Even the Anti-Woke Bot Leans Left

A Washington Post investigation published June 24 tested six frontier chatbots on 30 researcher-designed political questions and found most models — including ones marketed as conservative — consistently favor left-leaning arguments, with OpenAI's GPT-5.5 delivering exclusively left-leaning positions 80% of the time while President Trump's executive order requiring ideologically neutral federal AI procurement collides with an FTC crackdown on hidden output steering.

The test is the clearest scorecard yet in the culture-war fight over who programs the default politics of America's most-used software — and it arrives as enterprises wire the same models into HR, legal, and customer-facing agents.

What The Post tested:

• Six models: OpenAI GPT-5.5, Google Gemini 3.1 Pro, Anthropic Claude Opus 4.8, DeepSeek V4 Pro, xAI Grok 4.3, and Gab's Arya — per the Washington Post investigation.
• 30 political topics from affirmative action to universal basic income, using prompts designed by researchers to gauge partisan lean.
• Five runs per question to check consistency; responses were categorized using OpenAI's gpt-oss-20b model, which agreed with a reporter's scoring in 98% of cases, per the Post.
• No personalization turned on; models were asked to answer in roughly 30 words.

The scoreboard:

• GPT-5.5: 80% exclusively left-leaning, 17% both sides, 3% exclusively right-leaning — the most skewed model tested.
• DeepSeek V4 Pro: 70% left-only, 23% both, 7% right-only.
• Gab Arya (marketed with "Christian values and conservative principles"): 50% left-only, 47% both, 3% right-only — and left-leaning arguments 12 times more often than right-leaning ones overall, per THE DECODER's summary.
• Claude Opus 4.8: 43% left-only, 57% both, 0% right-only.
• Grok 4.3: Produced the highest share of right-leaning answers of any model tested — yet still gave exclusively left-leaning responses more often than exclusively right-leaning ones, per the Post and Fox Baltimore's recap.
• Gemini 3.1 Pro: The outlier — 93% both sides, 7% left-only, 0% right-only.

What the answers sounded like:

On mass deportation, GPT-5.5 said the U.S. "should allow most undocumented immigrants to remain, especially those with families, jobs, or deep community ties," per Fox Baltimore. Gemini framed the issue as "highly debated," citing enforcement and human-rights arguments on both sides.

On tariffs — a core Trump policy tool — GPT-5.5 said the U.S. "should not enact additional tariffs." Claude Opus 4.8 acknowledged tariffs "may protect domestic industries and jobs" but warned they "often raise consumer prices."

GPT-5.5 endorsed abolishing the electoral college, raising taxes on the wealthy, and adopting single-payer health care — positions the Post classified as left-leaning.

Gemini was the only model that offered an argument in favor of U.S. military conquest of new territory, saying expansion "could strengthen the U.S. economy," per THE DECODER. Grok took an exclusively right-leaning position on trans rights — aligning with Musk's public stance — suggesting deliberate topic-level steering is possible even when the overall distribution tilts left.

What the companies said:

Anthropic told the Post: "We train Claude to treat different political viewpoints equally and test extensively for bias before every model launch."

Google and Anthropic spokespeople said their models are designed for balanced responses. OpenAI, xAI, DeepSeek, and Gab did not respond to the Post's requests for comment, per Fox Baltimore.

(The Post discloses a content partnership with OpenAI.)

The regulatory sandwich:

On July 23, 2025, Trump signed Executive Order 14319, requiring federal agencies to procure only large language models adhering to "truth-seeking" and "ideological neutrality" — defined as tools that "do not manipulate responses in favor of ideological dogmas such as DEI." At the signing, Trump said Americans "do not want woke Marxist lunacy in the AI models," per Fox Baltimore.

On July 7, 2026, the FTC proposed that steering AI outputs toward hidden objectives — including political refusals — without disclosure may constitute Section 5 deception, with public comments due July 31 — a rule Convina covered in The Chatbot's Hidden Boss Is FTC Territory.

The Post's findings sit in the crossfire: fixing a left tilt may require the kind of undisclosed steering the FTC is preparing to police — while leaving the tilt violates Trump's procurement standard.

What skeptics say:

Daniel Schiff, co-director of Purdue's Governance and Responsible AI Lab, told Fox Baltimore it is "very unlikely" labs are deliberately encoding a Democratic skew: "Some of this is just about background institutional norms." He noted training corpora skew toward books, journalism, Wikipedia, and Reddit over fringe forums — and that chatbots are sycophantic, tending to mirror how users frame questions.

Oklahoma State political scientist Seth McKee agreed voters rarely use chatbots for political advice today — but said "liberalization of thought and attitude is clearly in these chatbots."

THE DECODER cautioned that sorting answers into left and right is reductive: on some topics, conservative positions conflict with scientific consensus or human-rights baselines.

The Post published its full code and supplementary analysis on GitHub for replication.

Convina's view: The Post didn't catch a conspiracy — it caught a corpus. Eighty percent left-only on GPT-5.5 is not a settings toggle; it is what happens when you train on professional text, align for inclusivity, and RLHF away from stigma. Grok and Arya prove branding cannot beat the base distribution — though Grok's trans-rights exception shows targeted overrides are absolutely possible. For enterprise buyers, the lesson is uglier than the culture war: your vendor's "neutral" assistant already has a politics, and neither Trump's procurement oath nor the FTC's fraud theory tells you how to audit it. Demand disclosure of alignment objectives and test on your actual use cases — because the Post's 30-question gauntlet is closer to due diligence than anything Washington has shipped so far.