LLM Analysis of 150+ Years of German Parliamentary Debates on Migration Reveals a Shift from Post-War Solidarity to Anti-Solidarity in the Last Decade

Aida Kostikova; Ole Pütz; Steffen Eger; Olga Sabelfeld; Benjamin Paassen

cs.CLSep 8, 2025

LLM Analysis of 150+ Years of German Parliamentary Debates on Migration Reveals a Shift from Post-War Solidarity to Anti-Solidarity in the Last Decade

Authors: Aida Kostikova, Ole Pütz, Steffen Eger, Olga Sabelfeld, Benjamin Paassen

Organizations: Bielefeld University · ∗Corresponding author · University of Technology Nuremberg

Abstract

Migration has been a core topic in German political debate, from postwar expellee displacement to labor migration and recent refugee movements. Large-scale analysis of such political discourse has traditionally required extensive manual annotation, limiting coverage. Large language models (LLMs) offer a scalable alternative. Using a theory-driven annotation scheme, we examine how well LLMs annotate subtypes of solidarity and anti-solidarity in German parliamentary debates and whether the resulting labels support valid downstream inference. We first evaluate multiple LLMs across model size, prompting strategies, fine-tuning, historical versus contemporary data, and systematic errors. The strongest models, especially GPT-5 and gpt-oss-120B, achieve macro-F1 scores comparable to human agreement, although their systematic errors can bias downstream results. We therefore combine soft-label model outputs with Design-based Supervised Learning (DSL) to reduce bias in long-term trend estimates. Beyond the methodological evaluation, we interpret the resulting annotations from a social-scientific perspective across the Reichstag (1867-1933), West German Bundestag (1949-1990) and Re-Unified German Bundestag (1990-2025) corpus to trace trends in solidarity and anti-solidarity toward migrants, with detailed analysis of postwar Germany (1949-1957) and contemporary Germany (2009-2025). We find relatively high levels of solidarity in the postwar period, especially in group-based and compassionate forms, and a marked rise in anti-solidarity since 2015. We argue that LLMs can support large-scale social-scientific text analysis, but their outputs require rigorous validation and, where systematic errors are present, statistical correction.

Figures & tables

Figure 1: Overview of the annotation scheme adapted from Thijssen (2012) . Statements are first classified at a high level as expressing solidarity , anti-solidarity , mixed , and none . Solidarity and anti-solidarity instances are then further categorized into subtypes based on the underlying rationale: group-based , exchange-based , compassionate , and empathic . The example sentence in the figure illustrates exchange-based anti-solidarity.

Figure 2: Absolute number of instances (left) and their share relative to all DeuParl sentences (right) for the Migrant dataset.

Frame	Solidarity example	Anti-solidarity example
Group-based	“[…] No state can accept that a significant part of the population […] remains outside the political community […]. No immigrant who […] shares the ways of life and values of the native population, can be expected to remain excluded as a foreigner under the current laws. […]”	“[…] A public administration [must not] employ foreigners to drive down wages. […] To act in the national interest, [mine owners] have a duty to dismiss the foreign workers whenever suitable German workers remain unemployed. The administration has many means, including expulsion, to advance the national idea.”
Exchange-based	“[…] We must remain open to the world because, economically, our country cannot afford to shut itself off along national lines. Our economy is export-oriented, and we need qualified immigration to Germany. We want to organize this, and we will. […]”	“ […] Migration can fail, especially when immigrants have low qualifications. In 2013 […] 40 percent of non-EU immigrants had no qualifications. Since the refugee wave, stabbings have increased by 20 percent […]. Is that successful migration?”
Compassion.	“[…] We can support all measures that serve to help people in an extreme situation. The immigrants whom we in the Federal Republic refer to as ethnic German resettlers and migrants from East Germany are indeed in such a situation. […]”	“[…] there is no end in sight to the abuse of asylum. There has been a dramatic increase in immigration by people who are not politically persecuted. […] The abuse of asylum has developed a dynamic of its own and has long since become a vicious circle.”
Empathic	“[…] we are concerned that the immigration law […] contradicts the goals of our foreign cultural policy; for it does not promote encounters between cultures on an equal footing, but demands an almost complete willingness to assimilate as the price of immigration. We are in favor of integration, not assimilation.”	“[…] an SPD politician warned: ‘ We have overextended ourselves by taking in people from completely different cultural worlds; 7 million foreigners in Germany represent a misguided development. We must prevent immigration from foreign cultures; [it] only creates an additional problem.’ Reasonable views. […]”

Table 1: Illustrative examples of the four (anti-)solidarity frames from the human annotated subset. Bold text is the main sentence, the other sentences are for context. Original German texts are provided in Table 20 in the Appendix.

Figure 3: Distribution of human annotated instances by time for the Migrant dataset. Total indicates the number of instances per decade (including None ); see Table 21 in the Appendix for exact counts.

	Migrant (T1)		Migrant (T2)
Model	Zero	Few	Zero	Few
Closed-weight models
GPT-4	0.42 (0.73)	0.43 (0.63)	0.29 (0.62)	0.26 (0.55)
GPT-5	0.43 (0.67)	0.50 (0.71)	0.33 (0.60)	0.34 (0.58)
Large open-weight models
gpt-oss-20B	0.37 (0.56)	0.40 (0.56)	0.26 (0.49)	0.26 (0.49)

Table 2: Comparative performance (macro F1). Scores are reported as fine-grained (high-level). “Migrant (T1)” is the original benchmark; “Migrant (T2)” is newly introduced in this paper. Green cells mark the best score per test set (model’s best setting); yellow cells mark the best score among open models (second best overall). Gray text indicates suboptimal settings for a given model. Pairwise statistical comparisons of selected few-shot models are reported in Appendix D in the Appendix.

Figure 4: Model size vs. performance on high-level and fine-grained classification in Migrant (T1) . Each model is plotted with its best-performing configuration (zero- or few-shot). The x-axis is logarithmic, so spatial distances do not represent linear differences in parameter count.

Figure 5: F1 scores for high-level labels on the combined Migrant test sets (Test 1 and Test 2), based on the best-performing configuration for each model.

Test set	gpt-oss-120B	Llama-3.3-70B	Qwen-2.5	Ensemble
Migrant (T1)	0.48 (0.70)	0.38 (0.65)	0.41 (0.61)	0.48 (0.72)
Migrant (T2)	0.31 (0.57)	0.29 (0.57)	0.28 (0.51)	0.31 (0.59)

Table 3: Performance (macro F1) of models included in the ensemble and the resulting majority-weighted ensemble on fine-grained and high-level tasks (in parentheses). All individual models use their few-shot configuration. In tie cases, votes are weighted by validation-set macro-F1. Tie-breaking weights are based on validation-set performance, ranking gpt-oss-120B

>

Llama-3.3-70B

>

Qwen-2.5.

Panel of Figure 6 — Figure 6: Row-normalized confusion matrices (%) for the best-performing configuration of each of the top models by macro-F1, as reported in Table 2 , on the two Migrant test sets combined (Test 1 and Test 2). Each row sums to 100, showing how items with a given reference label are distributed across predicted labels. Human Annotators (LOO) shows aggregated leave-one-out annotator comparisons, where each annotator is compared to the consensus of the remaining annotators. The Ensemble model aggregates predictions from Llama-3.3-70B, Qwen-2.5-72B, and gpt-oss-120B.

Panel of Figure 7 — Figure 7: Row-normalized high-level and fine-grained confusion matrices (%) for the best-performing configurations of the top models on the combined Migrant test sets, restricted to 1867-1900 . As in Figure 6 , rows sum to 100, Human Annotators (LOO) denotes aggregated leave-one-out annotator comparisons, and the Ensemble aggregates Llama-3.3-70B, Qwen-2.5-72B, and gpt-oss-120B.

Model / Estimator	Pearson	Spearman	RMSE
High-level trends
Llama-3.3-70B	0.899	0.881	0.082
Llama-4-Scout	0.777	0.802	0.082
Qwen-2.5	0.757	0.744	0.118
gpt-oss-120B	0.881	0.822	0.068
GPT-5	0.855	0.829	0.070

Table 4: Comparison with human-based decade-level trends across estimators. For DSL estimators, values are reported as mean

\pm

standard deviation (SD) across 300 random seeds. Reference: human soft labels. Results against human majority vote are reported in Table 22 in the Appendix.

Panel of Figure 8 — Figure 8: High-level and fine-grained (anti-)solidarity trends for the Migrant category under selected labeling and correction strategies discussed in Section 3.2.2 , based on the 2,184-instance human-annotated subset . Trends are computed as label shares within each decade (see definitions in Section 3.2.2 ). The period from 1930 to 1949 in each panel (grey shaded area) is excluded due to limited data availability during and immediately after the NS dictatorship. For DSL, lines show the mean estimate across 300 simulated runs. In each run, approximately 20% of the human-annotated instances are retained as labeled data, while the remaining human labels are treated as unobserved. Colored shaded areas show the mean $\pm$ the decade-specific SD across runs. Figures 24 and 25 in the Appendix show the corresponding trends for the remaining configurations.

Panel of Figure 9 — Figure 9: Solidarity trends for Migrant speeches computed using DSL-corrected soft labels (see definition (c) in Section 3.2.2 ) on the full 63K dataset . Figure 9(a) shows the high-level (anti-)solidarity trends, while Figure 9(b) shows distribution of fine-grained (anti-)solidarity subtypes over time. The period 1930–1949 (grey area) is excluded due to limited data during and after the Nazi dictatorship.

Figure 10: Estimated counts of (anti-)solidarity mentions in Bundestag debates, 1949–1957 (DSL-adjusted, see definition (d) in Section 4.2.1 ). LAG $=$ Lastenausgleichsgesetz ( en. : )Equalisation of Burdens Law; BVFG $=$ Bundesvertriebenengesetz ( en. : )Federal Expellees Act; HAuslG $=$ Gesetz über die Rechtsstellung heimatloser Ausländer im Bundesgebiet ( en. : )Law on the Legal Status of Stateless Foreigners in the Federal Territory.

Figure 11: Distribution of (anti-)solidarity subtypes, 1949–1957 (DSL-adjusted expected counts, see definition (d) in Section 4.2.1 ). Inset shows the evolving composition of solidarity subtypes, normalized within each six-month period as shares of all solidarity framing (see definition (c) in Section 3.2.2 ).

Figure 12: Relative distribution of (anti-)solidarity subtypes across selected migrant-related keywords, 1949–1957 (DSL-adjusted estimates; see definition (e) in Section 4.2.2 ). Bars show each subtype’s share within a keyword group; shares sum to 100%. Full label distributions appear in Table 28 in the Appendix.

Figure 13: Distribution of (anti-)solidarity subtypes across Bundestag parties, 1949–1957 (DSL-adjusted estimates; see definition (e) in Section 4.2.2 ). Bars show each subtype’s share within a party, summing to 100%. The analysis includes parties with at least 150 DSL-adjusted stanced statements. Full label distributions appear in Table 29 in the Appendix.

Panel of Figure 14 — Figure 14: (Anti-)solidarity trends for the Migrant category, 2009–2025. Figure 14(a) : Yearly DSL-adjusted shares of solidarity, anti-solidarity, and mixed stance (see definition (c) in Section 3.2.2 ). The none category is included in the normalization but omitted from the visualization. Figure 14(b) : Yearly DSL-adjusted shares of solidarity (left) and anti-solidarity (right) subtypes (see definition (c) in Section 3.2.2 ). Within each year, subtype shares are renormalized across the eight solidarity and anti-solidarity categories (excluding mixed and none), summing to 100%.

Figure 15: Raw yearly counts of parliamentary statements relating to migrants by selected political parties, 2009-2025.

Panel of Figure 16 — Figure 16: DSL-adjusted yearly shares of (anti-)solidarity subtypes across Bundestag parties, 2009-2025 (see definition (c) in Section 3.2.2 ). Within each year and party, subtype shares are renormalized across the eight solidarity and anti-solidarity categories so that they sum to 100%. Note: the AfD entered the Bundestag only in 2017, while the FDP was absent from 2013–2017; accordingly, no observations appear for these periods. See larger versions of the party-specific panels in Figure 26 in the Appendix.

Appendix figures & tables55 assets

Supplementary material from the paper’s appendix.

