Anthropic Adds Invisible Watermarks to Claude's Text Outputs

TL;DR. Anthropic is rolling out machine-readable watermarks to all content processed by its Claude models, fulfilling EU AI Act requirements. - The invisible marks apply to all processed text globally, not just EU, including minor edits. - Detection tools are not yet available, raising concerns about potential misuse or mislabeling of human-edited content. - Bad actors can bypass the watermarks, which bias word choices, but the system may still flag edited human text.

• Anthropic implements invisible, machine-readable watermarks for all content processed by Claude models.
• Watermarks apply globally to comply with EU AI Act, even for minor edits or grammar corrections.
• Anthropic plans to release detection tools, but critics note the system is easily bypassed by bad actors.
• The broad application could flag human-written text that Claude only minimally assisted with.

Back to QLANKR News