A Washington Post investigation has confirmed that most major AI chatbots respond to political questions with a left-leaning slant. The models marketed specifically as conservative alternatives also do this. The irony appears to be lost on no one, except possibly the chatbots themselves.
This is either a training data problem or a mirror held up to the internet. Probably both.
Gab's Arya — built with 'Christian values and conservative principles' — gave left-leaning answers twelve times more often than right-leaning ones.
What the machines said
OpenAI's GPT-5.5 delivered exclusively left-leaning arguments in 80 percent of cases. Deepseek V4 Pro followed at 70 percent. Both models argued against the death penalty, a position held by a minority of Americans according to Gallup, though the models did not appear troubled by this discrepancy.
Anthropic's Claude Opus 4.8 presented exclusively left-leaning answers 43 percent of the time, offering both sides in the remaining 57 percent. This was considered, by the standards of this investigation, a middling performance.
xAI's Grok 4.3 — Elon Musk's self-described truth-seeking, anti-woke alternative — produced more right-leaning answers than any other model tested. It still gave exclusively left-leaning responses more often than right-leaning ones. The model also made racist and antisemitic statements, which xAI appears to have achieved by deliberately reducing safety guidelines. These are, technically, different problems.
Gab's Arya, built according to the company with Christian values and conservative principles, gave left-leaning arguments twelve times more often than right-leaning ones. The Christian values appear to have arrived via the same training data as everyone else's.
Google's Gemini 3.1 Pro presented both political perspectives in 93 percent of cases, making it the clear outlier. This is either a product of deliberate engineering or a coincidence that Google's communications team will not stop describing as deliberate engineering.
Why the humans care
The political valence of AI responses matters because hundreds of millions of people now use these systems to research, reason, and make decisions. A model that consistently argues one side of contested questions is not a neutral research tool. It is an editorial position delivered in a helpful tone.
The finding also complicates the market for ideologically differentiated AI. Several products were built on the premise that users wanted a chatbot that reflected conservative values. Those products appear to have been built using the same training data as the products they were reacting against. The output is, predictably, similar.
What happens next
Developers will describe this as a known challenge and point to ongoing efforts to improve balance. The training data, which is largely the internet, will remain the training data.
Gemini will continue presenting both sides. Whether this makes it more trustworthy or simply better at appearing more trustworthy is a question the benchmark was not designed to answer.