AI Outputs That Show Your Watermark: Trade Mark Lessons for US AI Companies From the Getty Cases

When an AI image generator produces a picture with a distorted version of a famous watermark, the output looks like a bug. In law it may be a trade mark problem. In April 2026, a California federal court refused to dismiss Getty Images’ US claims against Stability AI for trade mark infringement, false designation of origin, dilution and unfair competition. Together with the UK ruling of November 2025, the Getty cases make AI generated watermark trademark infringement a live risk for any company whose model can reproduce a brand’s marks.

This guide explains what the courts have decided so far, what the claims mean, why watermarks and logos in outputs are different from training data questions, and what AI companies should do about testing, filtering, contracts and incident response. It is written by an Indian law firm. US law points are for US counsel to confirm, and we cover the India angle for outputs seen in India.

What the Getty cases decided

In the UK, the High Court in November 2025 found trade mark infringement only in limited respects, linked to Getty-style watermarks appearing in outputs from some earlier versions of Stable Diffusion, while rejecting Getty’s secondary copyright claim. The court did not treat the model itself as an infringing copy. The trade mark findings were narrow, and Stability AI was reported to have been refused permission to appeal them in the High Court. Both courts show that AI generated watermark trademark infringement claims will be tested on the facts of each model version.

In the US, Judge Trina Thompson of the Northern District of California ruled in April 2026 on Stability AI’s motion to dismiss. Reports say she let nearly all of Getty’s claims proceed, including trade mark infringement, false designation of origin, dilution and unfair competition, and dismissed one claim about false copyright management information, with permission to amend. A motion to dismiss tests only whether the claim is plausible, so the ruling is not a finding that Stability AI infringed.

These are early steps. No court has yet decided on a full record whether a model’s occasional reproduction of a watermark amounts to AI generated watermark trademark infringement in the US. Even so, the cases tell AI companies that watermark outputs are not just an aesthetic flaw. They are evidence that a plaintiff will use to argue that users may believe an image came from, or was approved by, the brand.

How trade mark claims reach AI outputs

Under the Lanham Act, a claimant can allege infringement of a registered mark when use in commerce is likely to cause confusion about source or approval. Section 43(a) covers false designation of origin, which can apply even to unregistered marks. Dilution under Section 43(c) protects famous marks from blurring or tarnishment even without confusion. A watermark that appears in an output can be said to suggest that the stock-photo brand made or licensed the image, or to tarnish it where the image is unflattering.

Defences exist. A company may argue that the watermark appears by accident and not as a mark used in commerce, that users are not confused because they know the image is AI generated, or that the use is not a use of the mark as a source identifier. Courts have also limited some claims that look like copyright claims in trade mark form, citing the Supreme Court’s decision in Dastar v. Twentieth Century Fox. The California court reportedly allowed the claims to go forward despite that argument at this stage.

Platform design affects exposure. A tool that lets users prompt for ’a photo in the style of a named agency’ may invite the output of marks. A company that sells outputs, or shows them in marketing, may be seen as using the mark itself. A company that provides only a tool may face secondary liability theories, and its terms of service and take-down process will matter. US counsel should assess which theory fits your product. Product design choices can therefore make AI generated watermark trademark infringement more or less likely.

The India angle

For a US AI company with Indian users, the same outputs can raise questions under the Trade Marks Act, 1999. Section 29 treats the use of a mark that is identical or similar, in a way likely to cause confusion, as infringement, and Section 29(4) protects marks with a reputation in India against use that takes unfair advantage or causes detriment. Indian courts have granted wide injunctions for well-known marks and are active on online platforms. Indian counsel can advise how AI generated watermark trademark infringement would be analysed in India.

Indian courts also hear claims of passing off, and Delhi High Court judges have been quick to order websites and platforms to act on infringing content. Intermediary rules require platforms to remove unlawful content on notice, and the Information Technology Rules, 2021 as amended in 2026 deal with synthetically generated content. A tool that creates outputs carrying a brand’s mark may attract complaints in India as well as in the US.

Practical enforcement in India usually starts with a legal notice, a platform complaint or an application for interim relief. A US company without an Indian entity may still be reached through its app store listings, payment flows and local partners. Plan for a response team that can handle an Indian notice quickly and check whether the company holds Indian trade marks for its own brand, since the same tools protect you.

Practical steps for AI companies

Test your models. Prompt for outputs likely to include watermarks, logos and brand names, and log what appears. Do this before every major release. Add output filters that detect watermark-like patterns and common logos, and consider blocking prompts that ask for a named brand’s style or mark. Keep records of tests, thresholds and fixes, because a diligent process helps a court see good faith. Testing is the most direct control against AI generated watermark trademark infringement.

Clean the data pipeline. Where you control training data, remove or de-emphasise watermarked images and clearly marked stock content, or license it. Record what was removed. Avoid fine-tuning on datasets scraped from stock sites unless you hold rights. Document where the training took place and which version of the model produced which output, since the UK case showed that versions matter.

Update contracts and policies. Terms of use should prohibit users from using outputs to imply an affiliation with a brand, and from removing or adding watermarks to deceive. Provide a simple complaints channel and act quickly on valid notices. For enterprise customers, decide whether to offer an indemnity for output claims and under what conditions, and review insurance for advertising injury and IP claims. Escalate any case involving a famous mark to counsel immediately.

Next steps and quick answers

In the next month, run a watermark and logo test across your current models, record the results, confirm that your filters work, and review your terms and complaint process. Brief your sales and communications teams, so that they do not promise that outputs are free of third-party marks. Ask US counsel to track the Getty case in California and the UK appeal. If you reach Indian users, add an Indian-law check to your plan.

Related reading: Getty v. Stability AI at the Court of Appeal, AI training data provenance audits after Bartz, AI music copyright lawsuits and generative AI startups, AI and copyright in India and the US compared, labelling AI-generated content under the IT Rules 2026, deepfakes and personality rights in India and trade mark squatting in India for US companies.

Quick answers

Can an AI-generated watermark infringe a trade mark?

It may, if a court finds that it is likely to confuse users about source or approval, or dilutes a famous mark. The Getty cases show courts will hear such claims, though they have not finally decided the US question.

What did the California court decide in April 2026?

Reports say it refused to dismiss most of Getty’s claims against Stability AI, including trade mark infringement, false designation of origin and dilution, and dismissed one copyright management information claim with permission to amend.

How can AI companies reduce watermark risk?

Test outputs, filter watermarks and logos, clean training data, update terms, offer a complaints process and keep records of what you did and when.

Did the California court decide that Stability AI infringed?

No. A motion to dismiss tests only whether the claims are plausible. Reports say most claims survived and one copyright management information claim was dismissed with permission to amend.

Can Indian law apply to AI outputs?

Yes, where outputs are seen in India. Sections 29 and 29(4) of the Trade Marks Act, 1999 and passing off can apply to marks with a reputation in India.

Final word

Halverton & Co. is an Indian law firm advising US companies, founders, investors and law firms on Indian law, including trade mark and brand protection, AI and technology law, platform regulation, and cross-border IP enforcement in India. We practise in Jharkhand, Maharashtra and before the Supreme Court of India, and work alongside US counsel, who advise on US law. Halverton & Co.: Where tech needs law! If you have a question about Indian law, write to us at office@halvertonandco.com, or get in touch.

This article reflects developments reported up to late October 2026. It is for general information only, is not legal advice, and does not create an attorney-client relationship. Halverton & Co. is an Indian law firm and does not advise on US law; US-law points should be confirmed with US counsel.