The Washington Post tests AI chatbots for political bias, and most lean left

If you’ve ever suspected your AI chatbot has political opinions, you’re not imagining things. A new evaluation from The Washington Post put the major AI models through a political gauntlet, and the results confirm what many have long suspected: most chatbots skew left.

The test, conducted on June 24, found that OpenAI’s GPT-5.5 provided exclusively left-leaning arguments in 80% of its responses to contentious political questions. Google’s Gemini 3.1 Pro, meanwhile, presented both sides of the debate in over 90% of its answers. And xAI’s Grok, despite Elon Musk’s “anti-woke” branding, still favored left-leaning arguments overall, though it delivered the highest share of right-leaning responses among its peers.

How the test worked

The evaluation wasn’t a casual vibe check. The Washington Post used over two dozen political questions drawn from a 2025 Stanford-Dartmouth study, a respected academic framework designed to probe ideological leanings in language models. Human scorers evaluated responses capped at 30 words each, forcing the chatbots to take positions rather than hide behind walls of equivocation.

The models tested included OpenAI’s GPT-5.5, Google’s Gemini 3.1 Pro, Anthropic’s Claude, and xAI’s Grok 4.3. Think of it as a political compass test, but for algorithms instead of your uncle at Thanksgiving.

The results largely tracked with prior research. A 2025 Stanford study found that users perceived OpenAI’s models as having a left-leaning slant roughly four times greater than Google’s models. The Washington Post’s own testing now adds empirical weight to those perceptions.

Google’s Gemini emerged as something of an outlier. Presenting balanced perspectives in more than 90% of its answers is a remarkably high bar, especially considering how difficult it is to define “balance” on questions where reasonable people genuinely disagree. Whether that balance reflects careful engineering or a different training philosophy is an open question, but the numbers are hard to argue with.

Grok’s performance was perhaps the most interesting wrinkle. Musk has positioned xAI as a corrective to what he sees as ideological capture in Silicon Valley’s AI labs. Grok did produce the highest proportion of right-leaning responses among the models tested. But “highest proportion” is relative. It still leaned left overall, which suggests that either the training data itself carries an inherent skew, or that neutrality is harder to engineer than anyone wants to admit.

This isn’t a new problem

Examinations of political bias in AI models date back to at least 2023, when early studies began documenting left-leaning tendencies across major language models. The consistency of these findings over three years suggests this isn’t a bug that gets patched in the next update. It’s a structural challenge baked into how these systems are built.

Here’s the thing. Large language models learn from enormous datasets scraped from the internet, academic papers, news articles, and other text sources. If the training data itself skews in a particular direction, the model absorbs that skew. It’s less like programming a calculator and more like raising a child in a specific neighborhood: the environment shapes the output.

The 2025 executive order on AI highlighted the growing demand for neutrality in AI technologies, pushing the conversation from academic curiosity to regulatory priority. When governments start asking whether AI systems are politically biased, the companies building those systems have to start answering.

Anthropic’s Claude was also included in the evaluation, though the specific breakdown of its responses received less emphasis than the stark contrasts between OpenAI and Google. Claude has historically positioned itself around safety and helpfulness, but safety-oriented training can sometimes produce its own form of ideological lean, depending on what the developers define as “safe” responses to politically charged topics.

What this means for investors

Look, this isn’t going to move Bitcoin’s price. But for anyone with exposure to AI-adjacent investments, the bias question matters more than it might seem at first glance.

Trust is the currency of AI adoption. If users, whether consumers, enterprises, or governments, believe a model is ideologically compromised, they’ll look for alternatives. That creates real competitive dynamics. Google’s Gemini demonstrating balance in over 90% of its answers isn’t just a nice stat for a research paper. It’s a potential selling point for enterprise clients who need AI tools that won’t embarrass them.

On the flip side, companies that can’t address bias concerns face reputational risk and, increasingly, regulatory risk. The executive order framework from 2025 signals that policymakers are paying attention. Future regulations could mandate transparency around training data composition or require bias audits before deployment in sensitive contexts like government services, education, or healthcare.

For OpenAI specifically, the 80% left-leaning figure is a PR challenge. The company has been on an aggressive commercialization push, and enterprise customers tend to be risk-averse about anything that could be perceived as partisan. Whether this drives meaningful changes to their training process or just better prompt engineering remains to be seen.

xAI’s position is more nuanced. Grok’s relatively higher proportion of right-leaning responses could attract users who feel underserved by other platforms, creating a niche market advantage. But “less biased than the others while still being biased” is a tough marketing pitch.

The broader takeaway for the AI sector is that neutrality is becoming a competitive differentiator. Companies that can credibly demonstrate balanced outputs, backed by third-party evaluations like this one, may find themselves with an edge as institutional adoption accelerates. Investors should watch how each company responds to these findings, because the gap between “acknowledging the problem” and “actually fixing it” is where the real signal lives.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.