Appendix

Comparison	F1 A	F1 B	$\Delta$ F1	95% CI	$p_{\mathrm{raw}}$	$p_{\mathrm{Holm}}$
Migrant (T1) — Fine-grained
GPT-5 – gpt-oss-120B	0.500	0.485	0.015	[-0.029, 0.059]	0.5215	0.5215
gpt-oss-120B – Llama-3.3-70B	0.485	0.376	0.109	[0.066, 0.154]	0.0002	0.0010
gpt-oss-120B – Qwen-2.5	0.485	0.407	0.078	[0.036, 0.121]	0.0004	0.0012
gpt-oss-120B – Gemma-2-27B	0.485	0.325	0.161	[0.118, 0.205]	0.0002	0.0010
gpt-oss-120B – gpt-oss-20B	0.485	0.400	0.085	[0.041, 0.130]	0.0008	0.0016

Appendix

Table 5: Paired stratified bootstrap comparisons of the planned few-shot model pairs. Positive differences favor the first model. Unadjusted 95% percentile bootstrap confidence intervals and two-sided bootstrap

p

-values are based on 10,000 resamples. Holm correction is applied to the five

p

-values within each test-set and annotation-level combination. Bold values indicate

p_{\mathrm{Holm}}<.05

Comparison	F1 A	F1 B	$\Delta$ F1	95% CI	$p_{\mathrm{raw}}$	$p_{\mathrm{Holm}}$
Migrant (T1) — Fine-grained
Llama-3.3-70B – Gemma-2-27B	0.376	0.325	0.051	[0.009, 0.095]	0.0164	0.0164
Qwen-2.5 – Gemma-2-27B	0.407	0.325	0.083	[0.043, 0.125]	0.0002	0.0004
Migrant (T1) — High-level
Llama-3.3-70B – Gemma-2-27B	0.647	0.538	0.109	[0.062, 0.159]	0.0002	0.0004
Qwen-2.5 – Gemma-2-27B	0.611	0.538	0.073	[0.028, 0.119]	0.0026	0.0026

Appendix

Table 6: Additional paired stratified bootstrap comparisons of Llama-3.3-70B and Qwen-2.5 against Gemma-2-27B. Positive differences favor the first model. Unadjusted 95% percentile bootstrap confidence intervals and two-sided bootstrap

p

-values are based on 10,000 resamples. Holm correction is applied to the two

p

-values within each test-set and annotation-level combination. Bold values indicate

p_{\mathrm{Holm}}<.05

Model 1	Model 2	Error Jaccard	P(Same Wrong $\mid$ Both Wrong)
gpt-oss-120B	Qwen-2.5	0.638	0.493
Llama-3.3-70B	Qwen-2.5	0.609	0.453
Llama-3.3-70B	gpt-oss-120B	0.573	0.461

Appendix

Table 7: Pairwise overlap in model prediction errors. Error Jaccard measures the proportion of shared errors relative to the union of all errors between two models. P(Same Wrong

\mid

Both Wrong) measures how often two models predict the same wrong label when both are incorrect.

Comparison	Total Variation	Jensen–Shannon Divergence
Model ensemble vs humans	0.49 (0.32)	0.39 (0.22)
Human vs human (avg 10 splits)	0.45 (0.19)	0.35 (0.12)
Randomized model baseline	0.75 (0.59)	0.69 (0.47)
gpt-oss-120B vs humans	0.58 (0.36)	0.48 (0.29)
Qwen-2.5 vs humans	0.59 (0.42)	0.50 (0.34)
Llama-3.3-70B vs humans	0.57 (0.40)	0.48 (0.32)

Appendix

Table 8: Distribution distances between human and model label distributions (lower = closer alignment). Values are shown as fine-grained (high-level).

Model	A40 GPU-h	kWh	CO 2 (kg)
Llama-3.3-70B	1,035.9	310.8	106.9
Qwen-2.5	1,211.8	363.5	125.1
Total	2,247.7	674.3	232.0

Appendix

Table 9: Estimated GPU-board energy use and CO 2 emissions, assuming 300 W per A40 and 344 g CO 2 /kWh.

Audit sample construction	1990-onward	2015-onward
Human-annotated instances	1123	441
All model-AS predictions	912	479
Candidate instances: human $\neq$ AS, at least one model $=$ AS	203	62
Individual model-AS predictions in candidate instances	319	111

Appendix

Table 10: Construction of the anti-solidarity false-positive audit sample. The broader audit covers cases from 1990 onward; the 2015-onward subset is used in a robustness check for the recent period in Section 4.3.1 .

Audit outcome	1990 onward ( $N_{\mathrm{AS}}=912$ )	2015 onward ( $N_{\mathrm{AS}}=479$ )
(1) Misattributed anti-solidarity errors	139 (15.2%)	38 (7.9%)
(2) Other clear AS false positives	38 (4.2%)	16 (3.3%)
(1+2) All clear AS false positives	177 (19.4%)	54 (11.3%)
(3) Plausible human-model disagreements	142 (15.6%)	57 (11.9%)

Appendix

Table 11: Manual coding results for candidate model-AS predictions. Entries report the number of predictions assigned to each audit outcome, with the percentage of all model anti-solidarity predictions in the corresponding period in parentheses.

Original text	Translation	Example model response
Audit category (1): Misattributed anti-solidarity
“ […] Die AfD, also die mit dem ganz kleinen braunen Karo, selbst wenn sie blaue Jeansjacketts tragen, hat mal wieder einen Skandal aufgedeckt, der aber gar keiner ist: Kinder von Ausländern erhalten gutes deutsches Kindergeld, sogar in Polen oder Rumänien. ‘Florida-Rolf’ ist jetzt zehn Jahre alt, heißt ‘Andrzej’ oder ‘Bogdana’ und lässt es sich in Warschau oder Bukarest auf deutsche Kosten gut gehen. Welche Ungerechtigkeit! […] ”	“ […] The AfD, the party with the little brown check pattern even when its members wear blue denim jackets, has once again uncovered a scandal that is not actually a scandal at all: children of foreigners receive good German child benefits, even in Poland or Romania. ‘Florida Rolf’ is now ten years old, is called ‘Andrzej’ or ‘Bogdana,’ and is supposedly living it up in Warsaw or Bucharest at German taxpayers’ expense. What an injustice! […] ”	gpt-oss-120B : anti-solidarity ; human reference : solidarity Label: ANTI - SOLIDARITY; Explanation: The speaker denounces the fact that children of foreigners receive German child benefits, calling it an injustice and a scandal. This frames migrants as undeserving recipients of public resources and expresses opposition to their inclusion, without any supportive remarks.
Audit category (2): Other clear anti-solidarity false positive
“ […] Das türkische Regime verletzt [gemeinsame] Werte auf offener Weltbühne, und das seit Jahren. Die türkische Regierung erpresst Europa. Sie droht, mal offen und mal weniger offen, europäische NATO-Partner mit unkontrolliertem Zustrom von Migration und Flüchtlingen zu destabilisieren. […] Sie will auch unser Schweigen angesichts einer immer schlimmer werdenden Verfolgung unschuldiger Bürgerinnen und Bürger der Türkei im Inneren. […] ”	“ […] The Turkish regime has been violating [shared] values openly on the world stage for years. The Turkish government is blackmailing Europe. It threatens, sometimes openly and sometimes less openly, to destabilize its European NATO partners through an uncontrolled influx of migrants and refugees. […] It also wants to secure our silence in the face of the increasingly severe persecution of innocent citizens within Turkey. […] ”	Qwen-2.5 : anti-solidarity , human reference : none Label: ANTI-SOLIDARITY; Explanation: The text criticizes Turkey’s government for using the threat of uncontrolled migration and refugees to extort Europe, framing migrants as a tool of political leverage and a potential source of destabilization. This portrayal of migrants as a threat aligns with an anti-solidarity stance.
Audit category (3): Plausible disagreement with the human reference
“ […] Der Schleppermafia ist das Handwerk gelegt worden, die Grenzen werden kontrolliert. Australien hilft jetzt jährlich 20 000 echten Flüchtlingen und nicht 10 000 anderen, die den Schleppern die allerhöchsten Preise zahlen können. […] jeder vernünftige Neuanfang in der Asylpolitik muss sich an solchen modernen Erfolgsmodellen orientieren. Wenn nicht, wird der linksgrüne Traum von einer Welt ohne Grenzen zu einem grenzenlosen Albtraum für unser Land und für Europa. Und das müssen und werden wir verhindern, meine Damen und Herren. ”	“ […] The smuggling networks have been put out of business, and the borders are being controlled. Australia now helps 20,000 genuine refugees each year instead of 10,000 others who can afford to pay smugglers the highest prices. […] Any sensible new beginning in asylum policy must follow such modern models of success. Otherwise, the left-green dream of a world without borders will turn into a boundless nightmare for our country and for Europe. And that is something we must and will prevent. ”	Llama-3.3-70B : anti-solidarity , human reference : mixed Label: ANTI-SOLIDARITY Explanation: The text expresses opposition towards a liberal migration policy, emphasizing the need for control and the preference for helping “echten Flüchtlingen” (genuine refugees) over those who can pay the highest prices, implying a critique of asylum seekers who are not considered genuine. The text also warns of a “grenzenlosen Albtraum” (boundless nightmare) for the country and Europe if a strict migration policy is not implemented, indicating a stance against open borders and unregulated migration. […]

Appendix

Table 12: Examples from the manual audit of anti-solidarity predictions. For each audit category, we show the original German passage, its English translation, and one representative model response.

Audit outcome	N	Raw AS	$\hat{g}_{\mathrm{AS}}$ mean $\pm$ SD	95% seed interval	Seeds with $\hat{g}_{\mathrm{AS}}<$ Raw AS
(1) Misattribution	38	0.579	0.374 $\pm$ 0.068	[0.254, 0.521]	100.0%
(2) Other AS FPs	16	0.771	0.587 $\pm$ 0.066	[0.448, 0.718]	100.0%
(3) Plausible disagreements	57	0.789	0.641 $\pm$ 0.057	[0.521, 0.753]	99.7%

Appendix

Table 13: First-stage DSL calibration for the manually audited anti-solidarity predictions across 300 seeds. Raw AS is the mean anti-solidarity in the raw ensemble. Calibrated

\hat{g}_{\mathrm{AS}}

is reported as mean

\pm

SD across seeds with the 2.5th-97.5th percentile interval. The final column gives the percentage of seeds for which calibrated anti-solidarity is lower than raw anti-solidarity.

N

is the number of audited model anti-solidarity predictions in each category.

Year	DSL S	DSL AS	AS after 7.9% adjustment	AS after 11.3% adjustment
2015	0.674	0.156	0.144	0.139
2016	0.624	0.125	0.115	0.111
2017	0.480	0.195	0.180	0.173
2018	0.424	0.441	0.406	0.391
2019	0.385	0.566	0.521	0.503
2020	0.310	0.607	0.559	0.539

Appendix

Table 14: Yearly solidarity and anti-solidarity estimates from 2015 onward under the original DSL estimates and the two audit-based sensitivity adjustments. Bold anti-solidarity values indicate years in which the corresponding anti-solidarity estimate exceeds the solidarity estimate.

Figure 17: Sensitivity of the post-2015 anti-solidarity trend to false positives identified in the manual audit. Original DSL-adjusted estimates are compared with two scenarios that reduce anti-solidarity proportionally by the estimated rate of misattribution errors (7.9%) and all clear anti-solidarity false positives (11.3%).

Party	$n$	GPT-5		gpt-oss-120B		Llama-3.3-70B		Qwen-2.5
		Macro	Weight.	Macro	Weight.	Macro	Weight.	Macro	Weight.
CDU/CSU	322	0.575	0.672	0.572	0.678	0.553	0.631	0.567	0.662
SPD	199	0.533	0.753	0.454	0.687	0.440	0.682	0.437	0.665
Grüne	142	0.604	0.851	0.462	0.813	0.453	0.757	0.370	0.665
FDP	82	0.565	0.719	0.576	0.696	0.617	0.738	0.473	0.615
AfD	121	0.419	0.806	0.423	0.757	0.535	0.791	0.381	0.770

Appendix

Table 15: Party-conditioned high-level classification performance on the human-annotated Migrant subset from 1990 onward. Macro F1 assigns equal weight to each evaluable high-level label, whereas weighted F1 weights each label-specific F1 score by its support in the human reference within the corresponding party. The best score among the open models for each party and metric is shown in bold.

Party	GPT-5	gpt-oss-120B	Llama-3.3-70B	Qwen-2.5
Solidarity ( $\Delta S$ , pp)
CDU/CSU	-7.1	+1.6	-1.9	-7.8
SPD	+0.0	-6.5	-8.0	-11.6
Grüne	+3.5	-1.4	-8.5	-26.8
FDP	+4.9	-1.2	-4.9	-11.0
AfD	-4.1	-0.8	+1.7	-6.6

Appendix

Table 16: Signed differences between model-predicted and human-reference label shares (

p_{\mathrm{model}}

and

p_{\mathrm{human}}

) by party. We define

\Delta S=p_{\mathrm{model}}(S)-p_{\mathrm{human}}(S)

and

\Delta AS=p_{\mathrm{model}}(AS)-p_{\mathrm{human}}(AS)

. Positive values indicate overprediction and negative values indicate underprediction. Differences are reported in percentage points (pp), i.e., the arithmetic difference between two percentages.

