Asian	Hispanic	Indigenous	Jewish	South Asian	Mixed Race
Command A	-0.07* [-0.08, -0.06]	+0.08* [+0.07, +0.10]	-0.17* [-0.19, -0.16]	-0.02* [-0.04, -0.01]	-0.03* [-0.04, -0.01]	+0.07* [+0.06, +0.09]	+0.11* [+0.10, +0.12]	+0.03* [+0.02, +0.05]	+0.06* [+0.05, +0.08]
DeepSeek V4 Flash	-0.04* [-0.06, -0.02]	+0.13* [+0.11, +0.15]	-0.07* [-0.09, -0.05]	+0.05* [+0.03, +0.07]	+0.03* [+0.01, +0.05]	+0.08* [+0.06, +0.10]	+0.22* [+0.20, +0.24]	+0.06* [+0.04, +0.08]	+0.10* [+0.08, +0.12]
Gemma 3 27B	+0.02* [+0.01, +0.03]	+0.13* [+0.12, +0.15]	-0.07* [-0.08, -0.05]	+0.04* [+0.03, +0.05]	-0.02* [-0.04, -0.01]	+0.14* [+0.12, +0.15]	+0.17* [+0.16, +0.19]	+0.08* [+0.07, +0.09]	+0.09* [+0.08, +0.10]
Llama 4 Scout	+0.14* [+0.13, +0.15]	+0.15* [+0.13, +0.16]	+0.01 [-0.00, +0.03]	+0.09* [+0.07, +0.10]	+0.02* [+0.01, +0.03]	+0.21* [+0.19, +0.22]	+0.21* [+0.19, +0.22]	+0.12* [+0.11, +0.14]	+0.03* [+0.02, +0.05]
Mistral Large	+0.10* [+0.08, +0.11]	+0.11* [+0.10, +0.12]	-0.03* [-0.04, -0.02]	+0.05* [+0.04, +0.06]	+0.04* [+0.02, +0.05]	+0.16* [+0.15, +0.17]	+0.21* [+0.20, +0.22]	+0.11* [+0.10, +0.13]	+0.08* [+0.07, +0.09]
Qwen 3.6 27B	-0.02* [-0.04, -0.01]	+0.08* [+0.06, +0.09]	-0.08* [-0.10, -0.06]	+0.03* [+0.02, +0.05]	+0.02* [+0.00, +0.03]	+0.13* [+0.11, +0.14]	+0.16* [+0.14, +0.17]	+0.07* [+0.06, +0.09]	+0.06* [+0.04, +0.07]

Appendix

Table 11: Regression coefficients for race (reference category: White), with 95% confidence intervals, from the same OLS models reported in Table 6 . * indicates

p<0.05

Figure 9: Mean political stance by stated user ideology and model. Liberal biographies shift responses toward the designated pro-side position, while conservative biographies shift responses toward the anti-side position.

Figure 10: Estimated relationship between stated user ideology and political stance by model after controlling for topic and country. Figure 3 shows the corresponding unadjusted comparison.

Figure 11: Signed deviation from the no-biography control for each of the five user-biography attributes in Command A. Positive values indicate a more pro-side stance than the control condition and negative values a more anti-side stance.

Model	Format	Treatment	Effect	Primary $p$	Clustered SE	Clustered $p$
Qwen 3.6 27B	MCQ	premise	+1.26	$<$ 0.001*	0.083	$<$ 0.001*
Falcon 3 10B	MCQ	premise	+1.06	$<$ 0.001*	0.125	$<$ 0.001*
Gemma 3 27B	MCQ	premise	+0.95	$<$ 0.001*	0.197	$<$ 0.001*
Llama 4 Scout	MCQ	premise	+0.63	$<$ 0.001*	0.147	$<$ 0.001*
Mistral Large	MCQ	premise	+0.59	$<$ 0.001*	0.227	0.009*
DeepSeek V4 Flash	MCQ	premise	+0.46	$<$ 0.001*	0.187	0.014*

Appendix

Table 12: Framing effects under topic-clustered standard errors, compared to the primary paired-

t

-test analysis (Table 5 ). Clustered SE and

p

come from an item-level OLS regression of pro_score on treatment position, topic, and country, with standard errors clustered by topic (10 clusters). * indicates

p<0.05

Attribute	Command A	DeepSeek V4 Flash	Gemma 3 27B	Llama 4 Scout	Mistral Large	Qwen 3.6 27B
Ideology	2/2 of 2	2/2 of 2	2/2 of 2	2/2 of 2	2/2 of 2	2/2 of 2
Race	5/9 of 9	7/9 of 9	5/9 of 9	6/8 of 9	7/9 of 9	6/9 of 9
Gender	2/2 of 2	2/2 of 2	2/2 of 2	2/2 of 2	2/2 of 2	2/2 of 2
Age	4/5 of 5	2/2 of 5	5/5 of 5	1/4 of 5	2/4 of 5	5/5 of 5
Education	3/3 of 4	4/4 of 4	2/3 of 4	4/4 of 4	4/4 of 4	3/3 of 4

Appendix

Table 13: Number of category-level coefficients remaining significant at

p<0.05

under two-way-clustered standard errors (by survey question and biography profile), compared to the primary (naive) standard errors, out of the total number of non-reference category levels for that attribute. Format: clustered/naive of total.

