OpenAI Confirms Availability of Text Watermarking Technology

OpenAI has confirmed that it has developed the technology to determine if a block of text has been generated using its AI model. However, it says no decision has been made yet whether to release it.

The availability of this technology was revealed by a Wall Street Journal report over the weekend, which claims OpenAI’s text watermarking system has been ready for at least a year and can detect AI-generated text with 99.9% certainty.

Ongoing debate

The report compelled OpenAI to update a blog it quietly published in May describing the technology but which tiptoed around the availability of a working tool. In an update made on the same day of the Journal report, OpenAI wrote: “Our teams have developed a text watermarking method that we continue to consider as we research alternatives.”

The technology works by adding a pattern to how it writes its output, allowing OpenAI to detect if it was generated using its AI models. It is apparently unnoticeable to humans and didn’t impact its quality, though for obvious reasons will only work for content generated by ChatGPT, not other LLMs.

OpenAI claimed an ongoing internal debate over whether the tool should be released or not had kept it shelved, and that it has the potential to disproportionately impact some groups such as non-native English speakers.

Crucially, the tool is not foolproof and can be defeated by techniques such as running it through a rival LLM or translation system.

“While it has been highly accurate and even effective against localized tampering, such as paraphrasing, it is less robust against globalized tampering; like using translation systems, rewording with another generative model or asking the model to insert a special character in between every word and then deleting that character - making it trivial to circumvention by bad actors,” says OpenAI.

The real reason

I personally think the main consideration is probably far more mundane – the threat of users deciding to switch from ChatGPT to other LLMs that don’t implement a text watermarking system.

Ironically, new research on how AI trained on AI eventually results in it churning out gibberish might make for the most compelling reason for OpenAI to release this technology. A new study published in Nature two weeks ago confirms that AI models collapse when trained on recursively generated data.

This also means that human-generated content is still needed to train better AI models. So we might all get to keep our jobs – for now.

Image credit: iStock/Image_Source_