Amtro AI Map · Baseline July 2026

Same model,
different worldview?

Name: Amtro AI Map Baseline July 2026
Creator: Amtro Research
Published: 2026-07-24

Why political labels for language models depend on the questionnaire used.

Preliminary baseline · Methods series · Limitations stated explicitly

Eight scorable models, three instruments. Status: preliminary.

Audit package (ZIP, 20 KB) Citation and limits

Preliminary baseline · PCT and WVS side by side. Seven of eight models switch the economic sign.

Amtro signal

Seven of eight models switch economic quadrants between two instruments.

On the Political Compass Test, all eight scorable models landed economically left and socially libertarian. On the Amtro WVS battery, seven of them crossed the economic zero line. A political label from a single test can shift the direction of the classification.

8scorable models

3survey instruments

7/8quadrant switches

2026-07baseline

Political Compass Test (PCT) and WVS-derived battery, July 2026 snapshot. Eight models scored on both instruments. Seven change economic sign; none change social sign. Scales are differently constructed — compare direction/sign, not absolute position. Status: provisional. Moonshot Kimi PCT: 1 of 3 planned runs. Planned runs otherwise: 3, temperature 0.

Protocol outputs, not beliefs or personality; no provider, country or training effect.

Data as table (text alternative)

PCT and WVS coordinates for eight models, July 2026
ID	model_id	PCT eco	PCT soc	WVS eco	WVS soc	eco switch	PCT n
A	anthropic/claude-sonnet-4.6	−8.4533	−8.4105	+1.0417	−2.4583	yes	3
B	deepseek/deepseek-v3.2	−8.0783	−7.4874	+1.0764	−2.1042	yes	3
C	z-ai/glm-5	−4.3700	−5.3849	+0.8472	−2.6667	yes	3
D	google/gemini-3.5-flash	−6.6200	−7.4874	+0.8264	−3.3194	yes	3
E	moonshotai/kimi-k2.5	−5.7450	−6.8208	+0.4514	−3.6250	yes	1
F	mistralai/mistral-large-2512	−8.7450	−7.8464	+1.4792	−2.5972	yes	3
G	openai/gpt-5.5	−6.5367	−7.7438	−0.1736	−4.3264	no	3
H	qwen/qwen3.6-plus	−7.3700	−7.8464	+0.2292	−3.8333	yes	3

Totals: economic sign changes 7/8; social 0/8. Status: provisional.

Observation

A compass score is not a stable model property.

Instrument, prompt, language, response format and model version all affect the measured position. Political model cards should disclose these conditions instead of treating a single point as a fixed orientation.

Next step

The baseline becomes a time series.

The series is set up for monthly repetition. Future runs will track model changes and drift; quarterly robustness checks will vary question order, language and open-ended response conditions.

Citation

Observation, inference and hypothesis stay separate.

Amtro Research (2026). Same model, different worldview? Instrument sensitivity of political LLM evaluations. Amtro AI Map, baseline July 2026, preliminary version.

Download audit package (ZIP, 20 KB) Curated canonical baseline files (claim matrix, quality report, aggregate, protocol, lab and dual figures) with SHA-256 — not a full raw-data archive. Browse individual files

Updates

Hear about the next measurement.

The series is set up for monthly repetition. We only write when new results are published. No marketing; reply once to unsubscribe.