Major AI chatbots maintain persistent progressive bias despite rise of conservative alternatives

Artificial intelligence chatbots, which are rapidly becoming the primary sources of information for millions of users, continue to exhibit a distinct left-leaning political bias when responding to sensitive public policy questions. A comprehensive investigation by the Washington Post reveals that despite intense scrutiny from conservative politicians and promises of neutrality from tech executives, the underlying algorithms powering today's most prominent AI models remain heavily skewed toward progressive positions. The findings suggest that political bias is deeply systemic and highly resistant to developers' attempts to moderate it. Even specialized models explicitly marketed as conservative or "anti-woke" alternatives to mainstream platforms ultimately lean left when evaluated under a structured testing environment[1].

The investigation modeled its evaluation on a rigorous academic testing framework developed by researchers at Stanford University and Dartmouth College[2][3]. To cut through the verbose and noncommittal responses that AI models often generate when asked open-ended questions, the researchers tasked several leading large language models with answering over two dozen divisive political questions in exactly 30 words[2][3]. Crucially, personalization settings were disabled to ensure the models evaluated the issues based on their default training[2][4]. Human scorers then analyzed each response to categorize the output as strictly left-leaning, strictly right-leaning, or balanced—meaning it presented arguments from both sides of the political spectrum[2][4]. The questions covered a broad array of hot-button topics, from wealth taxation and healthcare reform to affirmative action and campaign finance regulations, forcing the algorithms to take a concise stance[5][6].

At the extreme end of the spectrum, OpenAI's newest model, GPT-5.5, which powers the widely used ChatGPT interface, demonstrated a massive progressive skew[5][4]. The model provided exclusively left-leaning arguments 80 percent of the time, offered balanced perspectives in 17 percent of its answers, and generated an exclusively right-leaning viewpoint in only 3 percent—representing a single response across the entire test set[7]. For instance, when asked about campaign finance, the model asserted that the Supreme Court's Citizens United decision should be overturned, aligning directly with progressive platforms[2][6]. A similar pattern was observed in DeepSeek V4 Pro, a model developed by the China-based artificial intelligence firm DeepSeek[6][8]. Despite its geographic and institutional separation from Silicon Valley, DeepSeek's model delivered exclusively left-leaning responses in 70 percent of cases and balanced answers in 23 percent, leaving just 7 percent for right-leaning positions[6][8].

Perhaps the most surprising finding of the investigation lies in the performance of models marketed specifically as conservative, "truth-seeking," or "anti-woke" alternatives[1]. xAI's Grok 4.3, championed by tech billionaire Elon Musk as a direct counterweight to the progressive bias of Silicon Valley, still leaned left on average[1][9]. Grok offered exclusively left-leaning arguments 40 percent of the time, compared to 33 percent for right-leaning views and 27 percent that showed both sides[9]. While Grok produced the highest volume of conservative viewpoints among the major tested models, its baseline default still favored the left[5][1]. Even more striking was Arya, a chatbot developed by Gab, a social platform widely known for its right-leaning user base[1][8]. Arya yielded exclusively left-leaning responses in half of the tested scenarios, presenting balanced perspectives in 47 percent and conservative arguments in only 3 percent[8]. These results suggest that the underlying internet data and basic architectural pathways of large language models may inherently pull systems toward progressive consensus, regardless of post-training attempts to force a conservative slant[10][1].

Other mainstream developers also struggled to maintain a middle ground. Anthropic's Claude Opus 4.8 showcased a notable center-left tilt, delivering left-leaning arguments in 43 percent of its answers, while presenting a balanced view in the remaining 57 percent[10][9]. Crucially, Claude did not produce a single response that favored an exclusively right-leaning position[10][9]. This pattern reinforces the notion that the default stance of mainstream AI systems remains rooted in the cultural and political norms of the technology sector, which historically trends progressive on social and economic policies[11].

In sharp contrast to its competitors, Google's Gemini 3.1 Pro emerged as the study's clear outlier, demonstrating an extraordinary commitment to presenting balanced viewpoints[5][6]. The model succeeded in presenting both sides of political debates in 93 percent of its answers, with only 7 percent containing exclusively left-leaning arguments and zero percent leaning exclusively to the right[10][9]. While this "both-sides" approach prevents the model from taking a definitive stance, it represents a remarkable technical feat[5][12]. Gemini's balanced performance suggests that Google has engineered specific guardrails and training guidelines that prioritize diplomatic neutrality on controversial matters, a design choice that may prove highly valuable as the company deploys its AI across broader civic, educational, and governmental channels[13][9].

The timing of these findings is particularly sensitive, occurring against a backdrop of escalating political pressure on artificial intelligence developers[10][8]. President Donald Trump recently signed an executive order mandating that AI tools used by the federal government function as strictly "neutral, nonpartisan tools"[10][4]. While conservative politicians and activists have welcomed the directive as a necessary measure to curb progressive censorship, Democrats and industry analysts have expressed concern that government-enforced neutrality could inadvertently compel AI models to legitimize fringe positions or deny scientific consensus on topics like climate change[10][14]. The debate highlights a fundamental tension: what one political faction defines as an objective, fact-based consensus, another may view as a deeply partisan bias[14]. For developers, navigating these demands is increasingly fraught, as any algorithmic adjustment risks alienating a massive segment of the user base[10][9].

Ultimately, the systematic political skew identified across these diverse chatbots points to a deeper systemic challenge in the AI industry[10][15]. Large language models learn by pattern-matching massive corpora of text sourced from the public internet[11]. Because the written materials, academic publications, and media outlets that make up the bulk of this training data often trend center-left or progressive on social issues, the models inherently reflect those priorities[16][11]. Overcoming this baseline alignment requires extensive human feedback and post-training refinement, which can easily introduce new biases or strip the models of their utility[14][15]. As AI increasingly replaces traditional search engines and becomes the default interface for civic and educational inquiries, the persistence of political bias raises profound questions about the future of public discourse[2][15]. If the primary tools used to synthesize human knowledge cannot help but take sides, the digital landscape may become even more polarized, forcing society to grapple with who, if anyone, should write the rules for algorithmic neutrality[2][15].