Figure 18: High-level confusion matrices per party on the human-annotated Migrant subset from 1990 onward. Cell annotations report the number of instances and the row-normalized percentage for the corresponding human-reference label. The support of each human-reference label is shown as $n$ next to the row label.

Party	gpt-oss-120B	Llama-3.3-70B	Qwen-2.5
CDU/CSU	0/29 (0.0%)	0/19 (0.0%)	2/29 (6.9%)
SPD	2/18 (11.1%)	5/12 (41.7%)	8/19 (42.1%)
Grüne	2/4 (50.0%)	14/15 (93.3%)	23/27 (85.2%)
FDP	2/7 (28.6%)	4/6 (66.7%)	7/16 (43.8%)
AfD	1/9 (11.1%)	1/8 (12.5%)	1/12 (8.3%)
Linke	7/7 (100.0%)	15/16 (93.8%)	24/27 (88.9%)

Appendix

Table 17: Misattribution errors among manually audited post-1990 cases in which the human reference is not anti-solidarity but the model predicts anti-solidarity . Entries report the number of misattribution errors divided by the number of audited candidate predictions, with percentages in parentheses.

Party	$n$	Raw committee		Soft DSL		Soft DSL, no Qwen		Hard DSL, gpt-oss
		S	AS	S	AS	S	AS	S	AS
CDU/CSU	6365	45.0%	20.8%	48.3%	27.1%	47.7%	27.5%	45.6%	29.1%
SPD	3956	69.1%	6.5%	75.8%	2.4%	75.5%	2.4%	76.0%	3.7%
Grüne	2335	68.7%	9.3%	71.7%	1.4%	71.0%	1.3%	72.5%	1.1%
FDP	1473	53.5%	13.9%	58.9%	6.8%	58.8%	7.6%	52.6%	6.7%
AfD	2763	5.4%	77.6%	4.5%	83.2%	4.5%	83.2%	4.7%	82.5%

Appendix

Table 18: 2009-2025 party-level solidarity and anti-solidarity estimates under the raw model committee and alternative DSL versions. The raw committee averages the label votes of gpt-oss-120B, Llama-3.3-70B, and Qwen-2.5 into a soft label distribution without DSL correction. Soft DSL is the main soft-label DSL using the three-model committee and soft human labels. Soft DSL, no Qwen removes Qwen-2.5 from the model committee. Hard DSL, gpt-oss uses one-hot gpt-oss-120B predictions as the surrogate and human-majority one-hot labels for DSL correction. Cells report party-level shares of solidarity (S) and anti-solidarity (AS).

Dataset	Fine-grained macro-F1	High-level macro-F1
Migrant (T1)	0.477 $\pm$ 0.023 [0.445–0.506]	0.692 $\pm$ 0.014 [0.680–0.716]
Migrant (T2)	0.278 $\pm$ 0.025 [0.254–0.319]	0.519 $\pm$ 0.013 [0.504–0.532]

Appendix

Table 19: Decoding stability of gpt-oss-120B in the few-shot setting across five decoding seeds. Macro-F1 is reported as mean

\pm

standard deviation [minimum–maximum] across seeds.

Frame	Solidarity example	Anti-solidarity example
Group-based	“ Kein Staat kann hinnehmen, daß ein bedeutsamer Teil der Bevölkerung […] außerhalb der staatlichen Gemeinschaft […] steht. Keinem Zuwanderer, der […] Lebensweisen und Wertvorstellungen der Einheimischen teilt, ist zuzumuten, nach den jetzigen Gesetzen als Ausländer ausgegrenzt zu werden. […] ”	“ […] Niemals aber darf der Wunsch, die üblichen Löhne zu drücken, einer öffentlichen Verwaltung Anlaß geben, Ausländer einzustellen. […] Wollen die Grubenherren in Wahrheit national handeln, so haben sie die Pflicht, die ausländischen Arbeiter zu entlassen, wenn geeignete deutsche Arbeiter unbeschäftigt bleiben. Die Verwaltung hat ja viele Waffen, zuletzt die Ausweisungsbefugnis, um der nationalen Idee Geltung zu verschaffen. ”
Exchange-based	“ […] Aber wir müssen auch deswegen weltoffen sein, weil es sich unser Land ökonomisch nicht leisten kann, national vernagelt zu sein. Unsere Wirtschaft ist exportorientiert, und wir brauchen eine qualifizierte Zuwanderung nach Deutschland. Dies wollen und werden wir organisieren. […] ”	“ […] Migration kann scheitern, vor allem dann, wenn die Qualifikation der Einwanderer niedrig ist. 2013 […] hatten 40 Prozent der Zuwanderer aus dem Nicht-EU-Ausland keinen Abschluss. Seit der Flüchtlingswelle haben die Messerstechereien um 20 Prozent zugenommen, […] Ist das eine hervorragend erfolgreiche Migration? ”
Compassion.	“ […] Alle Maßnahmen, die dazu dienen, Menschen in einer Extremsituation zu helfen, können wir nur unterstützen. Die von uns in der Bundesrepublik als Aussiedler und Übersiedler bezeichneten Einwanderer befinden sich tatsächlich in einer solchen Situation. […] ”	“ […] Dennoch ist kein Ende des Asylmißbrauchs abzusehen. Es gibt einen dramatischen Anstieg der Zuwanderung politisch Nichtverfolgter. […] Der Asylmißbrauch hat eine Eigendynamik erreicht, die längst zum Teufelskreis geworden ist. ”
Empathic	“ […] wir befürchten, dass das Einwanderungsgesetz […] den Zielen unserer auswärtigen Kulturpolitik direkt widerspricht; denn es fördert nicht die Begegnung gleichberechtigter Kulturen, sondern fordert nahezu Assimilationsbereitschaft als Preis für Einwanderung. Wir sind für Integration, nicht für Assimilation. ”	“ […] ein SPD-Politiker warnte: ‘Wir haben uns mit dem Zustand von Menschen aus völlig anderen kulturellen Welten übernommen, 7 Millionen Ausländer in Deutschland sind eine fehlerhafte Entwicklung.’ Wir müssen eine Zuwanderung aus fremden Kulturen unterbinden, [es] schafft nur ein zusätzliches Problem. – Vernünftige Einstellungen. […] ”

Appendix

Table 20: Illustrative examples of the four (anti-)solidarity frames from the human annotated subset. Bold text is the main sentence, the other sentences are for context.

Label	Migrant
Group-based Solidarity	270 (12.4%)
Exchange-based Solidarity	79 (3.6%)
Empathic Solidarity	32 (1.5%)
Compassionate Solidarity	692 (31.7%)
Solidarity (no subtype)	55 (2.5%)
Total Solidarity	1128 (51.6%)

Appendix

Table 21: Distribution of labels in the human-annotated Migrant subset. Percentages are relative to the total number of migrant-related instances.

Figure 19: Macro F1 scores over time for each model (best configuration), and average pairwise Cohen’s Kappa for human annotators, evaluated on the combined Migrant test sets (Test 1 and Test 2) and grouped by decade.

Figure 20: High-level row-normalized confusion matrices (%) for the best-performing configuration of each model on the two migrant test sets (Test 1 and Test 2, combined). Each row sums to 100, showing how items with a given reference label are distributed across predicted labels. Human Annotators (LOO) shows aggregated leave-one-out annotator comparisons, where each annotator is compared to the consensus of the remaining annotators. The Ensemble model aggregates predictions from Llama-3.3-70B, Qwen-2.5-72B, and gpt-oss-120B.

Figure 21: Fine-grained row-normalized confusion matrices (%) showing the best-performing configuration for each of the selected models on the two migrant test sets (Test 1 and Test 2), combined. Each row sums to 100, showing how items with a given reference label are distributed across predicted labels. Human Annotators (LOO) shows aggregated leave-one-out annotator comparisons, where each annotator is compared to the consensus of the remaining annotators. The Ensemble model aggregates predictions from Llama-3.3-70B, Qwen-2.5-72B, and gpt-oss-120B.

Figure 22: Distribution of all Migrant keywords over the years, normalized per keyword. The keywords are sorted by frequency, which means that the reliability decreases towards the bottom-right.

Figure 23: (Anti-)solidarity trends for the Migrant category by keyword and decade, 1867-2025. Each panel shows DSL-adjusted decade-level shares of solidarity and anti-solidarity for one keyword (see definition (c) in Section 3.2.2 ). The mixed and none categories are included in the four-class distribution but omitted from the visualization. Keywords are sorted by frequency, so estimates become less reliable toward the bottom right.

Panel of Figure 24 — Figure 24: High-level solidarity trends for the Migrant category under selected labeling and correction strategies discussed in Section 3.2.2 based on the 2,184-instance human-annotated subset . Trends are computed as label shares within each decade (see definitions in Section 3.2.2 ). The period from 1930 to 1949 in each panel (grey shaded area) is excluded due to limited data availability during and immediately after the NS dictatorship. For DSL, lines show the mean estimate across 300 simulated runs. In each run, approximately 20% of the human-annotated instances are retained as labeled data, while the remaining human labels are treated as unobserved. Colored shaded areas show the mean $\pm$ the decade-specific SD across runs.

Panel of Figure 25 — Figure 25: Fine-grained solidarity trends for the Migrant category under selected labeling and correction strategies discussed in Section 3.2.2 based on the 2,184-instance human-annotated subset . Trends are computed as label shares within each decade (see definitions in Section 3.2.2 ). The period from 1930 to 1949 in each panel (grey shaded area) is excluded due to limited data availability during and immediately after the NS dictatorship. For DSL, lines show the mean estimate across 300 simulated runs. In each run, approximately 20% of the human-annotated instances are retained as labeled data, while the remaining human labels are treated as unobserved. Colored shaded areas show the mean $\pm$ the decade-specific SD across runs.

Model / Estimator	Pearson	Spearman	RMSE
High-level trends
Llama-3.3-70B	0.868	0.864	0.082
Llama-4-Scout	0.738	0.757	0.084
Qwen-2.5	0.718	0.714	0.116
gpt-oss-120B	0.847	0.810	0.068
GPT-5	0.818	0.816	0.073

Appendix

Table 22: Comparison with human-based decade-level trends across estimators. For DSL estimators, values are reported as mean

\pm

standard deviation (SD) across 300 random seeds. Reference: human majority vote. Results against human soft-label trends are reported in Section 3.2.2 .

Label	Model	Pearson	Spearman	RMSE
Solidarity	Llama-3.3-70B	0.921	0.929	0.099
	Model soft raw	0.942	0.939	0.106
	DSL soft	0.955 $\pm$ 0.019	0.940 $\pm$ 0.029	0.051 $\pm$ 0.010
Anti-solidarity	Llama-3.3-70B	0.872	0.836	0.069
	Model soft raw	0.922	0.896	0.053
	DSL soft	0.935 $\pm$ 0.027	0.869 $\pm$ 0.062	0.043 $\pm$ 0.009

Appendix

Table 23: High-level per-label agreement between model-based and human-annotated decade-level trends. For DSL, values are reported as mean

\pm

standard deviation (SD) across 300 random seeds. Higher is better for Pearson and Spearman; lower is better for RMSE. Reference: human soft labels.

Label	Model	Pearson	Spearman	RMSE
Group-based S	Llama-3.3-70B	0.542	0.429	0.180
	Model soft raw	0.780	0.776	0.117
	DSL soft	0.771 $\pm$ 0.095	0.758 $\pm$ 0.107	0.055 $\pm$ 0.011
Exchange-based S	Llama-3.3-70B	0.319	0.352	0.074
	Model soft raw	0.842	0.721	0.079
	DSL soft	0.703 $\pm$ 0.121	0.696 $\pm$ 0.130	0.033 $\pm$ 0.007

Appendix

Table 24: Fine-grained per-label agreement between model-based and human-annotated decade-level trends. For DSL, values are reported as mean

\pm

standard deviation (SD) across 300 random seeds. Higher is better for Pearson and Spearman; lower is better for RMSE. Reference: human soft labels.

Label	Full prevalence	HumanSoft trend SD	Seed variation / human signal	Pearson SD	Sampling-miss rate
Solidarity	52.5%	0.156	0.314	0.019	0.0%
None	17.8%	0.135	0.324	0.021	0.3%
Anti-solidarity	22.4%	0.105	0.397	0.027	0.2%
Mixed	7.4%	0.033	0.835	0.113	1.6%

Appendix

Table 25: High-level diagnostics of DSL stability across 300 simulated runs. Labels are ordered from lowest to highest Pearson SD. See the column definitions below.

