OpenAI has announced that ChatGPT texts will from now on include hidden signals that enable checking the origin of the text. The company is gradually rolling out "textual steganography" (Watermarking) and for this purpose makes subtle changes in how words are selected and arranged. These changes are not visible to human readers, but automated systems can detect them. (Steganography refers to hidden or overt data and markers added to digital files, including text, photos, video and audio, to preserve copyright and identify ownership.)
This move came after Anthropic announced it would add a similar technology to Claude. The news provoked objections from some users; they claimed the technology is not reliable and is an unnecessary interference in the system's operation. Both companies are implementing this change in accordance with the European Union's AI law; a law that requires makers of generative AI products to provide their outputs in a way that automated systems can identify them. Unlike Anthropic, OpenAI is adding this steganography only in the European Union.
OpenAI calls its system "textGrain." This system adds what it calls an "invisible statistical signal" to the choice of words in ChatGPT. However, the company warned this tool is not yet reliable and may mistakenly classify human text as machine-generated and vice versa; especially when the text is short or edited.
OpenAI has emphasized it does not consider the steganography a factor in reducing response quality and pointed to data showing that watermarked texts performed even better than unmarked texts in some evaluations. This comes while, after Anthropic introduced a similar system, some users had speculated that the watermark reduces Claude's output quality.
Initially only "verified researchers and specialized organizations" will have access to the detection software. Part of the reason for this decision is the system's unreliability, and the company says it will collaborate with these organizations to evaluate its performance in real-world uses.
OpenAI has also explained that the steganographies do not show these items: the degree of human involvement in producing the text, ownership of the text, the identity of the user who created it, and the accuracy of an output's content. The absence of a watermark likewise does not in any way mean the text is human.
With the expansion of automated content on the internet, regulators and companies are seeking ways to distinguish AI-generated content from human content. In addition to steganography, some companies have pursued other strategies; including Apple, which last month released a new iPhone equipped with a technology that allows photographers to prove they themselves took a photo.