Anthropic provides an official explanation on 'digital watermarks that indicate text was generated by AI.'
To comply with EU law, Anthropic will be adding watermarks to content generated by Claude. Anthropic has explained how it will add watermarks to text.
How Claude's text watermarking works \ Anthropic
The watermark in the text is not something that can be seen with the naked eye, but rather a pattern of word selection. Technically, Google DeepMind's ' SynthID Text ' is used.
In the first place, large-scale language models like Claude select the most appropriate word from multiple word candidates in order when generating text. For example, if the word candidates following the sentence 'Today's lunch is fried chicken and...' are 'onigiri,' 'omusubi,' and 'ramen,' the probability of ramen being selected is low, and in most cases, the first two will be judged as more appropriate. The meaning of the sentence itself does not change regardless of which of the first two is chosen, and the selection is determined by a random number generator.
When adding a watermark to text, the word is selected using a pre-existing word and a special 'key' (selection tendency), rather than a random number generator. While the words selected by the AI are still random, if a pattern of words matching the key appears repeatedly in the text, it can be determined that the text was likely generated by the AI.
However, it's important to note that this is merely a matter of possibility, and the detection of a watermark doesn't necessarily mean it's AI-made. Conversely, the absence of a watermark doesn't necessarily mean it's not AI-made.
The mechanism for detecting watermarks will soon be available as a 'Watermark Detection API.' This API detects keys managed by Anthropic from text, and its mechanism is fundamentally different from external tools that simply detect AI-like word patterns.
Furthermore, it appears that these watermarks do not affect the quality of the output. Anthropic's internal tests have reported no impact on the content, level of creativity, or readability of the content. Processing speed is negligible, and no additional tokens are consumed.
Anthropic's watermarking has several limitations. In addition to the aforementioned possibilities, it cannot identify content generated by AIs other than Claude because Anthropic does not know the keys used by those AIs. It also has difficulty identifying short sentences with limited word selection patterns, factual texts, and texts where only simple typographical errors have been corrected by AI.
Furthermore, detection can be avoided by having humans edit the text generated by the AI. Anthropic pointed out that 'minor edits may not completely remove the watermark, but a complete rewrite that replaces every word can remove it. Of course, in the latter case, there is room for debate as to whether the text can still be called AI-generated.'
As a disclaimer, Anthropic stated, 'The watermark only indicates that Claude likely had some involvement with the content. It does not distinguish whether Claude wrote it or heavily edited it. Also, if Claude translates the text, the watermark will be clearer because he selects every single word.'
Related Posts:
in AI, Posted by log1p_kr