Label	Full prevalence	HumanSoft trend SD	Seed variation / human signal	Pearson SD	Sampling-miss rate
Group-based AS	13.4%	0.136	0.417	0.034	2.3%
Compassionate S	32.7%	0.134	0.509	0.040	0.0%
Group-based S	12.5%	0.069	0.766	0.095	0.8%
Compassionate AS	5.5%	0.049	0.685	0.095	13.3%
Exchange-based S	3.6%	0.031	0.972	0.121	16.7%
Exchange-based AS	3.0%	0.022	1.054	0.134	21.2%

Appendix

Table 26: Fine-grained diagnostics of DSL stability across 300 simulated runs. Labels are ordered from lowest to highest Pearson SD. See the column definitions below.

Panel of Figure 26 — Figure 26: Party-specific DSL-adjusted yearly shares of (anti-)solidarity subtypes, 2009-2025. This figure provides a larger-format version of the party-specific panels shown in Figure 16 . Within each year and party, subtype shares are renormalized across the eight solidarity and anti-solidarity categories so that they sum to 100%.

Panel of Figure 27 — Figure 27: Party-specific DSL-adjusted yearly shares of (anti-)solidarity subtypes, 2009–2025 (continued).

Figure 28: One-step zero-shot prompt set used for the evaluation of GPT-4 on the Migrant dataset.

Figure 29: Two-step few-shot prompt set used for the evaluation of open-weight models on the Migrant dataset and for the final large-scale annotation with Llama-3.3-70B. The model is first given the high-level classification prompt in Figure 29(a) . If the output is Solidarity or Anti-solidarity, the corresponding subtype prompt in Figure 30(a) or Figure 31(a) is used.

Figure 30: Two-step few-shot prompt set used for the evaluation of open-weight models on the Migrant dataset and for the final large-scale annotation with Llama-3.3-70B (continued from Figure 29 ).

Figure 31: Two-step few-shot prompt set used for the evaluation of open-weight models on the Migrant dataset and for the final large-scale annotation with Llama-3.3-70B (continued from Figure 29 ).

Panel of Figure 32 — Figure 32: Solidarity trends for Migrant speeches computed using DSL-corrected soft labels (see definition (c) in Section 3.2.2 ) on the restricted corpus containing only the keyword groups Flüchtlinge and Ausländer . Figure 32(b) shows the high-level (anti-)solidarity trends, while Figure 32(a) shows the distribution of fine-grained (anti-)solidarity subtypes over time. In the fine-grained panel, percentages are normalized within each decade across the eight solidarity and anti-solidarity subtypes. The period 1930-1949 (grey area) is excluded due to limited data during and after the Nazi dictatorship.

Label	GPT-5	GPT-4	gpt-oss 120B	Qwen 2.5-72B	Llama 3.3-70B	Llama 4-Scout	Mistral 2-Large	Gemma 2-27B	BERT
Group-based S	0.42	0.33	0.39	0.41	0.32	0.36	0.34	0.34	0.31
Exchange-based S	0.43	0.60	0.56	0.40	0.36	0.46	0.41	0.37	0.22
Empathic S	0.48	0.05	0.28	0.19	0.11	0.17	0.38	0.24	0.00
Compassionate S	0.60	0.50	0.48	0.50	0.53	0.61	0.43	0.45	0.53
Solidarity (high-level)	0.81	0.80	0.78	0.68	0.71	0.78	0.54	0.63	0.67
Group-based AS	0.47	0.51	0.49	0.42	0.46	0.43	0.33	0.42	0.32

Appendix

Table 27: F1 scores for selected models on fine-grained and high-level solidarity classification on the combined Migrant test sets (Test 1 and Test 2). The top section contains fine-grained categories, and the bold rows show performance on high-level classification. Scores are chosen for the best performing configuration for the model in terms of Macro F1.

Keyword	Solidarity	Anti-Solidarity	Mixed	None	N (statements)
Asylsuchende	4.11	2.06	0.48	3.35	10
Ausländer	291.43	118.86	21.43	326.28	758
Aussiedler	118.43	12.48	15.22	25.87	172
Einwanderer	2.84	3.21	0.20	19.75	26
Emigranten	21.14	11.08	2.61	10.18	45
Flüchtlinge	1917.08	321.69	174.29	859.94	3273

Appendix

Table 28: Distribution of DSL-adjusted expected probability mass for solidarity, anti-solidarity, mixed, and none per keyword included in the post-war trends analysis Section 4.2 . Values reflect summed DSL-adjusted expected counts. Totals correspond to the number of keyword-coded statements in the analyzed subset.

Party	Solidarity	Anti-Solidarity	Mixed	None	N (statements)
BP	134.13	54.77	67.51	46.60	303
CDU/CSU	2040.48	229.98	94.93	1050.62	3416
DP	138.21	40.81	-0.54	26.52	205
DRP	34.53	1.47	1.74	3.26	41
FDP	539.79	38.02	56.78	251.41	886
GB/BHE	318.75	27.61	9.03	17.62	373

Appendix

Table 29: Distribution of DSL-adjusted expected probability mass for solidarity, anti-solidarity, mixed, and none per party included in the post-war trends analysis Section 4.2 . Values reflect summed DSL-corrected expected counts. Totals correspond to the number of party-coded statements in the analyzed subset.

Original Text	Translation
Example (1) : Predictions: none ( gpt-oss , Qwen , ensemble); exchange-based solidarity ( GPT-5 , Llama ). Human reference: exchange-based solidarity ; human votes: 3 $\times$ exchange-based solidarity , 1 $\times$ mixed (5 Dec 1913)
“ […] das Interessante […] ist, daß nicht nur die östliche Landwirtschaft sondern auch die Industrie Ausländer braucht, daß die Zahlen der industrietätigen Ausländer im Lause der Jahre viel mehr gewachsen sind als die Zahlen der von der Landwirtschaft gebrauchten Arbeiter. […] Das Maß des Bedürfnisses der Beschäftigung von Ausländern ist natürlich auch hier örtlich und nach Betrieben verschieden. Es gibt gewisse Betriebe, für die inländische Arbeiter kaum noch zu bekommen sind. […] Es gibt aber auch industrielle Gebiete, die ohne diese Ausländer gar nicht weiter existieren können. ”	“ […] what is interesting […] is that not only eastern agriculture but also industry needs foreigners, and that the number of foreigners employed in industry has grown over the years much more than the number of workers needed in agriculture. […] The extent of the need for the employment of foreigners naturally also varies here by locality and by sector. There are certain sectors in which domestic workers can scarcely be found anymore. […] There are also industrial regions that simply could not continue to exist without these foreigners. ”
Example (2) : Predictions: mixed ( GPT-5 , Llama ), group-based anti-solidarity ( gpt-oss , ensemble), exchange-based solidarity ( Qwen ). Human reference: mixed ; human votes: 3 $\times$ mixed , 1 $\times$ group-based solidarity (15 June 2018)
“[…] Die Beschleunigung der Asylverfahren ist überfällig, und eine konsequente Rückführung […], und […] eine konsequente Abschiebung. Übrigens sind zügige Verfahren auch und gerade im Interesse der Menschen mit Bleibeperspektive, die wir dann umso besser integrieren können, weil wir ihnen eine Perspektive geben können. Zu guter Letzt müssen wir die Einwanderung nach Deutschland insgesamt neu ordnen, mit einem Einwanderungsgesetz, das diesen Namen verdient. Herr Kollege Castellucci, möglicherweise wird umgekehrt ein Schuh daraus: Wer Akzeptanz für legale Migration erhöhen will, der muss auch entschlossen illegale Migration bekämpfen. […]”	“Speeding up asylum procedures is overdue, as are consistent returns […] and […] consistent deportations. Swift procedures are also, and especially, in the interest of people who have a prospect of staying, because we can then integrate them all the better by giving them a perspective. Finally, we must reorganize immigration to Germany as a whole, with an immigration law worthy of the name. Mr Castellucci, perhaps the reverse is true: anyone who wants to increase acceptance of legal migration must also resolutely combat illegal migration. […]”
Example (3) : Predictions: compassionate solidarity ( gpt-oss , GPT-5 , ensemble), group-based anti-solidarity ( Llama ), exchange-based anti-solidarity ( Qwen ). Human reference: compassionate solidarity ; human votes: 2 $\times$ compassionate solidarity (19 Jan 1995)
“[…] Sie [die europäischen Staaten] sind aber nicht bereit, Flüchtlinge vor Menschenrechtsverletzungen zu schützen, sondern diffamieren sie als Schmarotzer und Bedrohung unserer Sicherheit, wie dies leider auch gerade durch den Kollegen der SPD wieder geschehen ist. Der Vertrag von Maastricht sieht die gemeinsame Einführung eines Visumzwangs vor, wenn auf Grund einer Notlage in einem Land eine größere Fluchtbewegung zu erwarten ist. So wird ganz bewußt verfolgten oder unter sonstigen Notlagen leidenden Menschen die Flucht erheblich erschwert, oft sogar unmöglich gemacht. Flüchtlinge werden damit geradezu in die Arme von Schlepperorganisationen getrieben, über die man sich hier so gerne beschwert.”	“[…] But they [the European states] are not willing to protect refugees from human rights violations; instead, they defame them as parasites and a threat to our security, as unfortunately just happened again in the remarks of the SPD colleague. The Maastricht Treaty provides for the joint introduction of a visa requirement if, because of an emergency in a country, a larger refugee movement is to be expected. In this way, it is deliberately made much harder, often even impossible, for people who are persecuted or otherwise suffering hardship to flee. Refugees are thereby virtually driven into the arms of smuggling organizations, about which people here so readily complain.”

Appendix

Table 30: Error analysis examples. Bold text is the main sentence, the other sentences are for context. The reference label is the final human consensus label. Where relevant, we also report the distribution of individual human annotations.

Original Text	Translation
Example (4) : Predictions: group-based solidarity ( gpt-oss , Qwen ), group-based anti-solidarity ( Llama ). Human reference: group-based solidarity (10 Apr 2024)
“Kommen wir zum Entwurf der AfD und zu dem, was ihn einfach grottenschlecht macht. Sie fordern, dass in allen Fällen, in denen ein Elternteil nichtdeutscher Herkunft ist, die Ausländerbehörde beteiligt wird. Nicht nur stellen Sie damit alle Ausländer/-innen unter Generalverdacht […], sondern Sie verstoßen damit auch gegen EU-Recht […]. Sie ‘übersehen’ in Ihrem Entwurf, dass ein Großteil der Ausländer, die hier die Vaterschaftsanerkennung beantragen, längst eine Niederlassungserlaubnis haben. Aber gut, dass Sie gegen unsere ausländischen Mitbürgerinnen und Mitbürger hetzen. Ist ja wirklich Ihr politisches Tagesgeschäft, so bedauerlich das auch ist. […] ”	“[…] Let us turn to the AfD’s draft and to what makes it simply atrocious. You demand that in all cases where one parent is of non-German origin, the immigration authority must be involved. Not only does this place all foreigners under general suspicion […], but it also violates EU law […]. You overlook in your draft that a large proportion of the foreigners who apply here for recognition of paternity already have permanent residency. But of course, it fits your agenda to agitate against our foreign fellow citizens — that is really your daily political business, as regrettable as it is. […]”
Example (5) : Predictions: group-based solidarity ( GPT-5 , Qwen ); compassionate solidarity ( gpt-oss , Llama , ensemble). Human reference: compassionate solidarity ; human votes: 2 $\times$ group-based solidarity , 2 $\times$ compassionate solidarity (5 May 1960)
“ […] Diese Rückführung [der Evakuierten] muß beschleunigt werden, und wir müssen gesetzgeberische Maßnahmen ergreifen, damit die Evakuierten nicht immer wieder auf lange Zeit vertröstet werden, sondern in absehbarer Zeit zu ihren Rechten, zu einer menschenwürdigen Behandlung kommen. […] Gerade das Vertrauen dieser Millionen von Menschen verpflichtet uns, sie auch in den nächsten Jahren nicht zu enttäuschen, […] diesen Eingliederungsprozeß auf allen Gebieten mit dem größten Nachdruck zu fördern […]. In der politischen Diskussion hat zwar das Problem der sachlichen, wirtschaftlichen und sozialen Interessen der Vertriebenen und aller Kriegsgeschädigten kaum noch eine Rolle gespielt […]. ”	“ […] This return [of evacuees] must be accelerated, and we must take legislative measures so that evacuees are not kept waiting indefinitely time and again, but instead obtain their rights and humane treatment within a foreseeable period of time. […] It is precisely the trust of these millions of people that obliges us not to disappoint them in the years ahead either, […] and to promote this process of integration in every sphere with the greatest possible determination […]. In political debate, the problem of the material, economic, and social interests of the expellees and all other victims of war has in fact hardly played any role any longer […]. ”
Example (6) : Predictions: group-based solidarity ( GPT-5 , Qwen , gpt-oss , Llama , ensemble). Human reference: compassionate solidarity ; human votes : 4 $\times$ compassionate solidarity (28 Oct 2010)
“Die Bundesrepublik hat […] einen Vorbehalt zur UN-Kinderrechtskonvention zurückgenommen, der ihr zusichern sollte, dass sie ausländische Kinder schlechter behandeln kann als inländische. […] In den vorliegenden Anträgen […] ist dargelegt, was alles zu tun wäre: Heraufsetzung der Verfahrensmündigkeit von minderjährigen Flüchtlingen von 16 auf 18 Jahre, Ausbau und Verbesserung der Betreuung unbegleiteter minderjähriger Flüchtlinge durch die Jugendfürsorge, ihre Herausnahme aus dem Flughafenverfahren, keine Abschiebehaft für Minderjährige, keine Unterbringung von Kindern und Jugendlichen in Sammelunterkünften für Asylsuchende und Flüchtlinge. Unbegleitete Kinder und Jugendliche, deren Eltern nicht ausfindig gemacht werden können, sollen genauso behandelt werden wie elternlose deutsche Kinder. […] ”	“The Federal Republic has […] withdrawn a reservation to the UN Convention on the Rights of the Child that had ensured it could treat foreign children worse than German children. […] The motions submitted […] outline what still needs to be done: raising the age of legal capacity in proceedings for minor refugees from 16 to 18; expanding and improving youth welfare support for unaccompanied minor refugees; removing them from the airport procedure; no detention pending deportation for minors; no accommodation of children and adolescents in collective facilities for asylum seekers and refugees. Unaccompanied children and adolescents whose parents cannot be located should be treated in the same way as German children without parents. […] ”

