Invisible watermarks will now be included in any content generated by ChatGPT throughout the EU as part of an effort by OpenAI to ensure compliance with Article 50 of the new EU AI Act.
The ChatGPT invisible watermarker can also be enabled in all apps using the ChatGPT API from anywhere around the world. However it remains disabled until it is enabled.
As a means to avoid any misunderstandings regarding who the author of some piece of text is, OpenAI has decided to provide its verification method only to accredited researchers.
How the Invisible Watermark Fits into ChatGPT Text?
OpenAI calls this hidden marking system textGrain. Instead of changing how sentences sound or forcing weird phrasing it subtly picks specific word choices behind the scenes to leave a light pattern.
-
How it works? It tweaks word choices using math patterns that human readers cannot spot.
Internal tests show the quality of answers stays identical whether the mark is turned on or off.
App builders outside the EU can choose whether they want to turn this feature on for their projects.
How Well the Detector Spots Marked ChatGPT Text
How easy it is to spot an invisible ChatGPT watermark depends mostly on how long the text is. It also depends on what topic it covers and whether someone rephrased it later.
|
ChatGPT Testing Scenario
|
200 Words
(Tokens)
|
400 Words
(Tokens)
|
|
Psychology Topics (Flexible Words)
|
~80% Found
|
~95% Found
|
|
Math Topics (Fixed Words)
|
Hard to Spot
|
~60% Found
|
|
Swapping 10% of Words with Synonyms
|
Signal Weakened
|
66% Found
|
|
Swapping 25% of Words with Synonyms
|
Signal Lost
|
17% Found
|
Source: Official OpenAI Technical Evaluation Data
European Rules for AI Content and What Happens Next
This move ensures that OpenAI complies with the European regulations mandating labeling of the AI generated content. The company is proceeding in a conservative and region specific manner and is not implementing the feature on a global scale at this point in time.
The watermark does not indicate the name of the person who provided the prompt. and it also doesn’t tell if there was a human editing of the final document.
Given that the AI watermark can be erased during editing and that short pieces of writing don’t carry enough data for detection. OpenAI is planning to keep its detector technology confidential for the time being.