Anthropic’s Foolish Strategy: Watermarking AI Text
In a bid to “watermark” the text produced by Claude and adhere to the EU AI Act, Anthropic has introduced a plan that’s nearly as practical as a chocolate kettle. They’re adjusting their bots’ vocabulary so it’s distinguishable as AI-generated. Conventional watermarks are akin to designs on me mates’ banknotes, postage, or official papers that shout authenticity. However, in the digital domain, it offers a bit more leeway, doesn’t it? Anthropic’s method entails tweaking their models’ vocabulary, a tactic borrowed from Google DeepMind’s SynthID-Text research.
Understand It, Claude: Predictive Text Clarified
To put it simply, large language models are like that buddy who consistently tries to complete your sentences. When Claude types “The weather today was cold and…”, you might anticipate words such as “cold” or “gray,” but Claude’s feeling a bit artistic: “…crisp, the type of cold that bites at your fingertips and turns your breath into tiny puffs.” A charming fellow, but strip away the embellishments, and Claude is merely guessing the subsequent word. Anthropic asserts that in most instances, the sentence could conclude with “cold” or “gray,” and the meaning remains unchanged regardless.
Watermarking: Merely Random Nonsense
The watermark is produced by deviating from the anticipated word. They introduce randomness that can be detected with a digital key. DeepMind researchers explain “Generative watermarking” modifies text to embed subtle alterations, which leaves a statistical marker. When identified, it shows AI was involved. Anthropic claims this won’t change the meaning, asserting “no effect on creativity or readability.” However, they are relying on not watermarking crucial text. It’s somewhat like swapping out spice jars without altering your dish, isn’t it?
Confusion in Coding: Method Name Missteps
Within the coding sphere, watermarking can’t just indiscriminately change method names. Just picture if Claude opted to be all avant-garde in literature – “It was the best of times, it was the least of times…” Not particularly enjoyable, right? Serious authors might not take kindly to their work being subjected to AI treatment.
It’s Not the Apocalypse
Fortunately for us, Anthropic’s watermarking isn’t oppressive. It doesn’t insert any personal information and merely signifies Claude’s involvement in the text. Additionally, it’s somewhat effective. Some modifications may eliminate the watermark, but thorough rewrites will completely remove it. They’re not concerned if people circumvent the system. They’re just checking off a legal compliance requirement without increasing expenses. “Watermarking doesn’t hinder the models or inflate costs,” they state. Hey Claude, what’s another term for performative compliance? ®
Conclusion: Watermarking Antics
If you ask me, this watermarking is somewhat of a nuisance. Anthropic’s motivation stems from legal checkbox fulfillment, not from enhancing their AI. For a feature that isn’t going to intrude significantly or incur additional costs, it’s akin to spraying graffiti on your own shed. It might look interesting, but it won’t prevent the roof from leaking.
