Anthropic just started marking everything Claude writes. Here's what actually changed, what it can and can't prove, and the five-minute check to run before your next post goes live.
If your team uses Claude to draft social captions, blog posts or ad copy, something changed this week. On 11 August 2026, Anthropic published how it marks AI-generated content, and the short version is this. Claude now weaves an invisible watermark into the text it generates, and attaches signed provenance data to the image files it makes.
You won't see it. Your customers won't see it. But it's there, and it's worth understanding before someone runs your website copy through a detector and draws the wrong conclusion.
What Anthropic actually announced
Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on transparency of AI-generated content. That's a European rule, but the marking isn't limited to Europe.
The details worth knowing:
- →Claude models launched on or after 2 August 2026 support marking from day one. Older models are being backfilled.
- →Marking applies wherever Claude is offered, worldwide. Australia included.
- →It happens at the model level, so it doesn't matter which Claude product your team uses. The mark comes with the output.
- →Detection tools for the public are still coming. Anthropic says it will share technical documentation, but there's no public checker yet.
That last point matters. Right now there's a mark on your content that almost nobody can read.
Two different marks doing two different jobs
Text gets an invisible watermark. It's built into the words themselves, not bolted on afterwards. Anthropic says it doesn't change the meaning or readability of the output, it travels when text is copied and pasted, and it may survive some editing.
Ready to grow?
Book a free strategy session with a local specialist.
Files get signed provenance metadata. For supported file types like .png, .jpg and .svg, Claude attaches metadata using the C2PA open standard. It confirms a file was processed by Claude and flags whether the file has been tampered with since.
The two behave very differently in the real world. Text watermarks stick around through a copy-paste into your CMS. File metadata is fragile. Screenshot the image, convert the format, re-save it, or upload it somewhere that rewrites metadata on the way in, and the signature is gone.
A detected mark is a signal, not a verdict
This is the part most people will get wrong, and it's the part worth explaining to your staff before someone gets accused of something.
Anthropic is unusually clear about the limits, in both directions.
A mark can show up on work that isn't AI-written. People use Claude to proofread, translate and summarise material that already existed. If a Brisbane sparkie writes a service page from their own job notes and asks Claude to tidy the grammar, the output can carry a mark. The ideas, the pricing, the experience, all human. The mark doesn't know the difference.
No mark doesn't mean no AI. Content can come back clean if it was made by an older model, heavily edited or paraphrased, translated, mixed into other writing, or if the passage is simply too short to give a reliable signal. Strip a file's metadata and the provenance goes with it.
A watermark can't tell you who wrote something, whether it's true, or who owns it.
Treat it as one clue among several, and be very careful about treating anyone else's detection result as proof.
The Australian angle
Australia doesn't have an equivalent of the EU's labelling law. There's no local rule saying you must disclose AI use on a Facebook post about a tyre sale.
What hasn't changed is the rule that was always going to catch you. Under Australian Consumer Law, misleading or deceptive conduct is misleading or deceptive conduct, no matter what tool produced it. An AI-generated photo of a breakfast your cafe doesn't serve is a problem because it misleads customers, not because of anything hidden in the file. If you're unsure about disclosure obligations in a regulated industry like finance, health or property, get current legal advice rather than guessing.
The practical risk for most small businesses isn't a regulator. It's a client, a competitor or a job applicant running your content through a detector once those tools land, and reaching a conclusion you then have to unpick. If AI copy has crept across your website or your Google Business Profile, that's where to start checking.
Four rules to give your team
Short enough for an apprentice to remember.
- →Strip the personal details first. Names, addresses, phone numbers, registration plates, medical details. A mobile detailer in Penrith tidying a follow-up message should paste in the message, not the customer file.
- →Check every fact a customer relies on. Prices, trading hours, suburbs served, licence numbers, product names, offers and expiry dates. AI invents these with total confidence. One wrong Saturday trading hour costs a salon real bookings.
- →Real photos for real things. Your people, your premises, your finished jobs, your menu. Save AI images for concept illustrations and generic backgrounds, and label them as concepts when there's any chance of confusion.
- →A named person approves it. Not “marketing”. A person, whose name goes in the folder next to the final file.
Keep a light paper trail
You don't need a register for every caption. You do need one for content that makes a promise, gives advice or represents the business publicly.
For a finance broker's home loan ad, keep the final version, the source notes, the images used and who approved it. Filed by date is enough for a small team. For a Newcastle landscaper, keep before-and-after job photos in a separate folder from AI concept designs, so an imagined backyard never gets mistaken for completed work.
The five-minute check before you publish
- →Is every claim accurate for your actual business, including price, location and availability?
- →Does the image show something real, or is it clearly presented as a concept?
- →Have you removed customer details and confidential business information?
- →Can you point to a source for every promise, testimonial or result?
- →Has a named person read it and approved it?
- →Would a customer feel dudded if they found out AI helped write it?
If the last one gives you pause, that's your answer. Rewrite it.
Treat AI copy as a rough draft, not finished marketing
AI writes like it's never visited your suburb. It produces lines that sound fine right up until every competitor in town is saying the same thing.
Add what only you know. A Darwin air-conditioning business can talk about wet season call-outs, the brands it actually services, and answering the phone after hours in February. Then read it aloud. If it doesn't sound like something you'd say to a customer across the counter, it isn't ready.
The takeaway
Watermarks and provenance data will make AI content easier to spot, and harder to spot reliably, at the same time. They're clues. They're not proof of authorship, ownership or truth, and Anthropic says so plainly.
Use Claude to get a first draft moving. Keep a human in charge of the facts, the photos, the approvals and the customer trust. No watermark is going to do that part for you. If you'd rather hand the drafting to people who do this every day, our content and SEO team can help.
Source: Anthropic, “How Claude marks AI-generated content”, updated 11 August 2026. Accurate at the time of writing. This is general information, not legal advice.