What 'commercially safe' should mean
The phrase is a marketing label, not a legal standard, so ask what sits behind it. First, training data: licensed, owned or public-domain material is easier to defend than scraped web images. Under the EU AI Act, providers of general-purpose AI models placed on the EU market since August 2025 must publish a summary of their training content and keep a copyright policy that honours text-and-data-mining opt-outs; ask to see both. Second, output filters that block trademarks, real people and named living artists. Third, the contract: a licence for commercial use worldwide, and any IP indemnity, with its cap, its exclusions and the plans it applies to.
You may not own what you generate
Copyright follows human authorship in most places. The US Copyright Office's report of January 2025 says prompts alone generally do not make the user an author, while human selection, arrangement and modification of generated material can be protected; a federal appeals court upheld the human-authorship requirement in Thaler v. Perlmutter in 2025. The UK's Copyright, Designs and Patents Act 1988 protects computer-generated works for 50 years under sections 9(3) and 12(7), but how that applies to generative models is untested, and in March 2026 the government proposed removing that protection. EU law asks for the author's own intellectual creation. Where exclusivity matters, as with a mascot or key visual, build in real human creative work and document it.
Products must be reproduced, not redrawn
Image models re-imagine whatever they are shown. Labels turn into plausible nonsense, proportions shift, colours drift and a bottle gains a second cap. In advertising that is not a cosmetic flaw: an image that misrepresents the product can breach misleading-advertising rules such as the UK CAP Code or the EU Unfair Commercial Practices Directive. The dependable approach is to composite real product photography or 3D renders into generated scenes instead of generating the product itself, then inspect label text at full size. Synthetic White's product-lock keeps the product's geometry, labels and colours pixel-accurate while the scene changes, on commercially safe models by default.
Provenance and records
A rights question can arrive years after a campaign ends, and the answer depends on records made at the time. For every published image keep the model and version used, the date, the inputs and references, and who approved it. Embed provenance in the file as well: C2PA Content Credentials record how an asset was made and edited, and the IPTC digital source type field marks AI-made media. Since 2 August 2026 the EU AI Act has required providers of generative systems to mark synthetic images in a machine-readable way, with a grace period to 2 December 2026 for systems already on the market, and several platforms already read this metadata when deciding whether to label a post.
Updated 25 September 2026