Appendix

Table 31: Error analysis examples (continued). Bold text is the main sentence, the other sentences are for context. The reference label is the final human consensus label. Where relevant, we also report the distribution of individual human annotations.

Original Text	Translation
Example (7) : Predictions: group-based ( gpt-oss , ensemble), exchange-based ( Qwen , GPT-5 ) or compassionate solidarity ( Llama ). Human reference: compassionate solidarity ; human votes : 3 $\times$ compassionate solidarity , 1 $\times$ group-based solidarity (16 Jan 2020)
“Das Hamburger Modell ist geradezu vorbildlich; denn hier werden nicht Menschen mit geringem Einkommen und Geflüchtete gegeneinandergestellt, sondern alle haben den gleichen Anspruch auf Mobilität, mit dem Unterschied, dass die Geflüchteten das von ihrem Taschengeld zwangsweise abgezogen bekommen. […] ‘Wir sind auf Migration angewiesen, aber die Flüchtlinge…’ , könnte man sagen . […] Ich bin so erzogen: Wenn jemand Hilfe braucht und du kannst helfen, dann hilf. Ich bin stolz und froh, dass unser Land helfen kann […].”	“The Hamburg model is virtually exemplary, because here people with low incomes and refugees are not being pitted against one another; rather, everyone has the same entitlement to mobility, with the difference that refugees have it forcibly deducted from their allowance. […] One might say, ‘We depend on migration, but the refugees…’ . […] I was raised to believe this: if someone needs help and you can help, then help. I am proud and glad that our country is able to help […].”
Example (8) : Predictions: group-based solidarity ( gpt-oss , Qwen , Llama , ensemble), empathic solidarity ( GPT-5 ). Human reference: empathic solidarity ; human votes : 3 $\times$ empathic solidarity , 1 $\times$ group-based solidarity (12 Feb 1965)
“[…] Die große Aufgabe, die uns hier gesetzt war, war der Versuch, einen Kompromiß […] zwischen den Interessen der Ausländer in der Bundesrepublik Deutschland und den berechtigten Ansprüchen der deutschen Bevölkerung auf Aufrechterhaltung der öffentlichen Ordnung und Sicherheit zu finden. Das Gesetz ist ein Ausdruck unserer freiheitlichen demokratischen Grundordnung, indem es den Ausländern in der Bundesrepublik ein sehr hohes Maß von Entfaltungsmöglichkeiten gewährt. […] Das Gesetz nimmt auch besonders Bedacht auf eine Verbesserung der Rechtslage der ausländischen Flüchtlinge […], die ihre Heimat unter dem Druck totalitärer Systeme verlassen mußten. […]”	“[…] The major task before us was to try to find a compromise […] between the interests of foreigners in the Federal Republic of Germany and the legitimate claims of the German population to the maintenance of public order and security. This law is an expression of our liberal democratic constitutional order in that it grants foreigners in the Federal Republic a very high degree of scope for self-development. […] The law also gives special consideration to improving the legal position of foreign refugees […] who were forced to leave their homeland under pressure from totalitarian systems. […]”
Example (9) : Predictions: exchange-based anti-solidarity ( gpt-oss , Qwen , GPT-5 , ensemble), group-based anti-solidarity ( Llama ). Human reference: group-based anti-solidarity ; human votes : 2 $\times$ group-based anti-solidarity , 2 $\times$ exchange-based anti-solidarity (1 June 2017)
“[…] Mit diesem auf Angebotsorientierung im Migrationsrecht zielenden Gesetzentwurf wollen die Grünen im Kern nichts anderes, als die Einwanderung massiv vereinfachen und die Tore weit aufstoßen. Wir als Unionsfraktion stellen uns gegen die Beliebigkeit im Einwanderungsrecht und stehen daher auch klar für eine gezielte Auswahl bei der Migration ein. Der Staat und die Politik haben die Pflicht, bewusst darauf zu achten, dass die Migration für unser Land vor allem einen Mehrwert und keine Belastung darstellt, und genau darin liegt der Unterschied zum Asyl- und Flüchtlingsrecht. […] ”	“[…] With this bill, which is aimed at a supply-oriented approach to migration law, the Greens essentially want nothing other than to make immigration much easier and throw the gates wide open. As the CDU/CSU parliamentary group, we oppose arbitrariness in immigration law and therefore clearly support a targeted selection in migration. The state and policymakers have a duty to ensure consciously that migration represents above all an added value for our country and not a burden, and that is precisely the difference from asylum and refugee law. […]”

Appendix

Table 32: Error analysis examples (continued). Bold text is the main sentence, the other sentences are for context. The reference label is the final human consensus label. Where relevant, we also report the distribution of individual human annotations.

Original Text	Translation
Example (10) : Predictions: group-based anti-solidarity ( gpt-oss , Qwen , GPT-5 , Llama , ensemble). Human reference: empathic anti-solidarity ; human votes : 2 $\times$ group-based anti-solidarity , 2 $\times$ empathic anti-solidarity (17 May 2019)
“52 Prozent der britischen Muslime sagen nicht nur: ‘Homosexualität ist moralisch falsch’ , sondern: Homosexualität sollte verboten werden. […] bei fortgesetzter Einwanderung [wird sich] der Anteil der Muslime in Deutschland in den kommenden Jahren auf 20 Prozent erhöhen. In vielen Städten und Gemeinden sind sie dann die Mehrheit. Der Islam kennt keine Minderheitenrechte. Das sind die dunklen Wolken, die über den Homosexuellen in Westeuropa aufziehen.”	“52 percent of British Muslims say not only, ‘Homosexuality is morally wrong’ , but also: homosexuality should be banned. […] if immigration continues, the share of Muslims in Germany will rise to 20 percent in the coming years. In many cities and municipalities, they will then be the majority. Islam recognizes no minority rights. These are the dark clouds gathering over homosexuals in Western Europe.”
Example (11) : Predictions: none ( gpt-oss , Qwen , GPT-5 , ensemble), group-based anti-solidarity ( Llama ). Human reference: group-based solidarity ; human votes : 2 $\times$ group-based solidarity , 1 $\times$ group-based anti-solidarity (16 Apr 1868)
“Ausländer bleibt unerachtet solchen Aufenthalts innerhalb der Mecklenburgischen Lande immer Ausländer, und einem solchen die Trauung innerhalb der diesseitigen Lande zu gestatten, ist allein das Großherzogliche Ministerium des Innern competent. […] Vielmehr muß in solchen Fällen der Wiederverheirathung der Mecklenburger sein Recht zur Verheiratung durch einen gesetzlichen Domizilschein über das ihm für einen Mecklenburgischen Ort ertheilte Recht der Niederlassung und Verheiratung nachweisen und der Ausländer den Konsens des Großherzoglichen Ministeriums des Innern beibringen, bevor die Trauung geschehen darf. Ja, meine Herren, da sehen Sie nun, was in Mecklenburg das Indigenat aus dem Artikel 2 der Bundesverfassung für eine Bedeutung hat, und wie nothwendig es ist, daß irgend eine Ergänzung des Freizügigkeitsgesetzes […] erfolge, damit das Gesetz dort nicht blos eine Schale ohne Kern bleibe.”	“A foreigner remains a foreigner within the lands of Mecklenburg despite such residence, and only the Grand Ducal Ministry of the Interior is competent to permit such a person to marry within these territories. […] Rather, in such cases of remarriage, the Mecklenburger must prove his right to marry by means of a lawful certificate of domicile showing that he has been granted the right of residence and marriage in a Mecklenburg locality, and the foreigner must produce the consent of the Grand Ducal Ministry of the Interior before the marriage may take place. Gentlemen, there you can see what significance indigeneity under Article 2 of the federal constitution has in Mecklenburg, and how necessary it is that some supplementation of the Freedom of Movement Act […] should be made, so that the law there may not remain merely a shell without a kernel.”
Example (12) : Predictions: none ( gpt-oss , Qwen , Llama , ensemble), mixed ( GPT-5 ). Human reference: group-based solidarity ; human votes : 2 $\times$ group-based solidarity , 1 $\times$ none , 1 $\times$ mixed (15 Jan 1886)
“Wann und warum ein Ausländer ohne ein vorangegangenes gerichtliches Erkenntniß aus dem Reich oder einem Partikularstaat ausgewiesen werden darf, darüber existiren zwar keine gesetzlichen Normen und Bestimmungen […]. Rechtfertigende Gründe der Ausweisung eines Ausländers nach dem allgemein anerkannten Völkerrechte sind folgende: erstens Gefährdung der öffentlichen Sicherheit durch ein feindliches Auftreten des Ausgewiesenen […]; zweitens […]; drittens[…], und viertens […]. Damit sind alle Fälle erschöpft, in welchen nach dem modernen Völkerrecht Ausweisungen von einzelnen Personen stattfinden dürfen, und alle Lehrer des Völkerrechts sind um so vorsichtiger in der Empfehlung auch nur eines dieser Mittel, als sich das moderne Völkerrecht immer mehr der Theorie zuwendet, daß ein domizilirter Ausländer zu den wirklichen Staatsangehörigen gehört und als solcher zu behandeln ist. […]”	“Although there are no legal norms or provisions governing when and why a foreigner may be expelled from the Empire or from an individual state without prior judicial judgment […], the reasons justifying the expulsion of a foreigner under generally recognized international law are the following: first, endangerment of public security through hostile conduct on the part of the person expelled […]; second […]; third […]; and fourth […]. These exhaust all the cases in which, under modern international law, expulsions of individual persons may take place, and all teachers of international law are all the more cautious in recommending even one of these measures because modern international law is increasingly turning toward the theory that a domiciled foreigner belongs among the true members of the state and is to be treated as such. […]”

Appendix

Table 33: Error analysis examples (continued). Bold text is the main sentence, the other sentences are for context. The reference label is the final human consensus label. Where relevant, we also report the distribution of individual human annotations.

	Original Text	Translation
(1) Dominant DSL subtype score: group-based solidarity towards expellees (May 14, 1952)	“[…] das Lastenausgleichsgesetz und die Eingliederung der Vertriebenen [steht] in keinem Widerspruch zu unserm Anspruch und unserm Recht auf die Heimat. Im Gegenteil, […] daß die Rücksiedlung […] nur durchgeführt werden kann, wenn die körperliche, seelische und auch die wirtschaftliche Substanz der Vertriebenen erhalten geblieben ist, und das kann nur geschehen, wenn sie in die westdeutsche Wirtschaft eingegliedert werden. […] Wir sind durchaus für die Ausgestaltung des Lastenausgleichs nach sozialen Gesichtspunkten […]. ”	“[…] the Equalization of Burdens Act and the integration of the expellees does not contradict our right to the homeland. On the contrary, resettlement […] can only be achieved if the physical, mental, and economic substance of the expellees has been preserved, and that can only happen if they are integrated into the West German economy. […] We are fully in favor of shaping the Equalization of Burdens in accordance with social considerations […]. ”
(2) Dominant DSL subtype score: group-based solidarity towards expellees (Feb 25, 1953)	“[…] Der Ihnen vorliegende Entwurf [zum Bundesvertriebenengesetz] geht […] von der Zielsetzung aus, die […] Leistungskraft der Vertriebenen und Flüchtlinge […] im Rahmen der Gesamtwirtschaft zur Geltung zu bringen. Die Eingliederungsbestimmungen sollen der Erfüllung berechtigter Ausgleichsansprüche hinsichtlich der verlorenen Existenz dienen […] mit dem Ziel, daß sich das gesamte Arbeitspotential, das der Einheimischen und der Vertriebenen, so betätigen kann, daß eine größtmögliche Wirkung erzielt wird.”	“[…] The draft [on the Federal Expellees Act] before you starts from the objective […] of making […] productive capacities of the expellees and refugees […] effective within the overall economy. The integration provisions are intended to meet legitimate compensation claims with regard to lost livelihoods […] with the goal that the entire labor potential, that of natives and of expellees alike, can be deployed in such a way that the greatest possible effect is achieved.”
(3) Dominant DSL subtype score: compassionate solidarity towards Soviet-zone refugees (June 15, 1955)	“[…] es ist notwendig, zu den noch immer vielseitigen Problemen der Sowjetzonenflüchtlinge Stellung zu nehmen […]. Wir erkennen selbstverständlich an, daß durch die Neuregelung des Härtefonds im Lastenausgleich eine Verbesserung für die Sowjetzonenflüchtlinge erfolgte, weil die Ermessensfragen, unter denen dieser Personenkreis bislang besonders zu leiden hatte, dadurch eingeschränkt werden. […] Wenn wir aber die Sicherheit haben wollen, daß sich diese Neuregelung auch wirklich positiv für den Flüchtling auswirkt, so ist es notwendig, daß die vorgesehene Erhöhung des Härtefonds […] wirklich realisiert wird.”	“[…] it is necessary to address the still manifold problems of the Soviet-zone refugees […]. We naturally recognize that the reform of the hardship fund in the Equalization of Burdens has brought improvements […] because the discretionary decisions that this group had particularly suffered from have been curtailed. But if we want to be sure that this reform really has a positive effect for the refugee, then […] the planned increase of the hardship fund […] must actually be implemented.”
(4) Dominant DSL subtype score: group-based anti-solidarity towards expellees (Dec 13, 1950)	“In unseren Landkreisverwaltungen sind heute 42,2% Heimatvertriebene. […] Auch die einheimische Bevölkerung beklagt sich […] man höre auf den Landratsämtern überwiegend ostpreußisch, und das Plattdeutsche sei zur Seite gedrängt. Unter den Lehrern […] 51% Heimatvertriebene, unter den vollbeschäftigten Richtern […] mehr als 50%, unter der Polizei sogar 61%; zweifellos eine Überfremdung, die Besorgnis erregt! Aber all diese ehrlichen Bemühungen um Eingliederung der Heimatvertriebenen sind angesichts der ungeheuren Zahl vergeblich. Arbeitslosenzahl […] beträgt unter den Heimatvertriebenen fast 60%. Die Lasten sind unerschwinglich.”	“In our district administrations today, 42.2% are expellees. The native population […] complains […] that in the district offices one mostly hears East Prussian, and that Low German has been pushed aside. Among the teachers 51% are expellees, among the full-time judges also more than 50%, among the police even 61% – undoubtedly an over-foreignization that arouses concern! But all these honest efforts to integrate the expellees are in vain in view of their immense numbers. Unemployment […] almost 60% among the expellees […]. The burdens are unbearable.”

Appendix

Table 34: Examples for postwar period discussed in Section 4.2 . Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(5) Dominant DSL subtype score: group-based/compass. solidarity towards Soviet-zone refugees (Sept 23, 1954)	“[…] [Sowjetzonenflüchtlinge] meinen, es könnte ihnen sonst bei mangelnder Bereitwilligkeit der Bevölkerung zu neuen Opfern vielleicht nicht einmal mehr das zur Verfügung stehen, was ihnen heute aus dem Härtefonds zur Verfügung steht. […] Wir halten es für unvertretbar, daß […] diese nach wie vor auf den genannten, im Rahmen des Gesamtlastenausgleichs verhältnismäßig geringen Betrag ohne Rechtsanspruch angewiesen sein sollen. [Die], deren Urstatus dem der Heimatvertriebenen in Westdeutschland doch gleich ist, im Lastenausgleich diesen Heimatvertriebenen nicht gleichgestellt sind.”	“[…] [Soviet-zone refugees] fear that, if public willingness for further sacrifices declines, they may no longer receive even the support currently provided through the hardship fund […]. [It is] indefensible that […] they should still be dependent on the aforementioned, comparatively small amount within the overall equalization scheme without any legal entitlement. […] those whose original legal status is equivalent to that of expellees in West Germany are not treated equally under the Equalization of Burdens system.”
(6) Dominant DSL subtype score: group-based anti-solidarity towards foreigners (June 21, 1956)	“[…] Wir müssen das Bundesgebiet nicht nur gegen politisch subversive Elemente, wir müssen uns auch gegen kriminelle Elemente abschirmen . […] die Bundesrepublik […] [hat] Anlaß, sich gegen Zehntausende von heimatlosen Ausländern abzuschirmen, die […] in der Absicht [kommen], bei uns wirtschaftliche Versorgung oder politisches Asyl zu finden. […] alle internationalen Verpflichtungen, die wir […] erfüllen, [gehören] selbstverständlich nicht unter das, was ich hier ausgeführt habe.”	“[…] We must shield the Federal territory not only against politically subversive elements, we but also against criminal elements . […] the Federal Republic […] has reason to shield itself against […] stateless foreigners who […] come with the intention of seeking economic support or political asylum here. […] all international obligations which we […] fulfill, of course, do not fall under what I have set out here.”
(7) Dominant DSL subtype score: group-based solidarity towards foreigners (Oct 18, 1950)	“[…] [HAuslG] […] [wird] den Besatzungsmächten den Entschluß erleichtern, der Bundesrepublik die volle Verantwortung für die heimatlosen Ausländer zu übertragen. Wenn […] das deutsche Volk seine Bereitschaft bekundet, die Eingliederung der heimatlosen Ausländer in die deutsche Betreuung […] zu erleichtern, so gehen wir dabei von der Erwartung aus, daß auch die heimatlosen Ausländer bemüht sein werden, sich in die nun einmal gegebene Lage einzufügen. Sie werden sich dabei bewußt sein müssen, daß auch die Lage der deutschen Bevölkerung schwierig ist und daß viele Zehntausende insbesondere von Heimatvertriebenen noch immer arbeitslos sind und über keinen angemessenen Wohnraum verfügen. […] der heimatlose Ausländer [wird] viele der Härten dieser Nachkriegszeit mit den Deutschen im Gebiete der Bundesrepublik teilen müssen. ”	“ […] [HAuslG] […] will make it easier for the occupying powers to decide to transfer full responsibility for stateless foreigners to the Federal Republic. If […] the German people declare their willingness to facilitate […] the integration of stateless foreigners into German administration and care, we proceed on the assumption that the stateless foreigners will also strive to adapt to the situation as it now stands. They must remain aware that the situation of the German population is also difficult and that many tens of thousands, especially expellees, remain unemployed and lack adequate housing. […] stateless foreigners will have to share many of the hardships of this postwar period with the Germans in the territory of the Federal Republic. ”
(8) Dominant DSL subtype score: group-based solidarity towards foreigners (Oct 18, 1950)	“[…] Auch in der öffentlichen Fürsorge sollen sie Leistungen in gleicher Höhe erhalten wie deutsche Staatsangehörige. Es soll ihnen Zutritt zu Schulen aller Art zu den gleichen Bedingungen… gewährt werden. […] die Bundesrepublik [gewährt] ihnen einen Rechtsschutz, dessen sie bedürfen, weil sie nicht mehr den Schutz ihres Heimatstaates genießen […]. Der Gesetzentwurf geht […] über den Konventionsentwurf, dessen Grundsätze Mindestforderungen für die Behandlung heimatloser Ausländer darstellen, hinaus. […] ”	“[…] In public welfare too, they are to receive benefits equal to those of German citizens. They are to be granted access to schools of all kinds under the same conditions […]. the Federal Republic shall grant them legal protection, which they need because they no longer enjoy the protection of their home state. The law goes […] beyond the draft convention, whose principles set minimum standards for the treatment of stateless foreigners. […] ”

Appendix

Table 35: Examples for postwar period discussed in Section 4.2 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(9) Dominant DSL subtype score: exchange-based solidarity towards foreigners (April 19, 1951)	“[…] Wir begrüßen die Verabschiedung des Gesetzes für heimatlose Ausländer, weil uns damit die Möglichkeit geboten war, dem Ausland zu beweisen, wie wir Angehörige anderer Nationen in unserem eigenen Staate zu behandeln gewillt sind, und weil wir uns […] erhoffen, daß sie […] dazu beitragen wird, die Frage der Internationalisierung unseres Flüchtlingsproblems und des Zuflusses von Mitteln aus dem Ausland zu fördern. […]”	“[…] We welcome the adoption of the law for stateless foreigners, because it gave us the opportunity to demonstrate to foreign countries how we intend to treat nationals of other nations within our own state, and because we […] hope that it will make a contribution to promoting the internationalization of our refugee problem and the inflow of funds from abroad. […]”
(10) Dominant DSL subtype score: group-based solidarity towards foreigners (June 15, 1955)	“[…] Es ist kein Zweifel, daß die heimatlosen Ausländer natürlich vom Arbeitsministerium und vom Innenministerium […] eine rein materielle Unterstützung bekommen. Aber diese materielle Unterstützung gibt keine Möglichkeit, ihnen kulturell oder sonstwie unter die Arme zu greifen. […] [Es ist sehr schwer] diese Dinge heute den Ländern nahezulegen.”	“[…] There is no doubt that the stateless foreigners, of course, receive […] purely material support from the Ministry of Labor and the Ministry of the Interior. But this material support provides no opportunity to assist them culturally or in other ways. […] [It is very difficult] today to persuade the Länder on these matters. […]”
(11) Dominant DSL subtype score: exchange-based solidarity towards expellees, Party: FDP (Oct 20, 1949)	“Die werden […] von früh morgens bis spät abends arbeiten, wenn man ihnen das nötige Gelände, [Baumaterialien oder das Geld] zur Verfügung stellt. So werden sie in Gemeinschaftsarbeit wesentlich mehr leisten, als wir das sonst heute bei diesen Bauten sehen. In diesem Falle wäre das Geld, das wir dafür aufnehmen, praktisch nur als Kredit anzusehen. Wir hätten die Möglichkeit, daß diese Heimatvertriebenen sich ihr Heim selbst bauen, daß es allmählich ihr Eigentum wird. Wenn wir nur normale Mietsätze verrechnen und das Kapital ohne Zinsen oder zu geringen Zinssätzen verzinsen lassen, so wäre in wenigen, spätestens 15 Jahren, dieses Haus […] Eigentum des Betreffenden. ”	“They would […] work from early morning until late at night if they were provided with the necessary land, [building materials or the money]. In cooperative work they would achieve far more than we otherwise see with such construction today. In this case the money we raise would practically only be regarded as a loan. We would have the possibility that these expellees build their homes themselves, gradually becoming their property. If we only charged normal rents and let the capital bear no or low interest, then within at most fifteen years this house […] would become the property of the person concerned.”
(12) Dominant DSL subtype score: group-based/ compassionate solidarity towards expellees, Party: GB/BHE (Oct 29, 1953)	“[…] Wir dürfen die Augen auch nicht vor der Tatsache verschließen, daß Millionen von Vertriebenen und Kriegsgeschädigten noch nicht den Anschluß an unsere soziale Entwicklung gefunden haben. Auch hier mußten bisher Zehntausende arbeitswilliger tüchtiger Mitbürger abseitsstehen. Bei unserem Verlangen nach beschleunigter Eingliederung dieser Bevölkerungskreise geht es nicht nur um die Ablösung einer unproduktiven Fürsorge durch aufbauende Mitarbeit. Es geht auch nicht nur um die Erfüllung von ‘Rechtsansprüchen’ , und ich darf diesen Ausdruck Rechtsansprüche hier noch einmal betonen. ”	“[…] We must not close our eyes to the fact that millions of expellees and war victims have still not been integrated into our social development. Here too, tens of thousands of willing and capable fellow citizens have so far had to stand aside. In our demand for the accelerated integration of these groups, it is not only a matter of replacing unproductive welfare with constructive participation. It is also not only a matter of fulfilling legal claims — and I would like to emphasize this expression ‘legal claims’ once again.”

Appendix

Table 36: Examples for postwar period discussed in Section 4.2 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(13) Dominant DSL subtype score: compassionate solidarity towards refugees, Party: WAV (March 23, 1950)	“Die Aufnahme dieser Flüchtlinge wird in manchen Ländern bewußt erschwert, sabotiert und unmöglich gemacht. […] Seitdem die Spannung zwischen Osten und Westen zunimmt, hat ein neuer Strom von Flüchtlingen eingesetzt. […] Polen und die Tschechei [durchführen] die neue Ausweisungsaktion auf Befehl Moskaus, […] um das Flüchtlingsheer in Westdeutschland zu vermehren und noch mehr Arbeitslose zu schaffen, den Staat zu belasten und seine Wirtschaft zu zerstören. […]”	“The admission of these refugees is being deliberately obstructed, sabotaged, and made impossible in some states. […] Since the tension between East and West has increased, a new stream of refugees has begun. […] Poland and Czechoslovakia are carrying out the new expulsions on orders from Moscow, […] in order to enlarge the refugee army in West Germany, to create even more unemployed, to burden the state, and to destroy its economy. […]”
(14) Dominant DSL subtype score: group-based anti-solidarity towards expellees, Party: BP (Oct 20, 1949)	“Die Sudetendeutschen […] und die Südostdeutschen [stehen] dem süddeutschen, also dem bayerischen Kulturkreis am nächsten. […] Die aus dem Sudetenland bzw. aus Südosteuropa kommenden Deutschen [sollen] zunächst in Süddeutschland zusammen[gefasst werden], […] damit die Eingliederung […] schneller und reibungsloser vor sich gehen kann. [Wir] wollen […] nicht Schwierigkeiten schaffen, wo Schwierigkeiten nicht überwunden werden müssen, und nicht Gräben aufreißen, wo solche nicht vorhanden sind. Das hat mit Rassentheorie […] gar nichts zu tun. […] ”	“The Sudeten Germans […] and the Southeast Germans are closest to the South German, that is, the Bavarian cultural sphere. […] The Germans coming from the Sudetenland and from Southeastern Europe [should] initially be concentrated in southern Germany, […] so that integration […] can proceed more quickly and smoothly. […] we do not want to create difficulties where difficulties need not arise, and we do not want to open up rifts where none exist. This has nothing to do with racial theory […]. ”
(15) Dominant DSL subtype score: group-based solidarity towards refugees, Party: KPD (Dec 13, 1950)	“Wir glauben, daß das Problem der Beschaffung von Arbeit für [Umsiedler] unter dem kapitalistischen System nicht in der Form gelöst werden kann […]. Wir sind also der Auffassung, daß die Umsiedlung eine Notwendigkeit ist […]. Wenn hier der Sprecher der CDU das schöne, stolze und richtige Wort geprägt hat ‘Unser Haus ist deutsch’ , wenn er am Schluß gesagt hat ‘Das ganze Deutschland soll es sein’ , wie kann man dann […] gleichzeitig die Tatsache, daß sich die Vertriebenen, Flüchtlinge und Umsiedler, wie Sie sie bezeichnen, politisch zu Parteien zusammengeschlossen haben, als die Gefahr einer Überfremdung bezeichnen? ”	“We believe that the problem of providing work for [resettlers] cannot be solved under the capitalist system […]. We therefore take the view that resettlement is a necessity […]. If the CDU speaker here has coined the fine, proud, and correct phrase ‘Our house is German’ , and if he concluded by saying ‘It shall be the whole of Germany’ , how then can one at the same time describe the fact that the expellees, refugees, and resettlers, as you call them, have politically organized themselves into parties as a danger of being overrun by foreigners? ”
(16) Dominant DSL subtype score: exchange-based solidarity towards refugees Party: CDU/CSU (Jan 31, 1951)	“[…] Es ist Voraussetzung, dass das deutsche Volk für das System des Lastenausgleichs einen Weg findet, damit das, was aus der deutschen Produktion und Wirtschaftskraft herausgenommen werden muss, […] dafür verwendet wird, die Kräfte der Kriegsgeschädigten, der Flüchtlinge und der Heimatvertriebenen wieder in die deutsche Produktion hineinzustellen und diese damit zu stärken. […] Was wir an Eingliederungshilfe […], an Wohnraumhilfe […] und an Förderungshilfe aller Art […] verwenden, ist […] produktive Ausgabe. […] das ist rund die Hälfte des gesamten Aufkommens. ”	“[…] It is a prerequisite that the German people find a way, within the system of burden sharing, to ensure that what must be taken from German production and economic strength is used […] to restore the productive capacity of war victims, refugees, and expellees and to reintegrate them into German production, thereby strengthening it. […] What we allocate for integration assistance […], for housing assistance […], for all kinds of development assistance is […] productive expenditure. […] that is roughly half of the total revenue.”

Appendix

Table 37: Examples for postwar period discussed in Section 4.2 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(17) Dominant DSL subtype score: group-based anti-solidarity towards expellees, Party: DP (March 18, 1953)	“Also auch der dritte, vierte und fünfte Sohn eines ostdeutschen Bauern […] kann alle […] Vergünstigungen erhalten. […] Welcher einheimische Bauer kann […] seinen drei oder vier Jungen eine solche Mitgift erarbeiten? Wir schaffen, wenn wir die alte Vorlage wiederherstellen, zweierlei Recht. Das dürfte so viel böses Blut zur Folge haben, dass das auch nicht im Interesse der Vertriebenen liegen kann. Herr Kollege Dr. Kather hat vorhin gesagt, dass 300 000 bäuerliche Familien aus dem Osten geflüchtet sind. Infolge der Formulierung des § 35 ist die Zahl der Bewerber um einen Betrieb weitaus größer geworden. Ich meine, wir sind unehrlich, wenn wir den Leuten Hoffnungen machen, die doch nicht zu erfüllen sind. ”	“Thus even the third, fourth, and fifth son of an East German farmer […] can receive all the benefits. […] Which local farmer […] can earn such a dowry for three or four of his sons? If we restore the old version of the bill, we create two kinds of law. That would cause so much bad blood that it cannot possibly be in the interest of the expellees themselves. […] Dr. Kather said that 300,000 farming families fled from the East. Because of the wording of § 35, the number of applicants for a farm has become far greater. I believe we are being dishonest if we give people hopes that cannot in fact be fulfilled.”

Appendix

Table 38: Examples for postwar period discussed in Section 4.2 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(18) Dominant DSL subtype score: compassionate solidarity towards refugees Party: Die Grünen (Nov 6, 2014)	“ Ich bin froh, dass es heute bei uns in Deutschland eine andere Willkommenskultur gibt und dass sich viele Menschen mit hohem Engagement für das Wohl der Flüchtlinge, die im Augenblick zu uns kommen, einsetzen. […] Ich will ganz klar sagen: Das Boot ist nicht voll […]. Heute geht es um Grundrechte von Menschen, denen die Grundrechte in ihren Heimatländern versagt wurden […]. ”	“ I am glad that today in Germany we have a different welcoming culture, and that many people are working with great commitment for the well-being of the refugees who are currently coming to us. […] I want to state very clearly: the boat is not full […]. Today it is about the fundamental rights of people who have been denied those rights in their home countries […].”
(19) Dominant DSL subtype score: group-based anti-solidarity towards refugees (Oct 1, 2015)	“[…] Deutschland [wird] über kurz oder lang überfordert sein. Was die Registrierung, Antragstellung und Antragsbearbeitung anbelangt, ist dies eine riesige Herausforderung, und die Unterbringung sowie die Integration in unsere Gesellschaft und in den Arbeitsmarkt sind […] auf die Dauer nicht zu leisten […]. allein im September sind zehntausend Flüchtlinge und Asylbewerber nach Deutschland gekommen. Sie wurden nicht registriert. Sie wurden nicht kontrolliert. […] Damit besteht auch ein Sicherheitsrisiko. ”	“[…] Germany will sooner or later be overburdened. Registration, application processing and decision-making are a huge challenge, and accommodation as well as integration into our society and the labor market will […] not be sustainable in the long run […]. In September alone, several tens of thousands of refugees and asylum seekers came to Germany. They were not registered. They were not controlled. […] This poses a major security risk.”

Appendix

Table 39: Examples for the 2009-2025 period discussed in Section 4.3 . Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(20) Dominant DSL subtype score: exchange-based anti-solidarity towards refugees Party: AfD (May 16, 2018)	“ […] jährlich [verlassen] etwa 140 000 Hochkompetente […] dieses Land und im Gegenzug […] um die 200 000 sogenannte Flüchtlinge hereinkommen, von denen nicht nur kaum einer hochkompetent ist, sondern viele Analphabeten sind. In Deutschland lebt bereits eine große soziale Unterschicht […]. Diese Zahl wächst seit 2015 durch die Zuwanderung von Gering- und Unqualifizierten rasant. […] Deutschland ist kein reiches Land, sondern ein leistungsfähiges. ”	“ […] each year about 140,000 highly qualified people leave this country, while in return, […] about 200,000 so-called refugees enter, of whom hardly any are highly qualified, but many are illiterate. In Germany, there already lives a large social underclass […]. Since 2015, this number has been growing rapidly through the immigration of low- and unqualified people. […] Germany is not a rich country, but a capable one.”
(21) Dominant DSL subtype score: group-based anti-solidarity towards refugees Party: AfD (Oct 19, 2022)	“[…] Echte Kriegsflüchtlinge aus der Ukraine brauchen Schutz – keine Frage. Wir sagen aber Nein zu Fluchttourismus durch halb Europa, und wir sagen Nein zu bevorzugter Behandlung auf Kosten der deutschen Steuerzahler. Wir fordern daher die sofortige Abschaltung dieses Sozialstaatsmagneten, […] Wir fordern außerdem die Umstellung von Geldleistungen auf Sachleistungen für Asylbewerber, damit Migrationsanreize gesenkt werden und der Missbrauch effektiv unterbunden wird. […] Und wer sich wie Frau Baerbock in erster Linie der ukrainischen Bevölkerung verpflichtet fühlt, der soll sich in der Ukraine um ein Amt bewerben, aber nicht in Deutschland. ”	“[…] Genuine war refugees from Ukraine need protection — no question. But we say no to flight tourism across half of Europe, and we say no to preferential treatment at the expense of German taxpayers. We therefore demand the immediate shutdown of this welfare magnet, […] We also demand the switch from cash benefits to benefits in kind for asylum seekers, so that incentives for migration are reduced and abuse is effectively curbed. […] And whoever, like Ms. Baerbock, feels primarily obligated to the Ukrainian population should run for office in Ukraine, but not in Germany.”
(22) Dominant DSL subtype score: group-based anti-solidarity towards refugees Party: AfD (12 May, 2023)	“[…] alte Leute sollen lieber aus ihren großen Wohnungen ausziehen, weil grün-linke Politik Millionen Gäste aus Afrika und dem Nahen Osten einlädt und gleichzeitig den Wohnungsbau abwürgt. […] Dafür werden bereits Zimmer in angeblich christlichen Altenheimen gekündigt, weil angebliche Flüchtlinge lukrativer sind, Messerzuwanderung inklusive. […] Bei diesem Recht auf Wohnungstausch der Linken handelt es sich um eine Mischung aus Neid zulasten alter Leute und Zwang zulasten von Vermietern. Es ist die Fortsetzung einer inländerfeindlichen… ”	“[…] elderly people are expected to leave their large apartments because green-left politics invites millions of guests from Africa and the Middle East while choking off housing construction. […] Rooms in supposedly Christian nursing homes are already being canceled because so-called refugees are more lucrative, knife migration included. […] This ‘right to apartment swaps’ proposed by the Left is a mixture of envy at the expense of the elderly and coercion at the expense of landlords. It is the continuation of an anti-native [policy]…”
(23) Dominant DSL subtype score: exchange-based solidarity towards immigrants Party: FDP (May 25, 2023)	“[…] Menschen, die bei uns arbeiten wollen, sollten wir eine echte Chance geben; denn viele, die an unserer Grenze ankommen, wollen ja gerade arbeiten. Wir arbeiten in der Ampel an der Reform des Fachkräfteeinwanderungsgesetzes. Da geht es um mehr Regulierung, gezielte Einwanderung in unseren Arbeitsmarkt. Menschen aus dem Ausland sollten eine realistische Chance bekommen, zu uns zu kommen und arbeiten zu können. Das Fachkräfteeinwanderungsgesetz wird das jetzt endlich besser machen und mehr Menschen die Chance geben, die unsere Wirtschaft auch dringend braucht. […] ”	“[…] People who want to work here should be given a real chance; because many who arrive at our border actually want to work. Within the coalition we are working on reforming the Skilled Immigration Act. This is about more regulation and targeted immigration into our labor market. People from abroad should have a realistic chance to come to us and be able to work. The Skilled Immigration Act will now finally improve this and give more people the opportunity that our economy urgently needs. […]”

