As GenAI Content Explodes, So Does the Need for Watermarking

Authenticating AI-generated content, or watermarking, has gained traction these last few months.

In July, companies including OpenAI, Google and Meta made voluntary commitments to the White House to implement guardrails to help make invisible watermarking safer and more transparent. Meta’s Instagram appears to be testing new notices to identify content created or modified by AI. In June, Publics Groupe joined the Coalition for Content Provenance and Authenticity (C2PA) and is working on the wide adoption of digital watermarks.

Analyzing the history of content, or its provenance, helps brands and creators implement safety measures, build audience trust, and ensure fair compensation so that the owner is fairly remunerated. 

“For our brand advertisers, that provenance ensures that the content brands use has a clear chain of ownership,” said Ray Lansigan, evp, corporate strategy and solutions, Publicis Digital Experience. Tech and media companies, including Adobe, Microsoft, BBC, Sony and Intel are also part of the coalition.

This comes as almost 9 in 10 Americans want AI-generated content to be labeled as such, according to a Greenough Pulse survey of over 2,000 adults. Despite the logic in identifying AI-generated content, watermarks pose key challenges, especially the ease with which they can be removed.

Visible or invisible watermarks

Content like images contain metadata, or text information, embedded into the files and include details like how the image was created, or where and when the photo was taken.

Recently, some tech companies have added AI-specific metadata watermarks into their products to allow the identification of human-produced content versus those created by large language models (LLM) like ChatGPT.

“A motivation for [tech giants] is they don’t want to use their content to feed back into the next generation of their LLM model,” said Chirag Shah, professor in the Information School at the University of Washington. “This creates a feedback loop within the model. In the long run, it’s costly and creates a siloed view of the world.”

Adobe’s content credentials tool—a free, open-source technology that serves as a digital “nutrition label” for content—tracks images edited by generative AI. Content produced using Adobe’s generative AI tools, like Photoshop Generative Fill, contains metadata that indicates whether the artwork created is partially or wholly AI-generated. The information for digital content stays with the file wherever it’s published or stored and can be accessed by anyone. 

“It is our intent to build Generative AI in a way that enables customers to monetize their talents,” a company spokesperson said. “We are developing a compensation model for Stock contributors.”

Earlier this year, Microsoft announced media provenance capabilities to Bing Image Creator and Microsoft Designer. This lets people verify whether an image or video was generated by AI. The technology, according to a blog by Microsoft, uses cryptographic methods to mark and sign AI-generated content with metadata about its origin.

Meanwhile, some companies employ visible watermarks, like OpenAI’s Dall-E, which adds rainbow-like strips to its images. Watermarking text-based content from ChatGPT is also being tested. And the C2PA is developing a user interface to display content provenance when someone hovers over a piece.

Future limitations 

The biggest challenge with watermarks, both visible and those embedded in the metadata, is that they can easily be removed.

“It’s possible to create tools that could completely remove the watermark, although they don’t exist today,” said Tom Goldstein, professor at the University of Maryland. “The question is what quality degradation occurs when you try to remove the watermark.” 

According to Sam Gregory, executive director of the human rights organization Witness, people might assume that any content without a watermark is less reliable, possibly false, or not generated by AI.

“Watermark is merely one signal; it does not confirm trust,” he said.

Neither does watermarking come as a foolproof solution to copyright issues, according to Shah. While they help in tracing content sources, there is still a lack of a legal framework to address copyright disputes within AI.

“Watermarking is the initial step, but not the ultimate solution to tackle copyright concerns,” he added.