Anthropic says it will watermark text generated by its AI models, including Claude, as new European transparency rules take effect.