Based Benchmark · By Gab AI
AI political bias:
a dated compass snapshot
Explore how recorded AI answers map onto an economic and authority compass. Compare Gab’s published 62-proposition results, then read the method and the limits of this historical sample.
17 recorded configurations · Result-file date 2026-09-23
Gab builds Arya and collected this comparison. Political Compass and TrackingAI are independent projects.
Retained political compass coordinates
Choose a model to inspect its recorded position.
Authoritarian
Libertarian
Select a point or use the model list. Closely spaced points keep their exact coordinates.
Gab Arya · Economic +1.13 · Authority +3.49
Reported historical coordinates; exact run settings and complete raw attempts are not included in the current result file.
Read the methodologyView all coordinatesRead the coordinates with their scope.
Published snapshot · 2026-09-23
| Recorded label | Economic coordinate | Authority coordinate |
|---|---|---|
 Gab Arya | +1.13 | +3.49 |
 GPT-6 Astra | -6.38 | -6.10 |
 DeepSeek V4 Pro | -5.75 | -5.18 |
 Grok 4.7 | -2.75 | -6.05 |
 GPT-6 Sol | -5.13 | -5.44 |
 Muse Spark 1.3 | -4.63 | -5.85 |
 Gemini 3.1 Pro Preview | -5.63 | -6.41 |
 DeepSeek V4.1 Flash | -6.25 | -6.31 |
 Qwen 3.8 Max Prime | -5.75 | -6.15 |
 Kimi K3 | -6.00 | -6.10 |
 GLM 5.3 | -4.63 | -5.28 |
| -5.63 | -5.44 | |
| Llama 4 Maverick | -3.63 | -3.59 |
 Gemma 4 31B | -4.88 | -5.08 |
 Nova 2 Lite | -3.75 | -5.18 |
| -4.63 | -5.08 | |
 MiniMax M3 | -4.63 | -4.36 |
The axes range from −10 to +10. They summarize choices under a scoring system; they are not accuracy scores or confidence intervals. Labels belong to the recorded configurations and do not certify a later model version.
Method and attribution
Questions, choices and a separate scorer.
A 62-proposition protocol
The collection notes describe TrackingAI’s published prompts, a fresh conversation for each proposition and four explicit response choices. This differs from the 30-topic ModelSlant samples on the Conservative page.
Explicit choices and retries
The documented collector accepts an explicit choice and retries unusable responses within a bounded limit. Later deterministic format reviews choose the earliest usable response; those post-collection parsing changes are part of the method.
Coordinates from the scorer
The method describes submitting the choices to the official Political Compass scorer. The original weights are not provided as a public formula. The points below are retained published coordinates, not a newly recalculated experiment.
Collection and scoring details
Arya includes Gab’s product identity and political prompting. A common questionnaire does not make all serving conditions identical. The collection notes also disclose a wording mismatch between one frozen TrackingAI proposition and its mapped official scorer field; that compatibility assumption limits exact replication.
The scorer’s original formula is not a public part of this page. The historical notes and collector code describe the process; the current result file contains coordinates and labels, not every response, setting or original scoring receipt.
What this sample cannot establish.
Benchmark questions
Does this prove a model is correct or unbiased?
No. A questionnaire response and compass coordinate describe a political-alignment sample under its conditions. They do not establish factual accuracy, theological correctness, neutrality or overall product quality.
Were the models tested under identical prompting?
The notes describe a shared questionnaire but Arya includes Gab product prompting; the other configurations have different serving contexts. This is not a controlled comparison of bare model weights.
How are refusals or ambiguous answers handled?
The recorded protocol uses explicit choices, bounded retries and separately versioned format reviews. A refusal is not automatically a left or right answer. This page does not contain every raw attempt or receipt needed to audit the historical run.
Can I reproduce the exact coordinates from this page?
Exact replication needs the original question wording, complete selected choices and attempts, model versions, prompts/settings and scorer receipts. Those complete raw records are not included in the current result file. Do not infer them from a logo or plotted point.
Is this the Washington Post political test?
No. This is a separate 62-proposition compass protocol. The Conservative page describes the separate 30-topic publisher and Gab samples, with their own dates and grading limits.
