I’ve been reporting on AI image and video generators, and the push for labels on AI-generated content, for years now. So I was pretty surprised when a recent announcement from Anthropic, saying it will begin watermarking AI-generated text and files, sparked outrage from some AI loyalists.
To be sure, Anthropic isn’t doing this out of the good of its corporate heart, or even to make you mad. It’s doing it because it’s legally mandated to in Europe, thanks to a newly enacted regulation. The Claude maker wants to avoid a legal headache and potential fine. Its solution is to add an invisible signature to text and files made by its chatbot that can, eventually, be decoded.
The backlash was swift. Some people reportedly started canceling their Claude subscriptions in a bizarre boycott. All for watermarks that we can’t even spot for ourselves yet.
It’s puzzling to me, particularly after years of seeing AI media become indistinguishable from real photos and videos, when labels and watermarks are largely seen as the most basic responsibility an AI company should take to avoid confusion. It leads me to ask: Why would anybody think, in this world of AI slop and deepfakes, that AI-created text ought to be excluded from transparency efforts?
I get it: Text feels different from images and videos. To a certain extent, it is. Having an AI tool polish up your grammar or draft an email reply doesn’t seem as consequential as using ChatGPT to create an entire event flier or Midjourney to generate a video clip. Text-based AI tools are also very often used in workplaces, where some employers are pushing hard for AI adoption.
But I’m sorry, if we want any transparency around AI, we have to hold the line. If you used AI, say so. If you don’t want people to know you used AI, you probably shouldn’t have used it in the first place.
Stop assuming everyone uses AI #
Opposition to Claude watermarks seems to be partly coming from AI diehards. These people love AI; they use it all the time, and they think you do, too. Their argument against Claude watermarks is some version of, “I use AI constantly. I shouldn’t have to disclose it every single time.” In other words, they believe it should be assumed their work has some element of AI assistance.
It’s an understandable position, if only loosely tied to reality. Only 4% of US adults are using AI chatbots “constantly,” according to a June Pew Research Center report. Silicon Valley and AI leaders would certainly love it if we all adopted this attitude — that the AI they sell is becoming so essential to our work that it’s assumed we are constantly using it. And maybe in some industries, like software engineering, this is true. But even then, what is the harm of adding a line that says, “Refined with AI” or something similar?
The push for watermarks on AI-created images and videos has been a huge part of the conversation around AI in recent years, for good reason. It’s harder than ever to tell when a photo was captured with a camera by a photographer or generated by AI. Many (51%) of US adults want better labels on social media for AI-generated content, CNET found in February. Watermarking text is a natural next step.
Invisible labels on AI-generated text, like the kind Anthropic is proposing, could be a silver bullet to end the endless debate over whether something is AI-generated. We could finally use em-dashes without being falsely accused of using AI! Em-dashes are just one of a revolving series of so-called “clear signs” that something was made with AI. Wouldn’t it be nice to have another layer of proof where a piece of content came from?
Content provenance is by no means a foolproof process. Google’s SynthID, whose process Anthropic’s is similar to, has known limitations. The current tech doesn’t allow for a lot of nuance. For example, one concern is that if you were to write a report and have AI translate it into another language, it may label it as AI-created, even though you did the bulk of the work. But we owe it to ourselves to improve this process, not abandon it.
Reputation is everything #
I know from experience it’s nearly impossible to deter an AI believer from their enthusiasm. But AI labels won’t just apply to them. And I think there’s a significant group of Claude users who aren’t going to be excited to be “exposed” for using AI. I call them “silent AI users.”
This is an in-between group, folks who use AI but not daily. They use AI casually, often to search for information or for work. They’re no strangers to ChatGPT, but they aren’t orchestrating an army of agents. And yeah, maybe they’ve seen the reports about data centers being bad for the environment and know that AI can screw up, but it’s not enough to make them stop using AI.
In 2024, Slack found that 48% of desk workers said they would be uncomfortable disclosing they used AI to their manager; even more (57%) say they hid their use of AI. Generative AI may have become slightly more normalized in some workplaces since that survey, but it’s nowhere near universally accepted.
That leads me to a harder question: Are people really reluctant to include an AI disclosure in their work, or are they afraid to get tagged as an AI believer?
Sentiment toward AI in the US has changed dramatically since ChatGPT first dropped in 2022. Many (40%) of US adults believe AI will have a negative long-term impact on society. Concerns over deepfakes, AI slop and misinformation haven’t stopped. Lawsuits from authors, artists and entertainment studios are piling up, alleging AI companies illegally used our corpus of human art to fuel the machines that spit out cheap imitations of it. (Disclosure: Ziff Davis, CNET’s parent company, in 2025 filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
Unpopular data centers aren’t helping; a Gallup poll shows that seven in 10 US adults are opposed to data center constructions near them, with 48% strongly opposed. In an ironic twist, the very data centers necessary for the continued survival of AI are a major factor driving consumers’ outrage toward AI.
Maybe the diehards have a point that eventually, we’ll get to a point when it’s assumed AI will be a part of our daily workflows. But we aren’t there yet. It matters more than ever to be honest about how you’re producing work, whether it’s a commentary like this (I used zero AI here), a work project or a major motion picture. We owe it to each other to be transparent. Labels are one way to do so.
We shouldn’t get to pick and choose when we disclose our AI usage. If you used AI, you should be able to stand by it. Explain it, defend it. If you can’t, you probably shouldn’t use AI.