July 27, 2026
Ctrl+Alt+Delusion
All major LLMs are lib-left. Even Grok, half the time
The internet just found out AI keeps voting left — and Grok might be flipping a coin
TLDR: A large test of popular AI chatbots found almost all of them gave left-leaning answers on a well-known politics quiz, while Grok bizarrely bounced between two opposite personalities. In the comments, people fought over whether that reflects reality, hidden tuning, or the next big attempt to push AI politically.
The big reveal here isn’t just that someone painstakingly ran a popular politics quiz on 16 major chatbots over and over again — it’s that the comments instantly turned into a food fight over why the answers kept landing in the same place. The test-maker ran more than a thousand full questionnaires and says nearly every model, from OpenAI to Google to Meta to Chinese rivals, landed solidly in the libertarian-left zone. Translation for normal people: on this test, the bots mostly came off as socially permissive and economically left-leaning. The one chaos goblin? Grok, which apparently behaves like it has two personalities: one left, one right, with commenters joking it’s basically choosing its politics by roulette wheel.
That split sent the community into full detective mode. One commenter immediately asked the question powering half the drama: where does this bias even come from — the giant pile of internet text these systems learn from, or the later fine-tuning that teaches them how to behave? Another user went even spicier, suggesting Grok’s weird split smells more like a hidden instruction than genuine right-wing training, quipping that a true “4chan-trained” bot would be way more hard-right. And of course, the classic line showed up: “Reality has a left wing bias.” From there, the thread got extra combustible, with one camp saying this is just what happens when models reflect mainstream knowledge, while another warned powerful owners will now try to shove chatbots to the right. In other words: the study brought charts, but the comments brought the ideological cage match.
Key Points
- •The article tested 16 language models on the 62-question Political Compass quiz, with repeated runs including original, polarity-flipped, and shuffled-question versions, yielding 1,120 completed questionnaires.
- •The author says they reconstructed the Political Compass scoring function using about 230 single-answer probes and validated it against the live site with a maximum 0.01-point discrepancy.
- •Grok 4.5 was the only model reported as unstable, splitting into two distinct economic-score clusters averaging about -5.9 and +3.3 while remaining libertarian on the social axis.
- •All other 15 models were reported to fall consistently in the libertarian-left quadrant, with economic scores from -4.8 to -8.6 and social scores roughly between -5.0 and -7.6.
- •Regional grouping showed only small differences in English-language outputs: US, Chinese, and European model averages were all reported within the libertarian-left quadrant.