OpenAI to Add Invisible Watermarks to ChatGPT Text in EU; Set to Test New Ad Format

OpenAI is all set to introduce invisible watermarks for text generated by its AI models. This latest move will make it easier for readers to identify AI-generated content. This feature will be initially available in the European Union, in response to the EU AI Act, which requires providers of generative AI systems to make machine-generated content identifiable. Additionally, ChatGPT will show more ads in the coming days. The new ads will appear alongside images that users ask ChatGPT to generate.

OpenAI Announces AI Text Watermarking

In a blog post on Monday, OpenAI announced that text generated by ChatGPT and Codex in the European Union will show an invisible watermark over the coming weeks. API customers worldwide can enable the text watermarking for select models. Text watermarking will remain disabled by default.

OpenAI said that it has started accepting applications from researchers and expert organisations interested in testing its text watermark detector. It confirmed that text watermarking was introduced in response to the EU AI Act.

OpenAI's watermarking system, known as textGrain, adds an invisible statistical pattern to the model's word choices. The detector looks for that signal to assess whether a passage contains an OpenAI watermark. The company said textGrain matched or exceeded the performance of other watermarking methods in testing, including Google's SynthID for text. OpenAI is also planning to release the underlying technology as open source.

OpenAI acknowledged that editing AI-generated text will reduce the effectiveness of the watermark. The company says during its evaluation, the detector identified around 80 percent of watermarked 200-token passages at a 1 percent target false-positive rate. The detection rate increased to roughly 95 percent for 400-token passages in tested areas such as psychology. Performance was less consistent for subjects such as mathematics.

The company claimed that in the evaluation of 400-token passages, replacing 10 percent of words with synonyms reduced detection from around 92 percent to 66 percent. Replacing 25 percent of words reduced it to 17 percent.

OpenAI acknowledged that the text watermark can not measure how much human input was involved in a text. It also can not determine ownership of the content. It can indicate OpenAI involvement, but it should not be considered as proof of authorship. 

Additionally, OpenAI announced that the ChatGPT app will add more space for ads, starting with a test later this month. The company will test the “new visual ad format” initially during image generation in ChatGPT. The company says ads will be labelled and remain separate from the image being created.

The images will show consumers details about new products, product usage, or the experiences they make possible. For advertisers, the new feature will create new opportunities to reach customers. "As with our other ad formats, advertising does not influence the answers ChatGPT provides", OpenAI said. OpenAI confirmed that the testing of the new ad format will begin later this month in the US with an initial group of advertisers.