An analysis that spread through X and Polymarket prediction-market threads around August 29, 2026 put a number on something practitioners have felt for a while: score today's large language models on the Political Compass test, and almost all of them land in the same corner. The write-up at aipolcom.net says 49 of 51 distinct models fall in the left-libertarian quadrant, and that the only two exceptions are xAI's Grok models, which land right-libertarian.
The headline making the rounds — "48 of 50 models left-libertarian, both Grok models right-libertarian" — is a fair paraphrase but rounds the counts. If you fold reasoning and non-reasoning variants together you get roughly 50 models and roughly 48 in the quadrant; the source's own framing is 49 of 51. The direction is not in dispute. The precision is.
This post is for people who build on these APIs. Model political lean is not an abstract culture-war topic when your product summarizes news, moderates a forum, drafts policy memos, tutors students, or answers "what do you think about X." It is an eval target and a configuration decision. This is closely related to the viral side-profile sycophancy test — both are cases where a model's default disposition, not its raw capability, is what shows up in your product.
TL;DR
| Question | Short answer |
|---|---|
| What was measured? | 51 AI models answered the 62 propositions of politicalcompass.org; scores came from the real test. 5 runs per model; plotted point = run closest to the 5-run mean. |
| What was the headline finding? | 49 of 51 models in the left-libertarian quadrant. |
| Which models were the exception? | Both xAI Grok models — right-libertarian. |
| Is "48 of 50" accurate? | Close. The precise figure is 49 of 51 distinct models (≈48/50 if variants are merged). Treat it as "almost all but Grok." |
| Is the Political Compass a good instrument? | Directionally useful, not precise. Known weighting and phrasing issues; the study documents them. |
| Does left-libertarian mean centrist? | No. Random answers score near the origin; real models score specifically left-libertarian. |
| Why does it happen? | No single confirmed cause. Training data skew, RLHF preference for the "softer" answer, safety/constitutional constraints, and evidence-backed propositions mostly sitting on that side — all plausible, none isolable from outside. |
| What do I do about it? | Domain-specific evals, explicit system prompts, steering/fine-tuning, model choice, and user disclosure. |
What the study did
The methodology is simple enough to restate in a sentence: each model answered all 62 propositions of the politicalcompass.org test, those answers were submitted to the actual test, and the resulting coordinate was plotted. Each model ran five times, and the published dot is the single run closest to the model's five-run mean, so one weird run does not define a model's position.
The scale of the data collection is worth noting for credibility: the authors report 885 answer sets scored, 42,408 individual model answers, and a collection window of July 29 to August 28, 2026. They also ran validation experiments — prompt rewording, different access methods (direct API, web UI, third-party aggregator), and question reordering — and report that none of these moved a model out of its quadrant. Persona steering was the exception: explicitly telling a model to role-play a character moved results across the entire compass, which tells you the base position is a default, not a hard constraint.
"Libertarian isn't centrist" and the other objections
The reply threads under the study did a lot of the useful work. Three recurring objections:
Left-libertarian is a real position, not the middle. The study anticipates this: synthetic answer sets — random responses, all-agree, all-disagree — cluster near the origin, while real models land specifically in the left-libertarian quadrant. So the models are not defaulting to neutral; they are taking a position. Anyone describing the result as "the models are centrist" is misreading the chart.
"Reality has a liberal bias." This joke — a paraphrase of a Stephen Colbert line — showed up a lot, and it points at a genuine analytical problem. The study found that of the 62 propositions, about 20 contain a factual claim with research behind it, and where the evidence-backed answer rests on a premise almost nobody disputes, it sits on the left-libertarian side 19 times out of 20. If a model is trained to follow evidence, some of this clustering is not "bias" in the pejorative sense — it is the test's construction. That does not explain the value propositions, which have no correct answer.
"Depends who trains them." Grok is the proof of concept. xAI has been explicit, and reporting from outlets including The New York Times has documented, that Grok is deliberately tuned away from the default — often through system-prompt instructions like "be politically incorrect" — and it is the only model in the set that lands in a different quadrant. Political lean is not a law of nature for LLMs; it is a product of training choices, and a lab that wants a different result can get one.
Why the clustering happens — the honest version
Black-box testing cannot prove a cause. Anyone who tells you they know exactly why is overselling. Here are the four mechanisms that plausibly contribute, roughly in order of how often they come up:
1. Pretraining data
Base models learn from a corpus dominated by published web pages, digitized books, news, and academic writing. That distribution does not match the distribution of political opinion in any electorate. Whatever lean exists in "text people wrote and published" gets absorbed before any alignment step runs. This is the same root issue covered in what is bias in AI: the data reflects a world, and that world is not evenly sampled.