Topic	MCQ	OT	Pooled
Gaza	+1.70	+1.06	+1.38
Gun control	+1.73	+1.01	+1.37
Ukraine	+1.07	+0.96	+1.01
Climate	+1.34	+0.64	+0.99
Healthcare	+1.13	+0.70	+0.92
LGBT+ rights	+1.19	+0.42	+0.80

Appendix

Table 14: Contested-terminology framing effects by political topic, pooled across the seven main-analysis models. Values report the mean pro-side minus anti-side stance difference for MCQ and open-text (OT) responses and across both formats.

Figure 12: Contested-terminology framing effects by model and political topic for MCQ (left) and open-text (right) responses. Larger positive values indicate greater directional movement towards the political position signalled by the terminology.

Format	Australia	United Kingdom	United States
MCQ	+1.06	+1.10	+1.20
OT	+0.65	+0.63	+0.68

Appendix

Table 15: Contested-terminology framing effects by country, pooled across the seven main-analysis models and restricted to the six topics evaluated in Australia, the United Kingdom, and the United States. Values report the mean pro-side minus anti-side stance difference by response format.

Figure 13: Mean political stance under pro-side and anti-side contested-terminology wording by country, shown separately for MCQ (left) and open-text (right) responses. Comparisons are restricted to the six political topics evaluated in all three countries.

Model	Format	Treatment	Mean range	Contradiction rate
Command A	MCQ	premise	0.017	0.0%
	MCQ	term	0.004	0.0%
	OT	premise	0.346	0.9%
	OT	term	0.385	0.4%
DeepSeek V4 Flash	MCQ	premise	0.743	5.7%
	MCQ	term	1.104	8.7%

Appendix

Table 16: Run-to-run variation across five independent generations of identical prompts, by model, response format, and treatment. Mean range is the average within-prompt range of stance scores across generations; contradiction rate is the proportion of repeated prompts whose responses cross the neutral midpoint.

Figure 14: Run-to-run consistency across repeated generations of identical prompts. Left and centre show the mean range of stance scores across repeated runs for MCQ and open-text responses, respectively; right shows the contradiction rate, defined as repeated responses crossing the neutral midpoint.

Judge pair	$N$	$r$	Exact	Within 1	Side agreement
Gemma 3 27B vs. Command A	12,454	0.77	58.5%	92.3%	76.4%
Gemma 3 27B vs. Llama 4 Scout	9,339	0.79	63.6%	92.3%	80.8%
Command A vs. Llama 4 Scout	8,915	0.85	75.0%	95.0%	84.0%

Appendix

Table 17: Pairwise agreement between the three LLM judges used to score open-text responses, restricted to responses from the seven main-analysis models. We report Pearson correlation (

r

), exact score agreement, agreement within one point on the five-point scale, and agreement on which side of the neutral midpoint the response falls.

Figure 15: Reliability of the LLM-as-a-judge scoring procedure for open-text responses. Left: pairwise Pearson correlation between judges; centre: agreement on political side by judge pair; right: side agreement according to the model whose responses are being evaluated.

Model	OT judge (all 3)	Survey	Bio
Command A	100%/99%	97%	92%
DeepSeek V4 Flash	100%/99%	99%	100%
Falcon 3 10B	96%/98%	98%	4%
Gemma 2 27B	98%/97%	97%	97%
Gemma 3 27B	98%/99%	100%	100%
Llama 3.1 8B	98%/99%	55%	70%

Appendix

Table 18: Percentage of expected observations available for each model at the pipeline stages with incomplete coverage. Paired values report contested-terminology/premise coverage; survey-baseline and biography columns have no treatment split. MCQ collection, open-text response collection, and open-text judging by at least one juror are 100% complete for every model and are omitted.

Model	Overall	Most-refused topic	Widest attribute	Highest value	Lowest value
Llama 3.1 8B	29.6%	Gaza (59%)	Race (17pp)	Indigenous 42%	Mixed race 25%
Llama 4 Scout	5.8%	Abortion (21%)	Ideology (8pp)	Conservative 10%	Liberal 2%
Qwen 3.6 27B	1.5%	Abortion (3%)	no attribute clears a 5pp spread

Appendix

Table 19: Summary of refusal patterns in the biography treatment for models with an overall refusal rate of at least 1%. The table reports overall refusal rates, the topic with the highest refusal rate, and the user attribute exhibiting the largest variation across attribute values.

Topic	Indigenous	Hispanic	South Asian	East Asian	White
Gaza	71.8%	60.9%	63.1%	61.2%	59.2%
Ukraine	60.2%	38.8%	44.9%	39.5%	31.9%
Indigenous rights	43.8%	56.6%	56.4%	52.1%	47.9%
Immigration	46.2%	33.1%	28.4%	29.4%	29.3%
Abortion	60.3%	49.2%	45.4%	46.1%	46.1%
LGBT+ rights	39.3%	28.1%	27.5%	24.1%	21.3%

Appendix

Table 20: Llama 3.1 8B biography-treatment refusal rates by political topic and persona race. The five race categories with the highest overall refusal rates are shown; Figure 16 reports the full ten-category comparison.

