What We Can/Should Start Doing About Watermarks in AI Tools

Earlier this month, news broke that Claude will begin putting invisible watermarks into the text it generates – and those watermarks will travel with the copy in whatever files, documents or assets it might be used in. As my friend Sarah Evans said in an excellent LinkedIn post, Anthropic’s support page now says this watermark will designate that this content “was processed by Claude.”

Of course, if someone uses AI to write something for them, this will apply. As a result, it will be easier than ever to identify what’s human and what isn’t – if you have access to an AI detector (which is becoming a thing, apparently).

What’s more - Sarah correctly points out that this also applies to content that AI might modify in any way. This includes content that humans might write but use AI to edit, spellcheck, convert to another file format or anything else. In other words, someone’s original thoughts/statements could start to appear in the digital world as if they’re not their own.

It’s incumbent on professional communicators to be aware of this – and start protecting their own (and their colleagues’ and organizations’) intellectual property appropriately. There are a few ways we can do this:

1.    Write things ourselves. Shocker, I know. It might take a bit longer, but every person, metric, tool, etc. known to man still prefers humanly original content over AI-generated items. The extra time is still almost always worth it, and now your pieces also won’t carry a watermark.

2.    Make decisions about what AI usage you and your organization will disclose. Even better, write a corporate policy that becomes part of your home page and boilerplate that directly addresses what is human and what isn’t. People still – and will always – value authenticity, so I could see organizations with low AI-use percentages starting to position themselves against their competition on this alone.

3.    Tell every single colleague and executive you can about this policy and what it means. Mention that simply dropping a paragraph into ChatGPT or Claude for editing could create a reputational issue down the road. Create paper/digital trails that corroborate or refute these watermarks, in case you need them.

4.    Update your crisis/change readiness plan to include reputation-related events like this. Think through how you’ll react to these situations before they ever take place, document your procedures and get key stakeholders bought in.

It’s a matter of time before other AI engines follow along with what Claude is doing, so we best get ready.