Tech analyst Ben Thompson dismisses the 'clearly absurd' concept embedded in AI watermarking

Tech expert Ben Thompson says requiring AI companies to beryllium transparent astir the contented their models nutrient is missing the mark.

Thompson tore into the European Union AI Act, which Anthropic cited arsenic the logic it will add watermarks to contented successful Claude-generated files and text.

In his widely publication "Stratechery" newsletter, Thompson said specified AI transparency requirements are some overburdensome and improbable to beryllium beardown enough.

"I americium sympathetic to the impetus down watermarking: wouldn't it beryllium amended to cognize what is clone and what isn't?" he wrote. "What, though, is 'fake'? If an LLM states a existent fact, and a quality a fable, does it matter that the second doesn't person a watermark?"

Anthropic's plan, which it announced earlier this week, has profoundly divided the tech community. The AI elephantine said that Claude models launched aft August 2 would support watermarking from the commencement and it is moving to adhd specified support to older AI models.

Thompson took peculiar rumor pinch Anthropic adding a watermark moreover if Claude was utilized only for proofreading. The EU's regularisation allows for an editing exception, but technically speaking, Thompson said location was nary measurement for Anthropic to separate specified usage.

"I'm gladsome I ne'er developed the wont of copy-and-pasting proof-reading runs (I usage LLMs for proof-reading, but manually make each individual alteration successful my matter editor), but it hardly seems adjacent that anyone who wants to hole their grammar runs the consequence of having their original contented branded arsenic AI; the aforesaid vexation applies to translation," he wrote.

OpenAI, which presently watermarks images and audio, has besides pledged to grow the process to matter output, though the nonstop specifications of their scheme person not been released. OpenAI uses Google DeepMind's SynthID to watermark images and audio.

Regulators are turning to watermarks, traditionally seen successful ocular media, to thief nine discern erstwhile contented is AI-generated. Unlike ocular watermarks, the provenance markers for AI matter are not perceptible to the naked eye, but Anthropic and different AI companies person pledged to merchandise devices that will fundamentally service arsenic AI checkers.

Under the EU regulations, AI companies are required to guarantee "that the outputs of the AI strategy are marked successful a machine-readable format and detectable arsenic artificially generated aliases manipulated." Anthropic, OpenAI, Google, Meta, and Microsoft person each signed the applicable conception of the regulation. Notably, SpaceX has not. The transparency proviso went into effect connected August 2.

Anthropic did not bespeak that it would only enforce watermarking successful Europe. It remains to beryllium seen whether different companies will widen watermarking to users everywhere.

Overall, Thompson said watermarking unfairly credits AI pinch ideation erstwhile specified models are "(at slightest for now) a instrumentality that is wielded by humans."

"From this perspective, to insist connected watermarking is nary different than insisting that a ballpoint pen advertise itself arsenic the author, a conception that is intelligibly absurd," he wrote.