Hiding Flow’s mark is not unmarked
Turning off Gemini/Flow’s visible watermark leaves SynthID and C2PA in the file. Article 50 still wants a perceivable deployer label.
Hiding Flow’s visible watermark is not an unmarked file. On 14 August 2026 Google made the corner mark optional on Gemini and Flow. SynthID and Content Credentials stay. Article 50 still wants provider machine-readable marking and, for deployers of authentic-looking deepfakes, a label a person can perceive at first exposure.
TechCrunch, 14 August 2026 is the product source. Fastlane’s 24 August 2026 recap, citing that 14 August change, is a second dated source. This page is the legal and provenance meaning of “off.” It is not the toggle how-to, and it is not a DeepMind internals piece.
Visible chrome versus provider marking
Three layers get collapsed into “watermark.” They fail differently.
Visible chrome. A sparkle, a wordmark, a persistent overlay. A crop can remove it. A toggle can hide it. As of 14 August 2026, Gemini Settings → Media watermark and the matching Flow control do exactly that, except where Woodward said law requires the visible mark to remain. This is the layer clients complain about on a hero frame.
Invisible watermark. SynthID sits in the pixels or samples. Google applies it across the Gemini image and video families it named. A detector can still fire after the corner is empty. DeepMind’s own description of the scheme is that it is designed to stand up to ordinary crop, filter, frame-rate change, and lossy compression. You will not see it. You cannot inpaint the corner and call the file clean.
C2PA manifest. Content Credentials are a signed claim about one exact sequence of bytes. They travel until a tool re-encodes the file. Google said they stay embedded at generation whether the visible mark is on or off. They may die later, on a scheduler, an editor that is not C2PA-aware, or a platform ingest. Death of the manifest is not the same as “never marked.” It is a later hop.
Article 50 splits duties the same way the layers split. Providers of generative systems must machine-mark synthetic audio, image, video, and text so it is detectable. That obligation applied from 2 August 2026, with a grace period to 2 December 2026 for systems already on the market before 2 August, and watermark-detection interoperability running to 2 February 2027. SynthID plus C2PA is Google’s answer to that class of duty. Deployers who publish deepfakes — content resembling real people, places, or events that would appear authentic — must put a visible or audible label at first exposure. A hidden sparkle does not do that work.
The Commission’s transparency FAQ is the primary text (digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act). Icons cover fully AI-generated, AI-modified, and a basic-plus-interactive layer. Fines can reach €15 million or 3% of worldwide turnover. Those figures are a statute, not a Versely price.
Article 50 still wants a perceivable label
The user-facing mistake after 14 August is: “Google lets me ship without a mark, so I am disclosed.” You are not.
Provider marking is machine-readable. Deployer labelling is human-perceivable. Turning off Flow’s chrome removes a Google visible hint. It does not place your first-exposure label. Artistic, creative, satirical, or fictional works may label in a way that does not spoil the work — a credit, an end card, a mark that is perceivable without ruining a reveal. That is a placement courtesy in the FAQ, not a silence. Standard assistive editing is a different, actual exemption. Public-interest AI text is labelled unless a human exercised editorial control over the claims.
If the clip is a photoreal person, street, or event that would pass as real to an EU viewer, you still owe a line at first exposure. The disclosure template is the copy. Provider or deployer is the sorting post. You are almost always the deployer when you publish. Google was the provider when Flow generated.
Platform labels do not close the statute either. Instagram’s AI-generated profile badge (31 August 2026) is an identity control. Meta’s AI info on a Reel is a platform annotation. YouTube’s May 2026 overlay is YouTube’s. None of them is automatically the Article 50 first-exposure mark, and none of them fires because you left Flow’s sparkle on. Burn the line into the file if you need it to survive a mute autoplay session and a platform that draws its own chrome.
New York’s synthetic-performer statute (N.Y. GBL § 396-b, in force 9 June 2026) is a separate advertiser duty for digitally created humans who are not recognizable natural persons, with $1,000 / $5,000 civil penalties. It is not Flow’s toggle. Do not point a NY buyer at Gemini Settings.
C2PA can still die on upload
Keeping C2PA at generation is not the same as keeping it at distribution.
A hard-bound manifest hashes exact bytes. Instagram, Facebook, X, and most social transcodes change those bytes. ImgKey’s mid-2026 platform-policy snapshot is the practical map: X strips aggressively and shows nothing; Graph API and cross-post paths on Meta are less consistent than direct app uploads; LinkedIn is the network that most often preserves and displays Content Credentials. Do not assume the Flow export’s manifest is what a viewer’s downloaded file contains.
That is why “the file is marked” and “the posted copy is marked” are different sentences. SynthID is the layer built to survive the transcode. C2PA is the layer that often does not. Visible chrome is the layer you just hid. If your compliance story is “C2PA will tell Instagram,” you need a path that still has a manifest at ingest, and you still need self-disclosure when it does not. If your compliance story is “no visible mark, so we are fine,” you have dropped the only layer a person could see without a detector.
Archive the generator’s export. That is the copy whose manifest, if any, still hashes. Run a pre-publish check on the finished, captioned file — ratio, safe-area, captions, SKU spelling, and the perceivable label — not on the Flow preview.
What “unmarked” would actually mean
Unmarked, in production language, would mean:
- no visible chrome
- no invisible watermark a detector can read
- no C2PA or IPTC generative signal
- no platform self-disclosure
- no perceivable deployer label
Google’s 14 August change gives you the first bullet inside Gemini and Flow, sometimes. It does not give you the rest. Woodward said you can still ask Gemini or Search whether an image was AI-generated. That sentence is Google admitting the file is still marked.
Versely does not watermark outputs on any plan as a vendor chrome. That is a different fact from “the model provider skipped SynthID.” Veo and Omni-class outputs can still carry Google’s invisible mark because they came through Google. Catalog credits — Veo 3.1 at 40, and the rest of the board — do not buy an unmarked pixel. They buy a generate.
If you need the clip, the captions, and a first-exposure line in one place, sign up and treat Flow’s optional sparkle as furniture. Provenance and the deployer label are still on you.
FAQ
If Flow’s corner is empty, is the clip unmarked for Article 50?
No. Article 50’s provider duty is machine-readable marking; SynthID and C2PA remain. The deployer duty is a perceivable label at first exposure for authentic-looking deepfakes. An empty corner is neither “unmarked” nor that label.
Does hiding the mark strip Content Credentials?
Google says C2PA stays embedded at generation. A later editor, scheduler, or platform transcode can still invalidate the manifest. Check the file you will upload, not the Flow preview.
Can I tell a client the deliverable has no AI mark?
Not honestly, if it came through Gemini or Flow after 14 August 2026. You can say the visible Google chrome is off (where the toggle is allowed). You cannot say SynthID is off. You cannot skip a legal first-exposure label because the sparkle is gone.