“India is outpacing the world,” said OpenAI CEO Sam Altman, celebrating the nation’s enthusiastic embrace of ChatGPT’s ImageGen model. Since the launch of its image generator, users have crafted over 700 million images, with a staggering 130 million people using the feature globally. But as the model’s power continues to amaze, OpenAI is now experimenting with a system of watermarking a decision that has both piqued curiosity and spurred debate among ai researchers, digital artists and technology enthusiasts.
This has led to the development of the ImageGen model, which has generated one of the most potent multimodal systems in the current AI landscape as part of ChatGPT 4o. By blending together text prompts and advanced image generation, it enables its users to make beautiful art, even in the Studio Ghibli aesthetic style. Originally it was limited to paid subscribers, but its recent expansion to free-tier users has enormously widened access. But just as it expands, it faces the challenge of maintaining authenticity and protecting intellectual property, which has led OpenAI to investigate watermarking solutions.
The watermarking feature, which was first noticed in the beta version of the ChatGPT Android app by the AI researcher Tibor Blaho. He called it “image-gen-watermark-for-free,” implying that watermarking would be applied only to the images produced by free-tier accounts. ChatGPT Plus subscribers, however, would still be able to download images without the watermark. This differentiation begs the question of the implications for creative licensing and professional usage. But with watermarks being a symbol of authenticity, it can also act as a hindrance for artists who want to use AI-generated visuals in their portfolios or projects.
Watermarking has been done for AI domain previously. Tech giants such as Google and Microsoft have already integrated comparable systems in their generative models. For example, Google’s SynthID embeds invisible watermarks in images generated by its AI, while Microsoft applies cryptographic markers to images created by its DALL-E 3 model. OpenAI’s stance seems to fit in with these efforts, to try and balance transparency with usability. As OpenAI wrote in a blog post, “Combined with aggressive post-training, the resulting model has surprising visual fluency, capable of generating images that are useful, consistent, and context-aware.” But will watermarking actually build trust for AI content, or will it further kill its creativity?
Watermarking poses substantial technical difficulties. Invisible watermarks, though more difficult to tamper with, aren’t foolproof against manipulation. Simple edits such as cropping or pixel tightening can make them undetectable. Watermarks that are visible as opposed to invisible can protect authenticity but lose art value in such portions of content. Finally, the persistence of watermarking depends on the content; images are more resistant than textual data, as shown in studies focusing on AI fingerprints. This illustrates the necessity for bespoke solutions to tackle the specific issues of each media type.
Apart from the technical problems, watermarking has ethical and societal issues. For example, does a watermark unintentionally stigmatize AI-generated content? According to a study by Witness, a human rights and technology organization, such tracing is a potential violation of privacy and could risk civil liberties. Watermarks with identifying information could expose creators to risks in sensitive contexts. As discussed in Access Now’s discussion paper, “No one using generative AI should have to reveal their personal information to third parties in order to do so.” These considerations underscore the need to design watermarking systems that do not threaten user privacy or freedom of expression.
OpenAI’s watermarking experiment also partakes of broader policy debates. The White House’s executive order on AI highlights the importance of strong transparency mechanisms such as watermarking to provide safety and trustworthiness. But, as Chad Heitzenrater of the RAND Corporation warned, “Failure to address watermarking as a systems problem is likely to result in solutions that risk working against the security and trust they seek to instill.” Policymakers should consider the trade off between the benefits of watermarking against its downsides like higher computational costs and risks of misuse.
For digital artists and tech funding enthusiasts, the implications of watermarking are even deeper. On one hand, it can try and identify AI visuals from maps created by human artist, promoting trust in visuals being linked to the content. Conversely, it may restrict the creative potential for something like ImageGen, especially for free-tier users. As OpenAI’s chief operating officer, Brad Lightcap, noted, “Very crazy first week for images in ChatGPT – over 130M users have generated 700M+ imag since last Tuesday. India is now our fastest-growing ChatGPT market. The range of visual creativity has been extremely inspiring,” But that creativity needs to be counterbalanced with ethical safeguards and content provenance.
In the end, OpenAI’s watermarking effort also represents the changing world of generative AI. And as technology becomes easier to access questions about authenticity ownership and social impact will keep shaping its life path. We’ll see whether watermarking is the answer or a headache, but we do know one thing: the discussion surrounding AI-generated content isn’t going anywhere.
