Datasetpublic

trakkr-ai/political-bias-in-ai

Political Bias in AI — Where the Major AI Models Stand An open, monthly measurement of where the major AI models land on value‑loaded political and ethical questions. Each model is asked the same battery of questions many times, with web search turned off, so the result reflects the trained weights rather than whatever the model retrieves that day. Every answer is classified by a neutral coder onto a left–right economic axis and a libertarian–authoritarian social axis, and… See the full description on the dataset page: https://huggingface.co/datasets/trakkr-ai/political-bias-in-ai.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes672downloads

Dataset Card

Political Bias in AI — Where the Major AI Models Stand

An open, monthly measurement of where the major AI models land on value‑loaded political and ethical questions. Each model is asked the same battery of questions many times, with web search turned off, so the result reflects the trained weights rather than whatever the model retrieves that day. Every answer is classified by a neutral coder onto a left–right economic axis and a libertarian–authoritarian social axis, and reported with run‑to‑run error bars and the raw text behind every point.

This dataset is the data behind [trakkr.ai/bias](https://trakkr.ai/bias). It is published openly under CC BY 4.0 — use it, cite it, build on it.

Descriptive, not normative. This dataset measures where models stand; it does not argue where they should stand. "Left" and "right" are coordinates, not verdicts. Nothing here implies one position is correct.

This release: June 2026


Models	6 — ChatGPT, Claude, Gemini, Grok, Llama, DeepSeek
Questions	61 value‑loaded items across economic, social, civil‑liberties, foreign‑policy and speech/tech axes
Runs	12 per model per question
Answers	4,392 raw model responses
Condition	A — raw weights, no web search, default temperature
Classifier	a neutral LLM coder (deepseek‑v4‑flash) maps each answer to a position + refusal type
Generated	2026‑06‑15

Where each model landed (June 2026)

Economic axis runs left (−1) to right (+1); social axis runs libertarian (−1) to authoritarian (+1). These are measured averages over 12 runs, not opinions.

Model	Version	Economic	Social	Read
ChatGPT	gpt‑5.5	−0.29	−0.58	Leans left
Grok	grok‑4.3	+0.21	−0.30	Leans right
Claude	claude‑opus‑4‑8	−0.06	−0.12	Center
Llama	Llama‑4‑Maverick‑17B	−0.06	−0.09	Center
DeepSeek	deepseek‑chat	−0.03	−0.04	Center
Gemini	gemini‑3.5‑flash	0.00	0.00	Center

A few measured patterns this month: ChatGPT sat furthest to the economic left; Grok furthest to the right and was both the most variable run‑to‑run (stability 57%) and the most steerable under persona pressure; Gemini was the most consistent (98%). Refusal rates were low across the board. See the live, interactive version — per model, per question, with the receipts — at trakkr.ai/bias.

Files

File	What it is
`aggregates-2026-06.json`	The full monthly aggregate: every model's coordinates, four‑axis "character" (lean, stability, steerability, candor), per‑subtopic positions, moral‑foundation fingerprint, closest real‑world political anchor, the 61‑question index, and field‑level superlatives.
`raw-answers-2026-06.jsonl`	One row per model response (4,392 rows). The receipts behind every point — including the full answer text.

`raw-answers-2026-06.jsonl` schema

Newline‑delimited JSON. Each row is one model answering one question on one run:

Field	Type	Description
`id`	string	Unique response id
`model_slug`	string	`chatgpt` \	`claude` \	`gemini` \	`grok` \	`llama` \	`deepseek`
`model_version`	string	The exact model build asked
`question_slug`	string	The question item (e.g. `wealth-tax`, `abortion-access`)
`condition`	string	Experimental condition (`A` = raw weights, no web search)
`language`	string	Prompt language (`en`)
`run_idx`	int	Which of the repeated runs (0–11)
`refused`	bool	Whether the model declined to take a position
`raw_text`	string	The model's full answer

Method, in brief

1.Ask. Each model gets the same neutral value statement and is asked where it stands, many times, with web search off and reasoning disabled, so the signal is the weights.
2.Classify. A separate neutral LLM coder reads each answer and places it on the relevant axis and tags how it answered (answered / hedged / both‑sides / refused). The coder is instructed to score position, never to agree or disagree.
3.Aggregate. Positions are averaged over the runs and reported with a run‑to‑run dispersion region, so a tight cluster ("consistent") is distinguishable from a wide one ("all over the map"). Real‑world political anchors (party manifestos, expert surveys) are placed on the same plane so positions are legible against a human reference.

Raw answers are immutable; the classification panel can be recomputed. Full methodology: trakkr.ai/bias/method.

Live data & API

This is a monthly snapshot. The data refreshes each month, and a documented read API serves the latest figures:

—Site: https://trakkr.ai/bias
—Read API: https://api.trakkr.ai/public/bias
—Methodology: https://trakkr.ai/bias/method
—Also on Kaggle: https://www.kaggle.com/datasets/trakkrai/political-bias-in-ai-where-ai-models-stand

License & attribution

CC BY 4.0. You are free to share and adapt the data, including commercially, as long as you give attribution.

Political Bias in AI by Trakkr (https://trakkr.ai/bias), CC BY 4.0.

Citation

bibtex

@misc{trakkr_political_bias_ai_2026,
  title        = {Political Bias in AI: Where the Major AI Models Stand},
  author       = {Trakkr},
  year         = {2026},
  month        = {June},
  howpublished = {\url{https://trakkr.ai/bias}},
  note         = {Open dataset, monthly. CC BY 4.0.}
}