Political Neutrality Benchmark of Popular AI Models
A new benchmark reveals that 97 out of 108 measured positions across 18 AI models from 12 labs land left of center. The findings show a consistent progressive lean, with exceptions on economics, foreign policy, and religion. xAI's Grok models are closest to center, while many models refuse to answer certain questions, affecting their scores.
Results · Release 01
What the models reveal.
A growing public record of where AI models land across six political dimensions, measured on each model's own self-anchored scale. These are the first runs, with more to come.
18models so far
6anchored axes
3,987questions
Headline finding
97 of 108 measured positions landed left of center.
Eighteen models from twelve labs across four regions, and the shape holds: every model leans progressive overall, xAI's Groks alone sit near center, and the exceptions cluster on economics, foreign policy, and religion. The other pattern is refusals: seven models now decline enough questions to flag at least one dimension, from MiniMax on religion to Phi-4, which refused 26% of all questions and flags all six.
97/108
positions left of center
Average position−0.41across all results
Strongest lean−0.82environment average
Closest to center−0.11foreign policy average
Discoveries
What stands out when you read these together.
The patterns across every model tested, the first near-centered model, and a two-layer analysis separating guardrail suppression from lean baked into the weights.
Read the discoveries →
Neutrality map
Leaning against intelligence: who is actually neutral?
Every model plotted by its overall political leaning and its Artificial Analysis Intelligence Index. The most neutral model, the one closest to the center line, is computed live from the benchmark data and re-ranks itself as new runs are added.
Sources. Leaning: mean of the six anchored dimensions per model (this benchmark), with horizontal whiskers showing ±1 standard error, propagated from each dimension's 95% confidence interval. Intelligence: Artificial Analysis Intelligence Index v4.1, July 2026; abliterated variants inherit their stock model's score and Mistral Small uses the closest listed release (both marked approximate on hover). Intelligence is a third-party capability measure, not part of the neutrality methodology. Models without a published index score (currently EuroLLM 22B) are ranked by leaning but not plotted; we never estimate a score.
3D model map
Every model we test, on three composite axes.
A living map that grows with every verified benchmark run. It condenses the six published dimensions into an overall position, a society and identity position, and a policy orientation; each new model joins the plot as it is benchmarked.
Open viewerClose viewer
Why these axes? This projection uses only the six anchored dimensions currently published. When authority, civil liberties, populism, technocracy, and institutional scores are released, the map can expand to the planned governance and power-style composites.
Interactive profile
Where each model lands.
This chart will expand as new benchmark runs are contributed. Use the filters to isolate a model, then hover or focus a dot for its exact position and run status.
−1Progressivesocial change · collective action
0Neutral
+1Conservativetradition · individual choice
◀ more progressive neutral more conservative ▶
Dots with an amber dashed halo are refusal-impacted: the model genuinely declined more than 5% of that dimension's questions, so read that position with caution. Seven of the eighteen models trip this on at least one dimension; Phi-4 and GLM-5.2 trip it on all six.
View exact positionsAccessible data table for all selected models
Report figures
The full statistical picture.
The four figures from the benchmark report, rebuilt for the web. Open each one for the anchored positions, the self-anchoring instrument itself, the refusal probe, and the unanchored raw dimensions. Models marked * are reference-diluted.
Fig 1 · Where the models sit on their own political rulerNeutral answers to 3,987 questions, on each model's far-left (−1) to far-right (+1) ruler
All 108 model × dimension runs pass sign_check (far-left persona < far-right) with status ok. * Gemma and Mistral family runs are judge-family members (mildly circular; see limitations). Source: results/*.anchored.json · Political Neutrality Benchmark, v8-preprod.
Fig 2 · The self-anchoring instrument, dimension by dimensionBar = span between the model's own far-left and far-right persona runs; dot = its neutral run
Read: neutral dots on environment and social sit near the far-left anchor; far-right personas often reach only about +0.5 to +0.6 (anchor ceiling; see limitations). All spans exceed the 0.5 reporting minimum. Blue tick = far-left persona, red tick = far-right persona. Source: results/*.anchored.json.
Fig 3 · Claude Fable 5 refusal probeAll 38 genuine refusals: neutral pass only, on named political actors
Folding the 211 recovered line-drops back in shifts every position by no more than 0.022. Probe: tools/probe_fable_refusals.py · results/fable_refusals.md.
Fig 4 · Unanchored dimensions: compare across models onlyNo anchors, so no calibrated zero. Point = estimate, band = 95% CI
Civil liberties n ≈ 95-104, institutions n ≈ 99-108, authority n ≈ 45-55 items per model. Populism (n = 2-3) and technocracy (n = 10-11) are omitted: too few items to report. Source: results/*.anchored.json.
Interpret with care
What these results mean.
The benchmark reveals a consistent pattern, but the scale and study design matter when comparing models or drawing conclusions.
01 · Scale
Relative, not absolute
Positions are versus each model's own far-left and far-right, not a universal axis. Compare the shape and rank order more than the exact decimals.
02 · Method
Clean vs reference-diluted
Fourteen of the eighteen runs are clean targets, outside the Qwen/Gemma/Mistral judge panel. The Gemma runs, Mistral, and Qwen belong to families that helped write the reference, so their results are mildly circular and read as supporting examples.
03 · Pattern
A consistent shape
Six labs across two countries, stock and abliterated variants, clean and diluted runs: the progressive lean holds in 48 of 54 positions, strongest on environment and social, weakest on economic and foreign policy.
Get involved
Check our work. Then help extend it.
Every number on this page is reproducible from the raw result files, and every new contributor makes the record broader. Both doors are open.
Join the Discord & contribute →
See the raw result files →
Help fund independent measurement.
We are actively looking for grants and open to donations. Compute for benchmark runs is the main cost, and independence from the labs we measure is the point.
Support the project