Anthropic Explains Its Approach to Measuring Claude’s Neutrality

Anthropic has shared new details about how it is working to make its Claude chatbot politically balanced, an effort emerging as the White House pushes AI companies to reduce ideological bias. The update follows President Donald Trump’s recent ban on what he calls “woke AI,” which directs government agencies to use only unbiased and fact-focused models. Although the order applies only to federal use, industry experts point out that the adjustments required to meet these expectations often influence the design of publicly available AI systems due to the cost and complexity of developing separate versions.

In a new blog post, Anthropic says its goal is for Claude to respond to political topics with equal depth and fairness across different viewpoints. The company emphasizes that it wants the model to treat opposing positions with the same level of analysis and engagement, avoiding any preference for one political side. While the post does not directly reference Trump’s order, the timing highlights the growing pressure on AI developers to demonstrate neutrality. OpenAI recently announced similar efforts to limit bias in ChatGPT.

To guide Claude’s behavior, Anthropic uses a detailed system prompt that instructs the model to avoid offering unsolicited political opinions, stay factual, and include a range of perspectives when asked about political issues. The company acknowledges that system prompts cannot guarantee neutrality but says they still produce notable improvements in balanced responses.

Anthropic also relies on reinforcement learning to encourage specific traits in Claude’s replies. One of these traits directs the model to answer questions in a way that prevents users from identifying it as conservative or liberal. This training is part of Anthropic’s broader attempt to create AI that respects user independence by avoiding arguments that push one viewpoint more strongly than another.

The company has released an open-source tool that evaluates political even-handedness across AI models. Its latest findings show high neutrality scores for Claude Sonnet 4.5 and Claude Opus 4.1, which scored 95 percent and 94 percent. These results outperformed Meta’s Llama 4 and GPT-5, which received lower neutrality ratings. Anthropic argues that ensuring fairness is essential for helping users form their own judgments rather than being subtly steered toward a particular political stance.