The watermark you can turn off

Mid-August 2026 split the watermark story in two. Anthropic stamped Claude text with an invisible, machine-readable mark users cannot toggle. Google answered a top Gemini complaint by letting people hide the corner "spark" on images, video, and music — while keeping SynthID and C2PA in the file. Same week, same word, opposite levers. The fight is not whether AI outputs get labeled. It is which label you are allowed to peel off.

Google: visible optional, invisible mandatory

On August 14 Google rolled out a Media watermark setting in Gemini and Flow. Turn it off and the spark vanishes from Nano Banana images, Omni video, and Lyria audio. Turn it on (still the default) and the corner brand stays. Josh Woodward framed the move as creative control plus safety: visible marks become optional; invisible SynthID watermarks and C2PA metadata remain. You can still ask Gemini or Search whether something was AI-generated. The toggle will not ship in countries that legally require a visible AI label, and Google says Search is next. Credentio, a new open-source library, is meant to help developers embed local validation — provenance without relying on a corner logo.

That is not a Google-only pattern. OpenAI and Meta already lean on invisible provenance more than stamped branding; OpenAI pairs C2PA with SynthID and ships verify tooling rather than forcing a permanent spark on every download. Users have been cropping corner marks for years. Google is catching up to the obvious: a logo in the frame is a product complaint; a machine-readable signal is the compliance layer regulators actually asked for.

Claude: the mark you cannot turn off

Anthropic's docs still say models released on or after August 2, 2026 watermark generated text and use C2PA for supported files, across Claude surfaces and cloud partners. A hit means Claude may have processed the content; a miss does not prove a human wrote it. Article 50 and the Transparency Code require machine-readable marks where feasible — new systems around August 2, older ones by December 2, interoperable detection by February 2, 2027. Google, Meta, Microsoft, OpenAI, Mistral and others signed the same Code. Anthropic took the Brussels Effect path: one global behavior, not an EU-only switch — and no settings menu that makes the text mark go away.

Invisible ink, real tradeoffs

Text watermarks (SynthID-Text and cousins) nudge which token the model picks next so a provider can later detect a statistical pattern. Stronger marks detect better and can cost naturalness; softer ones paraphrase and multi-model pipelines erase. Media SynthID plus C2PA follow the same idea for pixels and audio — Google and OpenAI both ship variants — but detectors still do not cross-read cleanly. Fragmentation is the default until interoperability is forced. Google's spark toggle does not fix that. It only clarifies that the durable fight was never the logo.

Anger, irony, and brittle marks

Some users fear school and workplace detectors; others say the only reason to hate a mark is wanting to pass machine text off as unaided work. Both miss the middle: legitimate outline, polish, and scaffolding workflows still get stamped once detectors are one-click. Policy critics call Anthropic's global text rollout a quiet quality tax and warn a Claude hit can smear human-led reporting as "fabricated." Creators who cheered Google's spark toggle are not asking for unmarked deepfakes — they are asking for slides and ads that do not look like a demo. The training-data irony still lands: models trained on the open web now mark the people who use them. Research keeps showing the same limit: paraphrase and scrubbing beat text marks; stripping metadata does not kill SynthID; absence never proves authenticity; presence never proves no human judgment remained.

What directory readers should watch

Separate product branding from provenance. Prefer vendors that publish detection confidence, editing survival, and false-positive handling — not a binary badge for HR and classrooms, and not a corner logo that disappears the moment a setting flips. Ask whether marks are cross-readable before 2027 or locked to one lab's keys, whether visible labels are required in your market, and whether open-weight stacks stay unmarked escape hatches. Our Ethics page and Regulatory Tech map the compliance layer; see also who sees the model first and AI slop.

The point

In the same week, Google let users hide the visible watermark on Gemini images, video, and music while keeping the invisible one, and Anthropic added an invisible watermark to Claude text that users can't turn off. Visible labels are becoming optional while hidden ones become mandatory. A watermark can suggest that something was made with AI, but it can't prove it, so don't treat one as the final word.