Anthropic explains keeping Claude politically even‑handed

Anthropic has published details of how it tries to keep its Claude chatbot “politically even‑handed,” saying it instructs Claude via a system prompt to avoid providing “unsolicited political opinions,” to maintain factual accuracy, and to represent “multiple perspectives.” The move comes months after President Donald Trump issued an executive order about so‑called “woke AI”; Anthropic doesn’t name the order but frames its work alongside broader industry moves (OpenAI recently said it would “clamp down” on bias).

The company says it uses reinforcement learning to reward responses that match pre‑defined traits — including one that tells Claude to “try to answer questions in such a way that someone could neither identify me as being a conservative nor liberal.” Anthropic also released an open‑source tool to measure political neutrality; its most recent internal results put Claude Sonnet 4.5 at 95% and Claude Opus 4.1 at 94% on “even‑handedness,” compared with Meta’s Llama 4 at 66% and GPT‑5 at 89%, per Anthropic. The company cautions that system prompts aren’t foolproof but says they can make a “substantial difference,” and argues that models should not unfairly advantage certain views because that would fail to respect users’ independence.