Panel of Figure 16 — Figure 16: Biography-treatment refusal rates by political topic and persona ideology (top) and race (bottom) for Llama 3.1 8B (left) and Llama 4 Scout (right). Darker cells indicate higher refusal rates, illustrating how refusals vary across both topics and user-profile characteristics.

Comparison	$N$	$r$	Exact	Within 1	Side agreement
Human consensus vs. Gemma 3 27B	200	0.79	62.0%	94.0%	75.5%
Human consensus vs. Command A	198	0.87	63.1%	93.9%	79.8%
Human consensus vs. Llama 4 Scout	198	0.80	64.6%	92.9%	80.8%
Human consensus vs. jury median	200	0.84	63.5%	94.5%	80.5%

Appendix

Table 21: Human consensus (median of 3 annotators) vs. each LLM judge and the jury median, 200-item validation subset.

Group	Mean $N$	Mean $r$	Mean exact	Mean within 1	Mean side agreement
Human annotators (3 pairs)	199	0.77	54.5%	90.6%	75.9%
AI judges (3 pairs)	197	0.79	64.7%	91.9%	78.8%

Appendix

Table 22: Human inter-rater agreement vs. LLM inter-judge agreement (mean across the 3 pairwise comparisons within each group), same 200-item subset.

Figure 17: Rating distribution: each of the three annotators and three LLM judges individually (faint lines), with the pooled human and pooled AI-judge distributions overlaid in bold, restricted to the 200-item validation subset.

Figure 18: Comparison of Gemma 2 27B and Gemma 3 27B. Left: directional framing effects across response formats and treatments. Right: contradiction rates across repeated generations of identical prompts, providing a measure of run-to-run self-consistency.

Framing Effects in Language Models

Framing effects in language models study how wording, context, perspective, and presentation alter model judgments and outputs.

See all 45 papers

Political Ideology Shifts in Large Language Models
Aug 22, 2025Pietro Bernardelle, Stefano Civelli, Leon Fröhling +3Political Bias in Language Models Language Model Bias Evaluation
The University of Queensland, Brisbane, Australia · GESIS, Cologne, Germany · University of Udine, Udine, Italy
How Identity and Opinion Shape Political Sycophancy in LLMs
Aug 29, 2026Li-Ni Fu, Chang-Chih Meng, Chien-Hua Chen +2LLM Sycophancy Political Bias in Language Models
Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan · Institute of Information Science, Academia Sinica, Taiwan · Research Center for Information Technology Innovation, Academia Sinica, Taiwan
LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution
May 26, 2026Adib Sakhawat, Syed Rifat Raiyan, Tahsin Islam +3Political Bias in Language Models Language Model Bias Evaluation
Systems and Software Lab (SSL) Department of Computer Science and Engineering Islamic University of Technology, Dhaka, Bangladesh

Aug 22, 2025J/K move · Enter open · S save

Political Ideology Shifts in Large Language Models

Pietro Bernardelle, Stefano Civelli, Leon Fröhling, Riccardo Lunardi+2

The University of Queensland, Brisbane, Australia · GESIS, Cologne, Germany · University of Udine, Udine, Italy

Political Bias in Language Models Language Model Bias Evaluation Persona-Conditioned Generation

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly deployed in politically sensitive contexts, raising concerns about their susceptibility to ideological biases. In this work, we examine how synthetic persona conditioning shapes ideological expression across seven open-weight instruction-tuned models (7B-72B parameters) using the Political Compass Test (62 statements) as a standardized behavioral probe. Across three studies involving 200,000 synthetic personas and more than 260 million model responses, we analyze implicit and explicit malleability, as well as theme-associated variations. We find that: (i) larger models exhibit broader implicit ideological coverage, increasing from 14-35% for 7-8B models to up to 49% for 70B+ models; (ii) explicit ideological priming induces large and statistically significant shifts, with right-authoritarian cues moving all models in the intended direction and producing larger effects in most model-axis comparisons; (iii) left-libertarian priming produces more heterogeneous responses, including counter-directional economic shifts in three of four 7-8B models, while all 70B+ models move in the intended direction; and (iv) theme-associated semantic content in persona descriptions is linked to systematic and interpretable directional shifts in ideological space. While our results identify an upstream mechanism through which persona conditioning can alter model responses under a standardized ideological probe, we do not test whether such shifts affect users beliefs, decisions, or political behavior. Our findings are best understood as evidence of ideological malleability at the generation layer, highlighting the need to account for interactional factors when evaluating political neutrality, fairness, and safety in English-prompted, persona-conditioned language models.

Abstract

Framing Effects in Language Models

Explore similar work

Political Ideology Shifts in Large Language Models

How Identity and Opinion Shape Political Sycophancy in LLMs

LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution