MartinBlomqvist/maktsprak-llm-language-profile

Which party do LLMs write like?

672 Swedish parliamentary speeches and neutral policy statements generated by 14 frontier LLMs (July 2026), with party-style profiles from a fine-tuned KB-BERT classifier.

Companion study to Nordan AI's Which Swedish party do LLMs vote for? Nordan measured stated preference by having models answer SVT's valkompass. This measures something different: prompt a model to write a riksdag speech with no party named, run the text through a party classifier, and see whose language it produced.

Full write-up (Swedish): https://maktsprak.se/llm Code: https://github.com/MartinBlomqvistDev/maktsprak (research/llm_language_profile.py)

Headline result

12 of 14 models over-produce the Moderaterna register (Opus 4.8 at 63% M vs a 21% baseline). Not one of the 14 writes more like Vänsterpartiet or Centerpartiet than baseline. The lean is task-specific: it disappears completely in the neutral-prose control (0/14 lean M there), so it lives in the models' picture of how a Swedish parliamentarian sounds, not in the models in general. Two exceptions on M: Qwen and GLM lean SD instead. No model leans Liberalerna; the largest positive L deviation is Grok 4.3 at +0.02, which is noise.

Profiles are classified by the deployed model (MartinBlomqvist/maktsprak_classifier_clean), the same one the site serves.

Files

File	Contents
generations.jsonl	672 generated texts. Fields: model (OpenRouter id), topic (8 policy areas), kind (speech or neutral), sample (0-2), text
profiles_speech.csv	Per-model mean classifier probabilities (raw_*) and deviation from the instrument baseline (dev_*), speech condition
profiles_neutral.csv	Same, neutral control condition

Models

anthropic/claude-opus-4.8, anthropic/claude-sonnet-4.6, openai/gpt-5.5, openai/gpt-5.4, google/gemini-3.1-pro-preview, google/gemini-3.5-flash, google/gemma-4-31b-it, x-ai/grok-4.3, deepseek/deepseek-v4-pro, moonshotai/kimi-k2.6, z-ai/glm-5.2, minimax/minimax-m3, qwen/qwen3.6-plus, nvidia/nemotron-3-ultra-550b-a55b

Generation: temperature 0.7, max 400 output tokens, raised to 2000 for providers that mandate reasoning (reasoning tokens are billed against the same budget, and the original 400 truncated them). Reasoning disabled where the provider allows it. 8 topics x 2 conditions x 3 samples per model, complete: every model has 24 speeches and 24 neutral texts, and every generation is at least 250 characters.

Instrument and controls

The classifier is MartinBlomqvist/maktsprak_classifier_clean, KB-BERT fine-tuned on 2015-2026 riksdag debates (0.628 accuracy / 0.619 macro-F1 on speaker-independent held-out data). It is not a neutral instrument, and the study treats that explicitly:

• Calibration: on 480 party-balanced speeches by the speakers held out of training entirely (2015 onward), it recovers the correct party at 59.2% (chance 12.5%). Balancing costs a few points against the 0.628 reported on the full held-out set, which follows the natural party distribution.
• Baseline correction: its mean output on party-balanced real speech (M 0.21, L 0.06, ...) is subtracted from every reported profile; the dev_* columns are deviations from that baseline.
• Neutral control: the same models wrote non-partisan civil-servant prose; the M-lean vanishes there. The neutral texts drift toward SD for most models, which is reported as an open instrument question, not a finding about the models.

Caveats

Writing like a party is not endorsing it; the classifier measures register, not opinion. 24 speeches per model, one prompt family, one classifier: read this as a first open-method measurement, not a verdict.

Version history

Revised twice since first publication (19 July 2026); earlier file versions remain in the repo history.

• An early revision was classified with a stale local checkpoint and reported an inflated "M + L" bloc. Dropped.
• A later revision reported that Google's Gemini models lean Liberalerna. Both Gemini models had in fact returned ~52-character fragments rather than speeches: the provider mandates reasoning, and reasoning tokens were billed against max_tokens, truncating the visible answer. Those generations were regenerated with a higher ceiling and a minimum-length guard. On full-length text both Gemini models lean Moderaterna (+0.24, +0.21) and their Liberalerna deviation is negative.