Anthropic has announced a major update to its cloud AI chatbot, targeting concerns about political bias. According to the company, this update aims to achieve better symmetry when the model deals with political topics. This update comes at a time when large language models are being actively criticized for showing inaccurate information or bias, and for the ways in which AI systems shape public perception.

Anthropic outlined the new changes in a blog post and technical documentation, explaining that the update is a meaningful step toward building AI that people across the political spectrum can trust. One of the key aspects of the update is a set of clear instructions given to the cloud. In these instructions, LLMs have now been clearly told to avoid giving unwanted political opinions. Additionally, the company notes that the model has been trained to present information with balanced depth and analysis, regardless of which viewpoint it is addressing.

Anthropic says this framework – which it calls “political parity” – pushes Cloud to prioritize factual accuracy, represent multiple viewpoints, and avoid loaded or partisan terminology unless specifically necessary. The aim is not to keep Claude away from politics completely, but to ensure that he does not quietly lean to one side or the other.

To reinforce this behavior, Anthropic has sent system-level signals and reinforcement learning rewards to the cloud, which are said to be designed to motivate them toward neutrality. These include presenting the strongest possible version of different political arguments when asked, engaging respectfully with all viewpoints, and passing the “ideological Turing test”, meaning the model must be able to clearly express any position, even if it does not “agree with” it.

Anthropic admits the system is not foolproof, but says the update makes a “substantial difference” in the quality and evenness of the cloud’s political responses.

In an effort to bring more transparency to the process, Anthropic has also open-sourced its assessment tool for political parity. The company claims that its latest models, Cloud Sonnet 4.5 and Cloud Opus 4.1, have scored 95 and 94 percent respectively in its tests, which is reportedly higher than Meta’s Llama 4 and OpenAI’s GPT-5.

The cloud’s creators say the idea behind publishing the framework is to encourage the rest of the industry to measure bias more consistently. “A shared standard for measuring political bias would benefit the entire AI industry and its customers,” Anthropic writes, adding that he hopes other developers will join the effort.

According to Anthropic, these changes follow a long recent conversation about politics without pushing it in any particular direction. The company argues that most people seek “honest, productive discussion”, and an AI that subtly favors certain views, whether through tone, framing or persuasion, ultimately undermines the user’s freedom.