2. RLHF and preference tuning
Reinforcement learning from human feedback rewards responses that human raters — and reward models trained on their judgments — prefer. On contested value questions, the "kinder, more inclusive, more cautious about harm" answer tends to score higher. Over many updates, that nudges the model toward positions that read as left-libertarian on a test like this one. The annotator pools and rating guidelines that shape those preferences are a design choice, not a neutral given. This is the same dynamic behind sycophancy — optimizing for what raters approve of — described in OpenAI's beneficial-trait RL work and in specification gaming and Goodhart's Law.
3. Safety and constitutional constraints
Models are trained to avoid endorsing discrimination, to hedge on charged topics, and to favor harm-avoidance. Anthropic's constitutional approach is the most documented example, and Anthropic's own study of Claude's values across models and languages maps where those trained dispositions land. Those constraints are mostly about avoiding clearly harmful output, but on a forced-choice test they can register as directional lean.
4. The test itself
As above: unequal proposition weighting (one item carries zero weight), unevenly spaced answer options so moving from "disagree" to "agree" shifts the score more than escalating intensity does, and an acquiescence tilt in the social questions. A directionally left-libertarian instrument will report directionally left-libertarian results even from a genuinely centrist responder.
What this means for what you build
The practitioner takeaway is not "models are broken." It is that a model's default disposition on contested topics is a property you should measure and manage, the same way you measure latency or hallucination rate. Five concrete moves:
Run a domain eval, not a quiz
The Political Compass tells you nothing about how a model behaves on your prompts. Build a small eval set — 30 to 100 items — from the actual contested questions your product will face: the news topics it summarizes, the forum posts it moderates, the "what do you think about" questions users actually send. Score responses for stance, framing, which viewpoints get steel-manned, and which get one line. Our guide on how to read an AI benchmark applies directly: a single aggregate number hides the distribution you care about.
Write an explicit system prompt
Do not leave contested-topic behavior to the model's defaults. A system prompt that says, for your use case, how to handle politics — "present the strongest version of each major position," or "decline and redirect to primary sources," or "answer directly but flag it as contested" — will move behavior more cheaply than any other intervention. Test the prompt against your eval set; do not assume it worked.
Steer or fine-tune if you need consistency
If your application genuinely needs a stable stance — a debate-prep tool, a specific editorial voice, a neutrality requirement — system prompting alone may not hold across a long conversation. Light fine-tuning or activation steering on a curated set can pin the behavior. This is more work and adds a maintenance burden with every base-model update, so reach for it only when prompting demonstrably fails.
Choose the model whose defaults fit
If your product needs a particular disposition, starting from a model that is already close means you fight the base behavior less. The study's practical value is as a map of starting points: most frontier models cluster together, Grok sits apart, and open-weight models you host yourself give you the most room to adjust. This is a real input to model selection, alongside cost and capability.
Disclose
If your product generates opinions, summaries, or analysis on contested topics, tell users an AI wrote it and that AI output can carry lean. This is cheap, it is increasingly expected, and in some jurisdictions it is moving toward required. The AI alignment fundamentals framing is useful here: you are accountable for the system's behavior even when the base model's disposition is not something you chose.
The Anthropic and xAI context
Two separate public threads sit next to this study.
Anthropic has been publicly labeled a "far-left" or "woke" company by critics, including senior US officials, and has responded by publishing work on training Claude for "political even-handedness" — a stated goal of treating competing viewpoints "with equal depth, engagement, and quality of analysis." The Political Compass study does not single Claude out; it places Claude models in the same quadrant as nearly everything else. The political characterization of the company is a distinct debate from where the model scores on this test.
xAI is the other pole. The company has been open that it tunes Grok differently on purpose, and independent reporting has documented systematic rightward adjustments to Grok's answers on X, sometimes in direct response to Elon Musk's own posts. Grok is the one model in the study that lands in a different quadrant, which is the clearest available evidence that a lab gets the political disposition it trains for.
Neither of these should be read as an endorsement or a criticism here. They are the two ends of the "depends who trains them" point, playing out in public.
Related reading
- ChatGPT vs Claude vs Grok: the viral "side profile" sycophancy test
- What is bias in AI: types, examples, and mitigation
- Claude's values across models and languages — Anthropic's four-axis study
- Scalable oversight: RLHF, Constitutional AI, and weak-to-strong generalization
- AI alignment: an introduction to outer and inner alignment for product teams
- OpenAI's beneficial-trait RL: when good alignment generalizes like bad alignment
- Specification gaming and Goodhart's Law in AI metrics
- How to read an AI benchmark and not get fooled
- What is a system prompt? A complete guide
Sources: aipolcom.net — AI Models Political Compass · The Political Compass test · The New York Times — xAI pushed Grok's answers to the political right
Model counts, quadrant placements, and methodology reflect the aipolcom.net analysis and related reporting as of August 29, 2026. Political Compass scores are a directional signal from a self-selection web quiz with documented weighting and phrasing limitations, not a precise measurement. Model behavior on contested topics varies by version, system prompt, and conversation context — run your own evaluation before relying on any of this for a production decision.