The EU AI Act's transparency obligations took effect on 2 August 2026. Synthetic text, images, video and audio now have to be marked so that it can be detected as artificially generated.
If you are an Australian marketer using AI tools, the short answer is that most of this obligation does not fall on you. It falls on the companies that make the tools. Here is where the line actually sits, and the much smaller set of things you do need to think about.
Who the Marking Obligation Falls On
Article 50 of the AI Act splits its duties between two groups, and conflating them is the source of nearly all the confusion.
Providers are the organisations that develop an AI system or place it on the market. Under Article 50(2), providers of generative AI systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated. This is a technical provenance obligation, and it is the heavy one. OpenAI, Google, Anthropic, Meta, Microsoft and Mistral are all providers.
Deployers are organisations using an AI system in a professional capacity. That is you. Under Article 50(4), deployers must disclose deepfakes, and must label AI-generated or manipulated text published to inform the public on matters of public interest.
Note what the deployer obligation does not say. It does not require you to label every piece of marketing copy an AI helped you draft. It targets content that could mislead about a real person or event, and text published on matters of public interest. Ordinary commercial content is not the target of Article 50(4).
There is also an explicit carve-out at the provider level: the marking obligation does not apply where the AI system performs only an assistive function for standard editing, such as grammar correction, or where it does not substantially alter the input data or its semantics.
⚖️ Not legal advice. This is a marketer's read of a technical regulation, written to help you ask your adviser the right questions. If your business has genuine EU exposure, get an opinion from someone qualified to give one.
The Dates That Matter
- 1 August 2024: the AI Act entered into force, with provisions phasing in over several years.
- 20 July 2026: the European Commission published its finalised guidelines on implementing Article 50, alongside the Code of Practice on Transparency of AI-Generated Content.
- 2 August 2026: Article 50 transparency obligations became applicable.
- 2 December 2026: end of a limited deferral for generative systems already on the market before 2 August 2026.
- 2 February 2027: watermark-detection interoperability required.
Penalties for non-compliance with Article 50 reach up to 15 million euro or 3% of total global annual turnover, whichever is higher. Deliberately removing or altering an AI watermark is prohibited.
Does This Apply to an Australian Business?
The Act applies to organisations placing AI systems on the EU market, or using AI in a professional context within the EU, regardless of where the organisation is headquartered. So the test is exposure to the EU, not where your ABN is registered.
For a Sydney business selling to Australian customers only, direct legal exposure is limited. For a business with EU customers, an EU entity, or content that materially targets EU audiences, it is worth a proper look.
The more universal effect is indirect. The platforms you use every day are complying globally rather than building region-specific behaviour, which means the outputs of your tools change whether or not the law reaches you.
What the AI Companies Have Actually Done
Around 190 organisations had signed the Code of Practice on Transparency of AI-Generated Content by the end of July 2026, including Google, Meta, Microsoft, Mistral, OpenAI and Anthropic on the provider section covering machine-readable marking and detection.
Anthropic's implementation is the most publicly documented, and it is worth understanding precisely because it is often described inaccurately.
Text generated by Claude carries an imperceptible embedded watermark, and generated files carry digitally signed provenance metadata using the C2PA open standard. Marking applies globally, not only to EU traffic, which Anthropic attributes to not yet having a durable way to scope it by region. Coverage extends across its products, including the consumer app, the API, Claude Code and Claude Cowork, and through cloud partners.
Three corrections to the version circulating in marketing newsletters:
- It is not all Claude output. Coverage is by model. Models launched on or after 2 August 2026 support marking at launch. Earlier models are being addressed during the transition period the law allows.
- Detection is not public. The detection API is in private preview, available to categories of organisation the law requires, such as regulators, law enforcement, media, fact-checkers, researchers, educational organisations and EU civil society groups, plus enterprises with their own compliance obligations. You cannot currently check a document yourself.
- A mark is not proof. Anthropic states that a detected mark indicates Claude may have processed the content, not that Claude authored it, and that the absence of a mark does not establish that no AI was involved. Marks can also appear on human writing that Claude only edited or translated.
File-level provenance is more fragile than text watermarking. Format conversion, re-saving and screenshots can all strip C2PA metadata, so a file that once carried a mark may not after editing.
The Technical Standard Is Not Settled
Worth knowing before you plan around this: the Commission has stated that it has not identified any single marking or labelling technique that currently meets the Act's standard. The expected approach is multi-layered, combining digitally signed metadata with imperceptible watermarking, and optionally fingerprinting or logging as a fallback for short or heavily transformed outputs.
An EU icon for labelling and the use of open standards are being promoted to simplify compliance. In practice, the detection infrastructure that would make any of this enforceable at scale is still being built.
Will Google Use Watermarks to Demote AI Content?
This is the question we get asked most, and the honest answer is that there is no evidence it does.
Google's published position has consistently been that it rewards helpful, original content regardless of how it was produced, while treating scaled content abuse as spam. The distinction is between content produced at scale primarily to manipulate rankings and content that genuinely serves a reader. Production method is not itself the signal. We covered where that line sits in our piece on Google's AI search spam policy update.
Could that change? Machine-readable provenance data becoming widely available is a new capability, and it would be naive to assume no search engine will ever use it. But building your content strategy around a demotion that has not been announced, using a detection standard that regulators say does not yet meet its own bar, is speculation dressed as risk management.
What to Actually Do
None of this requires a compliance programme for a typical Australian business. It does justify a handful of habits that are good practice regardless:
- Name a human editorial owner for every published piece. Someone who has read it, checked the claims and is accountable for it. This is the single most valuable thing on this list and it has nothing to do with the EU.
- Keep a record of which tool produced which asset. A column in a content calendar is enough. If disclosure obligations arrive in Australia, reconstructing this later is painful.
- Disclose where you are depicting a real person or event. This is the deepfake category, and it is the one part of the deployer obligation that maps cleanly onto marketing work, particularly synthetic voice, likeness or video.
- Do not strip provenance metadata from generated files. Deliberate removal is prohibited under the Act, and there is no upside to it.
- Verify facts and figures independently. The reason to have a human in the loop is accuracy, not compliance theatre. Our AI content strategy guide covers how we structure editorial review.
Where Australia Sits
Australia has no equivalent AI content labelling law in force. The EU has a track record of setting templates that other jurisdictions adapt, and privacy regulation is the obvious precedent, so treating this as a preview rather than a foreign curiosity is reasonable.
That is an argument for building sound editorial process now, not for pre-complying with a rule that does not exist here. The businesses that will find future disclosure requirements trivial are the ones that already know who wrote what, with which tool, and who checked it.
Sources: Regulation (EU) 2024/1689 (EU AI Act), Article 50; European Commission guidelines on Article 50 transparency obligations (20 July 2026); Code of Practice on Transparency of AI-Generated Content; European Commission, "Quick Facts: Transparency rules for AI systems"; Anthropic, "How Claude's text watermarking works" and Claude Help Centre, "How Claude marks AI-generated content". This article is general information, not legal advice.
Nauman has spent more than 15 years in digital marketing across Australia, the UAE and Pakistan, working in-house and agency-side before founding DigiWolf. He holds an MBA in Finance & Marketing from the Institute of Business Administration, and works directly with every DigiWolf client on AI search, SEO and paid media strategy.