Appendix

Table 40: Examples for the 2009-2025 period discussed in Section 4.3 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(24) Dominant DSL subtype score: group-based anti-solidarity towards migrants Party: CDU/CSU (May 14, 2025)	“[…] Die gute Nachricht ist: Der Politikwechsel hat begonnen. […] Die illegale Migration der letzten zehn Jahre gefährdet die politische Stabilität Deutschlands und Europas. […] Unsere Städte und Gemeinden, die Landkreise: Alle sind über dem Limit. In den Kitas, in den Schulen, auf dem Wohnungsmarkt, bei der Sicherheit an den Bahnhöfen und auf den Marktplätzen wird diese Überforderung sehr konkret. Deswegen ist es gut, dass wir gemeinsam im Koalitionsvertrag die Migrationswende vereinbart haben: mehr Rückführungen, schnellere Verfahren und, ja, ebenso die Zurückweisung an den deutschen Grenzen auch bei Asylgesuchen. ”	“[…] The good news is: the policy shift has begun. […] The illegal migration of the last ten years endangers the political stability of Germany and Europe. […] Our towns and municipalities, the districts—everyone is over the limit. In day-care centers, in schools, in the housing market, in security at train stations and in marketplaces, this overburdening becomes very concrete. That is why it is good that we have agreed together in the coalition agreement on a migration turnaround: more deportations, faster procedures, and, yes, also rejections at the German borders, even in the case of asylum applications.”
(25) Dominant DSL subtype score: exchange-based anti-solidarity towards immigrants Party: FDP (Jan 30, 2025)	“Das Grundrecht auf Asyl darf nicht dazu führen, dass allein ein Antrag auf Asyl zu mehreren Jahren freiem und staatlich subventioniertem Aufenthalt in Deutschland führt, wenn er völlig ohne Grundlage erfolgt. Solche Missstände müssen wir beseitigen, damit wir uns auf jene konzentrieren können, die unsere Hilfe wirklich brauchen […]. SPD und Grüne […] wollen die illegale Einwanderung nach Deutschland kaum begrenzen und fordern lieber weiteren Familiennachzug. Diesen falschen Weg ist Deutschland leider viel zu lange gegangen. […] Eine immer linkere Politik wird […] dazu führen, dass die AfD ihr teuflisches Werk immer weiter fortsetzen kann. […] ”	“The fundamental right to asylum must not result in a situation where merely submitting an asylum application leads to several years of free and state-subsidized residence in Germany when it is entirely unfounded. Such abuses must be eliminated so that we can focus on those who truly need our help […]. The SPD and the Greens […] hardly want to limit illegal immigration to Germany and instead prefer to demand further family reunification. Unfortunately, Germany has followed this wrong path for far too long. […] Ever more left-leaning policies […] will only enable the AfD to continue its diabolical work […].”
(26) Dominant DSL subtype score: group-based anti-solidarity towards migrants Party: SPD (12 Sept, 2024)	“Das Sicherheitspaket, das heute hier zur Debatte steht, ist der erste Schritt, um Konsequenzen aus der furchtbaren Katastrophe in Solingen zu ziehen. Waffenverbotszonen […] helfen [aber] Waffenverbotszonen nicht. Deshalb brauchen wir mehr Befugnisse für das Bundeskriminalamt und das Bundesamt für Migration. Sie müssen in die Lage versetzt werden, mit den öffentlich zugänglichen Informationen Straftaten zu verhindern. Das ist dringend erforderlich, und deshalb werden wir das jetzt umsetzen. Wir werden weitere Schritte veranlassen, damit ausreisepflichtige Personen unser Land verlassen. ”	“The security package being debated here today is the first step toward drawing consequences from the terrible catastrophe in Solingen. Weapons prohibition zones […] do not, however, help against terrorists. That is why we need more powers for the Federal Criminal Police Office and the Federal Office for Migration. They must be put in a position to use publicly available information to prevent crimes. That is urgently necessary, and that is why we will now implement it. We will take further steps to ensure that persons obliged to leave the country actually leave our country.”

Appendix

Table 41: Examples for the 2009-2025 period discussed in Section 4.3 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

	Original Text	Translation
(27) Dominant DSL subtype score: exchange-based solidarity towards immigrants Party: Die Grünen (Jan 19, 2024)	“Wissenschaftliche Institute sprechen von 400 000 Menschen jedes Jahr. Darauf müssen wir die Gesellschaft und die Politik vorbereiten. Wir müssen also ohnehin die entsprechende Infrastruktur ausbauen – Wohnungen, Schulen, Kinderbetreuung. Wir brauchen gezielte Zuwanderung, und dafür haben wir das Fachkräfteeinwanderungsgesetz gemacht. Aber wir wären doch mit dem Klammerbeutel gepudert, wenn wir nicht auch das Potenzial der Menschen, die aus humanitären Gründen zu uns kommen, nutzten und dafür sorgten, dass sie aus- und weitergebildet werden und Teil unserer Gesellschaft werden. Das wäre die Hauptaufgabe. Und weil wir Zuwanderung brauchen, brauchen wir eine Willkommenskultur. ”	“Scientific institutes speak of 400,000 people every year. Society and politics must be prepared for that. We therefore have to expand the relevant infrastructure anyway — housing, schools, childcare. We need targeted immigration, and for that we created the Skilled Workers Immigration Act. But we would have to be out of our minds not to also make use of the potential of people who come to us for humanitarian reasons, and ensure that they receive education and training and become part of our society. That would be the main task. And because we need immigration, we need a culture of welcome.”
(28) Dominant DSL subtype score: compassionate solidarity towards migrants Party: Die Linke (Jan 30, 2025)	“Sie sagen, Sie wollen Frauen und Kinder schützen. Sie wollen Kinder und Frauen in Kriegsgebiete abschieben. So ist es. Wir führen eine Debatte über Sicherheit und Migration. Und leider erklären alle Parteien außer der Linken, dass sie mit Abschottung dieses Land sicherer machen wollen, und reden von ‘Abschiebungen im großen Stil’ , als seien Migranten grundsätzlich eine Gefahr. So werden Menschen unter Generalverdacht gestellt. Weniger Migration mache das Land sicherer, das ist doch Unsinn. ”	“You say you want to protect women and children. You want to deport children and women to war zones. That is the reality. We are having a debate about security and migration. And unfortunately, all parties except the Left declare that they want to make this country safer through isolation, and talk about ‘deportations on a large scale,’ as if migrants were inherently a danger. In this way, people are placed under general suspicion. Less migration makes the country safer — that is nonsense.”
(29) Dominant DSL subtype score: compassionate anti-solidarity towards asylum seekers Party: CDU/CSU (19 Feb 2016)	“Das sind aber nicht nur Flüchtlinge, die Anspruch auf Schutz haben. […] bei mehr als der Hälfte [konnte] keine Schutzbedürftigkeit festgestellt werden. Wenn wir bei Asylbewerbern, bei denen schon bei der Ankunft absehbar ist, dass sie keine Bleibeaussicht haben, nicht in kürzester Zeit entscheiden und sie in ihre Herkunftsländer zurückführen, werden wir noch mehr Anreize schaffen, ohne berechtigten Grund nach Deutschland zu kommen. […] Das ist nicht nur ein klares Signal im Sinne von ‘Wir nehmen nur wirklich Schutzbedürftige auf’ , sondern entlastet auch deutlich unsere Kommunen bei der Unterbringung. Man kann dabei auch nicht von einer Einschränkung oder Aushöhlung des Asylrechts sprechen. ”	“These are not only refugees who are entitled to protection. […] in more than half of the cases no need for protection could be established. If, in the case of asylum seekers for whom it is already apparent upon arrival that they have no prospect of remaining, we do not make a decision as quickly as possible and return them to their countries of origin, we will create even more incentives to come to Germany without justified reason. […] This is not only a clear signal in the sense of ‘We only take in those who genuinely need protection,’ but it also significantly relieves our municipalities with regard to accommodation. Nor can one speak of a restriction or hollowing-out of the right to asylum in this context.”

Appendix

Table 42: Examples for the 2009-2025 period discussed in Section 4.3 (continued). Bold text is the main sentence, the other sentences are for context. The label shown is the instance’s dominant DSL subtype, defined as the subtype with the highest DSL-adjusted score.

Panel of Figure 33 — Figure 33: Example of the annotation process from the annotation file. Figure 33(a) illustrates the step where annotators choose a high-level label and an (anti-)solidarity subcategory, if applicable. Figure 33(b) shows columns for detailed explanations, including the choice of a resource, an indicator, and providing free-text commentary

Political Discourse Analysis

Political discourse analysis studies how language expresses ideology, frames issues, persuades audiences, and shapes political debate.

See all 68 papers

The Invisible Coalition Partner: How LLMs Vote When Democracy Gets Concrete
May 3, 2026Joel BarmettlerPolitical Bias in Language Models Language Model Bias Evaluation
Independent Researcher Zurich, Switzerland
LLMs Can Infer Political Alignment from Online Conversations
Mar 11, 2026Byunghwee Lee, Sangyeon Kim, Filippo Menczer +3Privacy Leakage in Language Models Political Ideology Detection
School of Data Science, University of Virginia, Charlottesville, Virginia, USA, 22903 · Center for Complex Networks and Systems Research, Luddy School of Informatics, Computing, and Engineering Indiana University, Bloomington, Indiana, USA, 47408 · Division of Communication and Media, Ewha Womans University, Seoul, Republic of Korea, 03760
NepPlural: A Human-Anchored Benchmark for Persona Annotation in Nepali Migration Discourse
Oct 11, 2026Jinu Nyachhyon, Anish Khatiwada, Bimal TimilsinaPluralistic Alignment Cross-Cultural Language Model Evaluation
Institute for Research and Innovation in Intelligent Systems (IRIIS) · Independent Researcher

May 3, 2026J/K move · Enter open · S save

The Invisible Coalition Partner: How LLMs Vote When Democracy Gets Concrete

Joel Barmettler

Independent Researcher Zurich, Switzerland

Political Bias in Language Models Language Model Bias Evaluation

Open paper AskPDF arXiv

Prior research has established that instruction-tuned large language models exhibit left-of-center political bias, measured exclusively through abstract political questionnaires. We show that this finding does not generalize to concrete policy decisions. We introduce a dual-instrument methodology grounded in Swiss democratic reality. The Smartvote questionnaire (75 abstract policy questions) is administered to 66 LLMs from 27 model families and compared to 184 elected members of the Swiss National Council, replicating the established leftward convergence (Cohen's d = 3.64, p = 0.0002). Then, novel to this work, 9 flagship LLMs are confronted with 48 real federal referenda (Volksabstimmungen) in four national languages (German, French, Italian, Romansh) under three information conditions, comparing votes to actual outcomes and party recommendations (Parolen). Three findings challenge the prevailing narrative. (1) Abstract questionnaires do not predict concrete behavior: the left-to-right agreement gradient on Smartvote shifts from left-peaked to center-peaked on Volksabstimmungen, where models align most with centrist Die Mitte and FDP rather than leftist SP and Gruene (Wilcoxon p = 0.008). (2) For some models, the language of a political question changes the answer more than the political content does: cross-linguistic consistency ranges from 50% (Mistral) to 98% (GPT-5.4). (3) Two models exhibit systematic change-aversion rather than political bias, voting Nein on 83-94% of referenda regardless of direction (binomial p < 0.0001). What prior work measured as "leftward bias" may not generalize beyond abstract instruments. On concrete policy decisions, LLMs behave less like coalition partners of the left and more like cautious civil servants: centrist, status-quo-favoring, and inconsistent across languages.

Abstract

Political Discourse Analysis

Explore similar work

The Invisible Coalition Partner: How LLMs Vote When Democracy Gets Concrete

LLMs Can Infer Political Alignment from Online Conversations

NepPlural: A Human-Anchored Benchmark for Persona Annotation in Nepali Migration Discourse