OpenAI starts placing invisible watermarks on texts under the EU's Artificial Intelligence Law

Artificial intelligence company OpenAI is starting to place invisible watermarks on the texts it produces, following Anthropic, which has taken the same step in recent months in order to comply with transparency obligations under the European Union (EU) Artificial Intelligence Law.

In its statement, the company stated that starting from the coming weeks, it will add invisible statistical marks called "textGrain" to the text outputs of ChatGPT and Codex users within the EU borders.

Global API customers will be able to optionally enable this system starting today for certain models, but the feature will remain off by default on the API.

The company is also working to include this feature in model outputs accessed through cloud partners.

AS WORDS CHANGE, THE POWER OF DETECTION BECOMES WEAK

According to the announced technical data, textGrain technology places statistical signals imperceptible to the human eye in the word choices of the model.

However, the system shows obvious weaknesses in the face of interventions made on the text.

In the test results shared by the company, with a targeted false positive rate of 1 percent, the detection rate, which was 95 percent in 400-token texts suitable for flexible word selection such as psychology, decreased to approximately 80 percent in 200-token texts.

Detection rates remained significantly lower in fields where word options were limited, such as mathematics.

Text edits also rapidly reduce the strength of the signal. When 10 percent of the words in a text of 400 tokens are replaced with synonyms, the detection rate decreases from 92 percent to 66 percent. When 25 percent of the words are changed, the detection rate drops to 17 percent.

OpenAI noted that it would not directly release the detection tool to the public due to these technological limitations and the risks of producing both false positive and false negative results.

Access to the verification tool will initially be provided only through application to approved researchers and expert bodies to evaluate and improve the system.

The company announced that in the tests carried out on its latest flagship model, Astra, the watermark application did not create a significant performance difference in output quality.

According to various benchmark tests, the system; In the artificial intelligence analysis index, it received 49.57 points in the unwatermarked text and 49.76 points in the watermarked text.

Similarly, in the GPQA Diamond test, the result without watermark was 94.44 percent, while the result with watermark was measured as 93.94 percent.

OpenAI also made detailed warnings about what the watermark can and cannot show.

According to the statement, invisible marks do not measure the proportion of human contribution in the text, do not assign copyright or responsibility, do not reveal the user, account or entry that produced the text, and do not verify the accuracy of the content.

The absence of a watermark in a text does not necessarily prove that that content was written by a human; Because the text is short, translated, edited, or taken from an unsupported model, it may prevent detection.

The company stated that open verification tools based on SynthID and C2PA standards used in visual and audio content will continue to remain publicly available.