ChatGPT text will now include hidden signals that allow users to check where it came from, OpenAI has said.
The company is rolling out “text watermarking” that subtly changes the wording of output in an attempt to show that it was generated with artificial intelligence. The changes are intended to be invisible to human readers, but can be picked up by automated systems.
It comes after Anthropic said that it would add similar technology to Claude. That prompted outcry among some users, who claimed that it would be unreliable and represented an unnecessary interference with how the system works.
Both OpenAI and Anthropic made the change in keeping with the EU’s AI Act, which requires companies making generative AI products to make their outputs identifiable by automated systems. Unlike Anthropic, OpenAI is only adding the watermark in the European Union.
OpenAI calls its own system textGrain. It adds what OpenAI calls an “invisible statistical signal” to the way that ChatGPT chooses words.
It warned that the tool is still unreliable, falsely identifying human text as automated and vice versa. That is especially true when text is short or has been edited, it said.
The company stressed that it did not believe that the watermarking would reduce the quality of ChatGPT’s responses, and pointed to data that showed that it actually performed better than unwatermarked text on some evaluations. After Anthropic announced its similar system, it drew criticism from some users who speculated that it would mean that Claude’s outputs would be less good because of the watermarks.
OpenAI said that it would initially only allow access to its detector software to “approved researchers and expert organizations”. That is partly because the system is still not reliable, and it said it would work with those organisations to evaluate its performance in real-world use.
The company also stressed that watermarks cannot give an indication of how much a human contributed to the text, that it does not say anything about who owns it, will not identify the user who created it and that it is no indication of the accuracy of a given output. It also stressed that the absence of a watermark is not any indication that it was produced by a human.
As automatically generated content has spread across the internet, regulators and companies have scrambled for ways to check whether a piece of text, image or other media has been made by an AI system or a human. In addition to watermarks, some companies have pursued other strategies – including Apple, which last months launched a new iPhone that includes technology intended to allow photographers to prove that they actually took a photo.
Read MoreWikipedia blames rogue OpenAI bots for rare outage
Ex-Anthropic whistleblower sounds the alarm about dangerous AI companies at hearing
OpenAI CEO sends chilling warning: ‘World should accept some bad things happening’
