AI Transparency Suggestions

The most common mistake here, and one this tool made until 3 August 2026, is assuming a deepfake has to copy a specific real person. It does not.

The Guidelines say a subject counts as "existing" if it exists, can plausibly exist, or could plausibly have existed. So a photorealistic invented person is inside the definition, because such a person plausibly could exist. The same goes for realistic depictions of invented objects, places, animals, and events.

What falls outside is the impossible: a sphinx flying over the Eiffel Tower, dragons, elephants driving cars. Things that defy nature or biology are not pretending to be real, so nobody could be misled about reality by them.

The working heuristic: photorealistic and plausible is inside, stylized or impossible is outside. That is a much better predictor than asking whether a real person was copied.

Practical upshot: the EU, New York, and the platforms all agree that an invented human in an ad needs disclosure. They arrive there by different routes, but there is no gap to exploit.

Mostly you are outside these rules, but not entirely, and the exceptions are the ones that bite hardest. Three different tests are running, and only one of them cares which tool you used.

• EU Article 50: does not apply. The rules govern AI systems. A manual composite, however elaborate, is not an AI system, so there is no marking or labeling duty.
• Platform AI labels: do not apply. YouTube, TikTok, and Meta are all asking about AI-generated or AI-altered content. Hand compositing does not trigger the AI toggle.
• FTC deception: fully applies. This one is tool-agnostic. If the finished image misleads a buyer about a real product, it is deceptive whether you made it with AI, in Photoshop, or with scissors and glue. Nobody at the FTC is asking which software you opened.
• New York's synthetic performer law: probably applies. The statute covers a digital asset created "using generative artificial intelligence or a software algorithm." A CGI human in an ad is caught by that wording even with no AI involved. The duty attaches where you have actual knowledge a synthetic performer is in the ad.

The short version worth remembering: the AI rules follow the tool. The deception rules follow the result. New York's rule follows the output, meaning a fabricated human, regardless of how it was made.

So switching from AI to manual work removes the labeling paperwork. It does not make a misleading image safe, and it does not get a fake spokesperson past New York.

They follow the audience, not your address. They apply where the business is based in the EU, sells there, or the asset is aimed at an EU audience. Without one of those, everything below Tier 3 is a trust choice rather than a legal duty, and it should never be presented to a client as a legal duty.

Tier 3 and Tier 4 still matter everywhere regardless. Misrepresenting a real product is a false advertising problem in any market, and the Tier 4 lines are a studio policy rather than something copied from one regulator.

The Digital Omnibus did postpone deadlines for high-risk AI systems, and that got most of the coverage. Article 50, the transparency and labeling rules this tool is about, was not delayed. It applied on 2 August 2026 as scheduled.

There is one narrow grace period, and it does not help you. Systems already on the market before 2 August 2026 get until 2 December 2026 to meet the machine-readable marking requirement. That is a provider-side duty sitting with the AI tool vendors. It does not extend any deadline for the person publishing the content.

Almost every point of confusion in this subject comes from treating these as one thing. They are not.

• Naming your file (_ai-gen or _ai-mod) is a habit for your own organization. No law asks for it. It exists so you can scan a delivery folder and swap an AI asset later without hunting.
• Notifying the audience is the legal question. Does the law require a visible disclosure where people will see it. A file can carry the tag and need no public label at all.
• Notifying the platform is a third thing, usually a checkbox in the upload flow. It is contractual, not legal, and the threshold is often lower than the law's.

So do not read "it got tagged" as "the law says disclose this." And do not read "the law does not require it" as "the platform will not flag it."

It comes down to who decided to use AI, not who pressed publish.

• A contractor posting to a client's account, on that client's instruction, does not become separately responsible. The client stays responsible.
• A contractor who decides on their own to use an AI tool for the job does take it on.

So a clear record of who chose to use AI matters more than who clicked the button. When that is murky, label it rather than argue about it later.

There are two layers of marking, and only one of them is yours.

• Hidden provenance data baked into the file is the AI company's legal duty, whether that is Adobe, Midjourney, or another vendor. Using tools that already do it is enough on your end. It is not something you need to build.
• The visible disclosure is yours. That is the whole subject of this tool.

Worth knowing that the platforms read that hidden data, which is how they label your content automatically whether or not you declare it.

• The business-to-business exemption in the EU rules is narrow and does not really cover marketing or advertising. Do not use it to move an asset down a tier.
• A client's own written no-AI policy overrides this entire tool for that client. Human-made and light editing only, with no AI-generated or AI-modified assets, until they say otherwise in writing.

AI literacy, Article 4. If you use AI tools in your work and you have EU exposure, you are required to make sure the people using them understand what they are and what can go wrong. This is separate from disclosure and it has nothing to do with labels.

The Commission's own FAQ answers the exact case: a company whose staff use ChatGPT to write advertising copy or translate text does have to comply, and should make sure those people are informed about specific risks such as hallucination.

It is lighter than it sounds. There is no certificate, no mandated training format, no AI officer, and no governance board required. Keeping an internal record of whatever training or guidance you gave is enough. It also reaches contractors and freelancers working on your behalf, not just employees.

Timing worth knowing: the obligation has technically applied since 2 February 2025, but enforcement only began 2 August 2026. Note also that the Commission proposed shifting this duty onto Member States in the Digital Omnibus of 19 November 2025, so it may change. Watch it rather than treating it as settled.

The headline figure is up to EUR 15 million or 3% of worldwide turnover, whichever is higher. For SMEs, including start-ups, that flips: the cap is whichever of the two is lower. For a small practice that means the percentage, not the fifteen million.

Enforcement sits with national market surveillance authorities in each member state, not with Brussels directly, and they are required to be proportionate, weighing the nature and gravity of the infringement and whether it was intentional or negligent.

Short answer: usually yes, and this is the single most misunderstood point in the whole subject. Trademark and copyright are separate systems with separate rules, and AI lands differently in each.

Trademark: generally available. The USPTO does not ask who or what created a mark. It asks whether the mark identifies the source of your goods, whether you are actually using it in commerce, and whether it conflicts with something already registered. AI-generated logos have been registered. Nothing about using AI disqualifies a mark on its face.

Copyright: not available for AI-only work. US law requires a human author. The Copyright Office refuses registration for work created without human creative input, courts have upheld that, and the Supreme Court declined to reconsider it in March 2026. Where a human meaningfully shapes, selects, arranges, or edits AI output, that human contribution can be registered, and the AI-generated parts have to be disclaimed on the application. The Copyright Office has registered hundreds of works on that basis.

So why avoid AI for a finished logo anyway? Not because it is unprotectable. Because of what you give up and what you take on:

You lose the copyright layer, so you can still stop a competitor using your mark as a brand identifier, but you have a weaker position against someone simply copying the artwork.
Distinctiveness gets harder. Similar prompts produce similar results, and a mark that looks like everyone else's is a weaker mark by definition.
Conflict risk goes up. AI works from patterns in existing material and can produce something close to a mark already in use.

Whatever you do, a professional trademark clearance search before committing to a mark is not optional, and that is true for human-made logos too. This tool does not replace one.