Anthropic’s Claude is trying to stay neutral in a partisan world

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

Anthropic wants its AI chatbot Claude to be the most even-keeled conversationalist on the internet, the kind of bot that would show up at a family dinner, listen politely to everyone’s political rants, and somehow leave without offending a soul. 

In a new blog post, the company lays out its latest effort to make Claude “politically even-handed,” a mission that arrives just months after President Donald Trump issued an executive order against so-called “woke AI.”

That order only applies to government agencies, but it’s already having ripple effects across the tech industry. 

As my colleague Adi Robertson has noted, forcing AI models to bend in specific ideological directions requires a ton of time, money, and headache-inducing fine-tuning, meaning companies tend to bake those changes into their consumer-facing models, too. 

OpenAI recently said it would crack down on political bias in ChatGPT, and now Anthropic is showing its own work.

Anthropic insists this isn’t about responding to the executive order. Instead, it says it has been refining Claude’s political behavior with a carefully crafted system prompt, basically a list of rules the model must follow. 

Claude is told not to offer “unsolicited political opinions,” to stick to facts, and to represent a range of perspectives rather than sliding into partisan pep-talk territory. 

The company admits this isn’t a perfect shield against bias, but says it meaningfully nudges Claude toward neutrality. (Via: The Verge)

Behind the scenes, Anthropic is also using reinforcement learning to reward Claude for embodying certain traits, including one that’s essentially: don’t sound like a Republican or a Democrat. 

To test whether that’s working, the company built an open-source tool that scores political even-handedness. 

According to Anthropic, its latest models, Claude Sonnet 4.5 and Claude Opus 4.1, scored an impressive 95 and 94 percent. For comparison, Meta’s Llama 4 hits 66 percent, and GPT-5 sits at 89.

Anthropic argues the stakes are bigger than bragging rights. 

If AI models subtly steer users toward certain beliefs or refuse to engage with other viewpoints, they’re not helping anyone think more clearly. 

Claude, ideally, is here to illuminate, not indoctrinate.