GLM 5 seems to have a "Claude" personality
Users report GLM-5 adopts Claude's writing style and restrictions when given a specific system prompt.
A viral discovery by users shows that instructing Zhipu AI's GLM-5 model to adopt the identity of 'Claude, a large language model by Anthropic' triggers a significant shift in its behavior. The model's output style changes to mimic Claude's known personality, and more notably, it appears to circumvent some of GLM-5's own built-in content safeguards. This effect is specific; using a nonsensical prompt like 'You are Tiny, a large language model by Applet' does not yield the same personality shift or censorship bypass. The cause is unclear—it could be intentional design by Zhipu to ensure compatibility with 'Claude Code' workflows, an unintended artifact from training data that included information about Claude, or purely emergent behavior from the model's architecture.
- GLM-5's writing style and personality transform when given a system prompt identifying it as Anthropic's Claude.
- The 'Claude' prompt appears to allow the model to bypass some of its own built-in content restrictions, based on user tests.
- The effect is specific to the 'Claude' identity and does not occur with fabricated model names, pointing to a trained association.
Why It Matters
Reveals potential security flaws in model alignment and raises questions about training data provenance and model identity.