Another newsletter about artificial intelligence — I'm sorry, because I suspect many people are experiencing a kind of “AI fatigue”, given the astonishing and sometimes worrying innovations that arrive almost weekly, often accompanied by articles about the amazing or terrible future ahead of us.
This is perhaps the fourth or fifth issue dedicated to AI. The indecision comes from the fact that I always try to enlarge the focus, framing the “AI revolution” — spoiler alert: I think it's a big change, but I'm not sure it's a revolution that subverts our world — within topics like our moral landscape, our ideas about authorship, or how we know things.
What topic do I wrestle with this time? My intention was to talk about the “AI watermark” that Anthropic is deploying in its models: an invisible trace on text generated by Claude. That was the intention, at least: as usual, I ended up reasoning about something more general, and the watermark turned out to be the way in rather than the destination. Other AI detectors work by spotting the quirks of LLMs — the large language models that are the engines behind popular apps like ChatGPT, Gemini or, indeed, Claude — and are not reliable; from direct experience, they detect unedited texts generated with a vanilla prompt and little else. This watermark, similar to the SynthID included in Google's models, is based on a hidden bias towards certain words. Roughly explained: let's say the LLM is generating the description of the sky on a cloudy day and has to choose between “gray”, “dark”, “overcast”, “leaden”, “slate”. The shortlist itself is not arbitrary: meaning and style have already ruled out everything that doesn't fit. But once the field is narrowed to those few plausible candidates, in a normal situation the pick among them is random, like rolling a fair die; with the watermark the die is loaded, and certain results become slightly more frequent than others. If you know how the die is loaded, and if you have enough words to analyse, you can detect the mark. More information, with interactive examples, here.
I consider this a bad decision, for two reasons. But first, the reason I don't share: I don't fear that the texts I publish, here or elsewhere, will test positive for this watermark. I regularly use LLMs to check my writing, to complete some paragraphs, to find examples, to add trivial information I don't want to retrieve while I'm writing, like a book title or a publisher. I use LLMs to correct my English, which is not my mother tongue and which I pack with odd expressions and grammatical errors, but also my Italian that often has overcomplicated sentences. I also use AI for fact-checking, asking it to highlight all errors and inaccuracies (usually working from external sources rather than the model's internal knowledge). It's possible that these operations add the watermark to my texts, and I'm ready to take the blame: what I publish is what I think, maybe in a clearer and more pleasant form thanks to an LLM.
An AI Watermark Won't Tell You What to Trust
