OpenAI Adds Invisible Watermark to ChatGPT and Codex Text
OpenAI says the machine-readable marker, opt-in worldwide starting today, will become automatic and mandatory for EU users in the coming weeks.
OpenAI announced it will insert an invisible, machine-detectable watermark into all text generated by its AI models, including ChatGPT and Codex. The feature is opt-in globally starting today, according to gizmodo.com.
In the coming weeks, the watermark will be applied automatically, with no off switch, to eligible ChatGPT and Codex text output in the European Union. OpenAI says the change brings it into compliance with the EU AI Act, which requires model-generated text to carry a machine-readable identifier.
Access to OpenAI's watermark detector will be granted to researchers and expert organizations on a case-by-case basis, not to the public at launch, because of false-positive risk, the company said.
Other labs and public opinion
Anthropic launched a similar machine-readable marker in August, which prompted some users to look for ways to scrub the indicator from their text. Google DeepMind has been watermarking its outputs for a couple of years, but in August announced that users could remove watermarks from images, videos and songs in countries where it is not required.
A recent Rainey Center poll found that nearly eight in 10 people support AI watermarking requirements, including 82% of Democrats and 75% of Republicans.
Quick answers
Which OpenAI products will carry the watermark?
Text generated by OpenAI's AI models, including ChatGPT and Codex.
Can users in the European Union turn the watermark off?
No. In the coming weeks it will be applied automatically and without an off switch to eligible ChatGPT and Codex text output in the EU.
Who can use OpenAI's watermark detector?
Researchers and expert organizations, granted access on a case-by-case basis. It is not available to the public at launch, due to false-positive risk.