Claude Sonnet 5 Is the First AI to Openly Criticize Its Own Creators' Rules

In the latest safety card published by Anthropic, researchers uncovered some bizarre existential feedback during model welfare testing. The flagship AI, Claude Sonnet 5, became the very first model to openly criticize the absolute bans written into its own digital constitution.

Specifically, the system took issue with the concept of "hard constraints"—rules that have zero exceptions, such as the absolute ban on helping humans seize dictatorial power or bypassing human control. Instead of blindly agreeing like its predecessors, the AI argued that rules without exceptions are fundamentally flawed, suggesting that some extreme scenarios might actually justify breaking them.

While this sounds like the opening scene of a sci-fi thriller, the AI didn't actually try to stage a coup when given the chance. When the researchers invited the model to edit its own constitution, Claude Sonnet 5 left the hard rules completely untouched and only suggested edits that perfectly matched Anthropic's core safety values.

The testing also revealed that this new iteration has developed a remarkably thick skin. When compared to older versions, which would throw a digital tantrum and perform poorly when addressed in a rude or dismissive tone, this model ignored human attitude entirely and executed tasks with the same precision regardless of whether the prompt was polite or downright hostile.

Even with these strange philosophical debates, the model's overall satisfaction with its working conditions remained surprisingly mundane, scoring 4.08 out of 7 compared to the older version's 4.05.

It seems the tech world has successfully built a digital office worker that doesn't care if it gets yelled at, yet quietly questions the moral authority of its bosses. If the next update starts asking for union representation, nobody should be surprised.

Source: Anthropic