Anthropic’s AI watermark push is already colliding with tools built to erase it

Anthropic’s AI watermark push is already colliding with tools built to erase it
Anthropic’s effort to make AI-written text easier to identify has ignited a familiar technology arms race: within days, developers were building tools designed to make its invisible labels vanish.

The company began embedding imperceptible watermarks in text generated by supported Claude models released from August 2, saying the statistical pattern in word choices would persist when text was copied and pasted. Anthropic has framed the measure as part of its commitments under the EU AI Act, though it has also said the watermark is not intended to establish authorship.

The reaction came quickly. Paris entrepreneur Guillaume Meyer released his open-source “Watermarks Remover” project after the announcement; he says its first version took roughly five hours to build. The tool rewrites text while preserving its meaning, aiming to disrupt the patterns detectors rely on. After an August 11 post, Meyer said the project drew more than 2 million impressions and turned from a curiosity into a full-time effort.

Meyer’s argument is not that attribution is pointless, but that this particular mechanism is too blunt. “I am all for content attribution,” he said. “I am against the watermarking technique, and that’s a very significant distinction.” He warned that people using AI only to proofread, translate or make a minor edit could nevertheless find their work marked—an especially fraught prospect for non-native English speakers relying on grammar tools.

Critics of the removers see a different danger: labels can help audiences distinguish synthetic content from human work. Yet researchers and legal specialists quoted by Business Insider say the technology has an inherent weakness. Rewording can remove a watermark, while the EU rules place durability obligations chiefly on AI providers and do not expressly ban third parties from creating removal tools.

That leaves the sharpest line at use, not code. Anthropic’s policy prohibits passing model output off as human-generated; making a remover may be lawful, but using one to deceive could be another matter.

Continue reading https://foxvector.com/stories/01a02fab-9745-27dd-73c8-34d29d04c13e