The abundance of AI-generated content has led technology companies to develop methods to inserting invisible watermarks in images and videos that identify their synthetic origin, but doing so in texts is a greater technical challenge. Although there are various tools on the market that claim to detect texts generated by AI, the truth is that their results are usually unreliable. Now it is OpenAI that begins to introduce watermarks in the texts generated by ChatGPT, but only within the European Union.
The reason this is so is the AI Law European Union, which requires providers of generative AI models to make it possible to identify content generated by artificial intelligence through machine-readable systems. Starting today, OpenAI API customers will be able to activate watermarks for certain models, and over the next few weeks, the company will start automatically adding invisible watermarks to texts generated by ChatGPT and Codex within the EU.
How textGrain works
The ChatGPT watermark will not add hidden characters, unusual punctuation marks, invisible spaces, or additional tokens to the text. Instead, OpenAI’s textGrain technique subtly modifies the statistical pattern of the words and tokens selected by the model.
When an AI generates text, it assigns different probabilities to the possible tokens that may appear next; These are small units into which you divide the text and which can correspond to a word, a part of it, or a punctuation mark. TextGrain slightly modifies those probabilities in a secret pattern to encourage the model to select certain alternatives. It is a similar idea to the one advanced Anthropic for Claude in August.
OpenAI claims that its textGrain technique matches or exceeds the results of other methods it has tested, including SynthID for text Google. The company will also make this technology available to researchers and other organizations so that they can develop tools to detect texts generated by artificial intelligence.
OpenAI explains how to fool AI detection in texts
OpenAI itself recognizes, however, that watermarks in texts They are much less reliable than those used in images or audio. The shorter a text is, the more difficult it is to determine if it has been generated by an AI.
According to the company’s internal testing, its detector was able to identify around the 80% of the watermarked fragments of 200 tokens and approximately 95% of the fragments of 400 tokens.
But even small modifications can significantly weaken the watermark. In tests carried out by OpenAI, Replacing only 10% of the words with different ones reduced the detection rate from approximately 92% to 66%. By substituting 25% of the words, the rate dropped to just 17%..
To prevent false positives from causing unjustified accusations, OpenAI warns that Detecting a watermark should not be considered proof that someone created text with ChatGPT. A flag may indicate that an OpenAI model generated or processed a portion of the content, but It does not allow us to determine how much the user subsequently modified it or what proportion of the final result corresponds to human work..