Can your business train AI on content scraped from the internet?

Used under a Creative Commons Licence

Can your business train AI on content scraped from the internet?

If you run a website, online platform or AI‑driven business in Australia, you might be wondering: “Can I train AI on content scraped from the internet?” or “Does an AI model ‘contain’ the data it was trained on?”

The UK High Court’s recent decision in Getty Images v Stability AI EWHC 2863 (Ch) gives some important clues – and some new questions – for anyone building or using internet‑delivered AI tools.

When Legal Letters Backfire

If you’ve ever been on either side of a legal spat over IP—whether trademarks, copyright, or patents—you’ve probably considered sending, or received, a strongly worded letter. But did you know that under Australian law, making the wrong kind of threat could land you in hot water? “Unjustified threats” claims are a real risk for businesses and creatives seeking to enforce their rights or defend their brands.

What happened?

In that case, Getty Images took legal action against the company behind Stable Diffusion, claiming it had scraped Getty’s online photos and used them to train the model without permission. The court then had to decide two key things: whether the model itself counted as an unlawful copy of those photos (which could expose the developer to serious copyright liability), and whether the appearance of the “Getty Images” watermark in some AI‑generated pictures amounted to trade mark infringement (which could trigger separate trade mark remedies and damages).

While this is a UK decision, it is already influencing how lawyers and regulators around the world think about copyright, privacy and AI – including here in Australia.​

As the US judge Learned Hand famously said, “The law has no choice but to adapt itself to new conditions of life,” and AI is testing that adaptability every day.

Key facts: web scraping, models and watermarks

Stability AI developed Stable Diffusion, a generative AI model that turns text prompts into images. Getty alleged that:​

1. millions of its online images and captions were scraped and used to train the model without a licence;
2. some AI‑generated outputs reproduced substantial parts of Getty photos;
3. some outputs even carried distorted versions of the “Getty Images” watermark; and
4. by distributing the model, Stability AI was effectively importing an “infringing article” into the UK.​

Getty eventually dropped its main “training” claim in the UK because it could not prove that the training itself happened there, and also parked its “primary” output‑based copyright claim for another day. That left the court to focus on two things: whether the model itself was an “infringing article”, and whether the use of the watermark in outputs infringed Getty’s trade mark.​

So the judge had to look at two issues: whether the model file itself was an illegal copy, and whether showing “Getty Images” in some AI‑made pictures broke trade mark law.

Trade marks, watermarks and online outputs

Getty did better on the trade mark side of the case.

Some of the images generated by Stable Diffusion showed the “Getty Images” watermark – or something very close to it – in the corner, as if they were genuine Getty photos.

Getty argued that, to an ordinary viewer, those images would look like they came from Getty’s library, even though they did not.

The judge agreed that, when the watermark appeared in that way, it was being used as a brand, not just as a random graphic.​

The court found that the watermark was being reproduced in a commercial context where it could indicate the origin of the images, and that the model was being used “in the course of trade” because Stability AI was distributing it for business purposes.

For certain older versions of the model, that was enough for trade mark infringement. In other words, even though the model itself was not a copyright‑infringing copy of Getty’s photos, using Getty’s brand inside AI‑generated images without permission still crossed the line.​

For any business using AI to create content for a website, ad campaign or social media, this has clear consequences. If your AI‑generated images show someone else’s logo, watermark or distinctive brand elements, you may be treading into trade mark infringement or passing off, particularly when those images are used to promote your goods or services online.

Because online content is shared, reposted and taken out of context very quickly, a single AI‑generated image can easily create confusion about who is behind it – and that is exactly what trade mark law is designed to prevent.​

As the saying goes, “Trust is hard to win and easy to lose” . On this note, one could say that nothing erodes user trust faster than mishandled data in an AI‑driven service.

For Australian AI developers and online platforms what does this mean?

If your model has been trained on images, text or other material scraped from websites, are you confident that this lines up with what your customers – and the law – would see as fair? How would you answer if a client, investor or regulator simply asked, “What exactly did you train this on, and did you get permission or pay anyone?”

There is also a basic fairness question around using online content in AI projects. Do you see web content as “free to use”, or do you think in terms of asking, licensing and having clear rules? If someone else scraped your own website to train their model, how would you feel about that – and are you ready for your customers to ask whether you are doing the same thing with their content?

Getty also shows how important it is to know the basics of your AI set‑up.

Could you explain, in plain terms, what sorts of data went into the models you use, who collected it and on what basis? Inside your business, is there a shared plan for how to deal with user data, visible logos or watermarks, and what to do if an AI‑generated image looks very close to someone else’s work or accidentally exposes personal information?

FAQs: online AI use

1. Does this decision mean AI developers can freely scrape images from the internet?
No. Getty dropped the main “training” claim in the UK on jurisdictional grounds, so the court did not rule that scraping is lawful.

2. Does an AI model “contain” the data it was trained on?
The UK court said Stable Diffusion does not “contain or store” training images in a way that makes the model itself an infringing copy. However, regulators in Europe see some models as potential “data stores” for personal data, so this issue is still evolving.​

3. Can AI‑generated content infringe trade marks?
Yes it would seem so! In Getty’s case, outputs that included the “Getty Images” watermark were found to infringe trade mark rights in certain circumstances.

4. How does this affect Australian businesses using AI tools from overseas?
If you use overseas AI tools on your Australian website, you still need to comply with Australian copyright, consumer and privacy law.

5. Should my business change anything now?
Yes – you can tighten your contracts, review your training data sources, and implement internal AI governance and privacy controls to reduce risk.

Please note the above article is general in nature and does not constitute legal advice.

Please email us info@iplegal.com.au if you need legal advice about your brand or another legal matter in this area generally.