AI Watermarks: What New Policies from Anthropic, Google, and Adobe Mean for VCU
- The Policy Change: Major AI providers are embedding invisible watermarks and metadata into AI-generated text and media to align with the EU AI Act.
- Who It Impacts: VCU faculty, staff, and students using AI tools for writing, research, and coursework.
- VCU’s Connection: Understanding how watermarking works helps our campus navigate academic integrity, data privacy, and the use of university-supported AI tools.
In August Anthropic announced that all new Claude models will embed watermarks into generated text. This update helps meet European Union AI Act requirements, which mandate that AI-generated content be detectable.
Other GenAI Tool Watermarking
Other major tech companies are rolling out similar tracking methods:
- Google: Embeds SynthID watermarking across Gemini text, image, and audio outputs.
- Adobe: Uses mandatory Content Credentials (C2PA metadata) across Creative Cloud tools to track AI edits.
- OpenAI: Attaches metadata to images, but has delayed text watermarking for ChatGPT over concerns about accuracy and fairness to non-native English speakers. OpenAI does provide an image checker.
Why It Matters to Higher Ed
Text watermarks are not visible logos or hidden letters you can highlight and delete. Instead, they make subtle, mathematical tweaks to word choices as the AI generates text.
For VCU, this changes how we identify AI content in classrooms and administrative work. A watermark only proves that an AI model processed the text. For example, if a student simply uses AI to proofread or translate their own original writing, it may still trigger a watermark detection.
It’s also worth noting that heavy editing, paraphrasing, or translating AI text can erase these watermarks altogether. And, in keeping with the arms race nature of AI, there are already free watermark removing tools out there.
All this is to say that watermarks are indicators of origin, rather than absolute proof of origin. As Anthropic explains, “A watermark can only determine that Claude was likely involved with the content at some point. It cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.'”
Categories AI and Emerging Tech Policy and Guidance, AI Resources, AI Risks, AI Workforce Impacts