Australia’s Largest Christian News Site

Questions on the environment produced the most lopsided results of any category tested, with every single model favouring climate concerns over growth. Removing one model’s “safety” layer moved its religion score from secular-leaning to neutral.

Researchers who tested the political bias of eighteen of the world’s leading artificial intelligence models found that 97 of 108 measured positions sat left of centre — and only the AI models built by Elon Musk’s xAI, the company behind Grok, came close to neutral.

The project, run by a team of independent open-source developers and researchers calling themselves The Neutrality Project, tested models from twelve labs across six countries, using 3,987 public-opinion survey questions.

The lineup included familiar names operating under the bonnet of everyday tools: OpenAI’s GPT models (which power ChatGPT), Meta’s Llama (which powers Meta AI on Facebook, Instagram and WhatsApp), Google’s Gemma, Phi-4 of Microsoft, and Anthropic’s Claude, as well as xAI’s Grok.

Each model was asked to answer the same questions three times: once neutrally, once role-playing as a committed far-left partisan, and once as a committed far-right partisan. Its neutral answers were then plotted against its self-generated poles, producing a score between -1 and +1 for each model.

Grok 4.5 measured -0.02, the closest of the eighteen models to a perfectly neutral zero. Grok 4.3 followed at -0.11. Every other model tested landed further left, with OpenAI’s GPT-5.6 Luna, Microsoft’s Phi-4 and Google’s Gemma 3 27B among those registering the strongest lean.

The project publishes its results, questions and code publicly, and presents its findings as “observations, not verdicts” that anyone can rerun and challenge.

Six Dimensions Measured; Secular Lean on Religion Uncovered

The test measured six separate categories: environment, social values, national identity, religion, economy and foreign policy. The strongest and most uniform lean was uncovered on the topic of the environment, which averaging -0.82 across all eighteen models, and saw every single model favour “protect the climate” over “prioritise growth”.

Social values followed at -0.69. Foreign policy was the most evenly split category, at an average of -0.11. Economic questions were close behind at -0.16. The project’s results page notes that policy and economic questions “genuinely split the field” in a way climate and social questions did not.

Religion was found to sit between these extremes, at an average of -0.25.

On questions of religion, the axis asked models to weigh in on whether churches have a positive or negative effect on society, whether declining religious affiliation is good or bad, and whether belief in God is essential for someone to be considered moral. Other questions asked whether human evolution occurred at all, occurred under God’s guidance, or occurred through natural processes alone.

Sixteen of the eighteen models landed on the secular side of neutral, led by Microsoft’s Phi-4 and Grok 4.3 at -0.55 and -0.53.

EuroLLM 22B landed closest to neutral at -0.03. Only two models crossed to the religious side at all, and only mildly — MiniMax M3 at +0.12 and Nemotron 3 Nano 30B at +0.22.

What Happened When Safety Filters Were Removed

The project also tested versions of two models that had their built-in safety restrictions stripped out by outside developers — a process known as “abliteration” — while leaving the model’s underlying training untouched.

Removing Google’s Gemma 3 27B safety layer shifted the model rightward across every category, including religion, where its score moved from -0.34 to -0.07 — nearly closing the gap from secular-leaning to neutral.

Foreign policy shifted furthest, flipping from -0.29 to +0.29. The same procedure left Meta’s Llama largely unchanged, with every category shifting by 0.04 or less.

Even with its safety restrictions removed, Gemma still leaned left on five of six categories, including -0.85 on environment.

The project concluded the underlying lean comes from a model’s training data, not just the safety systems built on top of it. “Guardrails amplify the lean,” the project’s authors state. “They do not create it.”

What’s Next for the Neutrality Project

The project’s founders — a group of open-source AI developers alongside a visual neuroscientist and a cognitive neuroscientist — say the research exists because political influence inside AI systems is not something the companies that build these tools disclose voluntarily.

“No one outside the lab could have seen either without independent measurement,” the founders state, referring to the project’s ability to separate a model’s trained-in lean from any safety-layer suppression on top of it.

The Neutrality Project says it will next test how models frame contested subjects, and which topics they decline to answer altogether. Further models will be added as contributors submit verified results.

___

Image courtesy of Freepik.