What are the intellectual property risks of training a custom AI model on my products?

Quick Answer

The two main risks are input infringement and data leakage. If the training set includes images you do not own (competitor photos, scraped Pinterest boards, unlicensed stock), the resulting outputs are derivative works that can infringe copyright. The second risk is platform-level: on tools without strict tenant isolation, your private product imagery can influence outputs for other users. A third, narrower risk applies if recognisable people or third-party trade dress appear in the inputs, which can trigger right-of-publicity or trademark claims even when you own the camera file.

This is not legal advice

This article is general information about IP exposure when training AI on product imagery. It is not legal advice. The training-data legal landscape is moving quickly, with several large cases active in 2026. For decisions that affect your business, talk to a lawyer licensed in your jurisdiction.

Risk Checklist

• Input ownership: only train on images you own or have licensed. Your own studio shots and lifecycle photography are usually defensible, but check the model release and any third-party logos, packaging, or props that appear in frame. Scraped reference boards are not.
• Platform privacy: confirm the tool isolates your uploaded images from the public base model and from other tenants. The contract language to look for is an explicit restriction on using your inputs and outputs for training shared models without opt-in.
• Trade secrets: be cautious uploading unreleased prototypes to web-based AI tools until you have reviewed their data handling and retention terms.
• Likeness: if your training set includes recognisable people, you need scoped written consent for AI use, not just a copyright clearance. Older "all media now known or hereafter devised" releases are usually insufficient for AI-generated derivatives, and New York's Fashion Workers Act now requires explicit AI consent for model agreements.
• Trademark and trade dress: third-party logos, packaging, or distinctive product shapes that appear incidentally in your training set can carry through into outputs. The cleanest input is your own product against a controlled background.

What the live cases tell ecommerce brands

The training-data legal question is mostly an upstream-provider problem rather than a downstream-brand problem, but the rulings shape what counts as defensible procurement.

• Bartz v. Anthropic (US): settled for USD 1.5 billion, the largest copyright settlement in US history. The final fairness hearing is rescheduled to May 14, 2026, with nearly 120,000 author claims filed.
• Andersen v. Stability AI (US): trial set for September 8, 2026.
• Getty Images v. Stability AI (UK): the High Court's November 4, 2025 ruling largely rejected Getty's copyright claims; limited trademark infringement was found for some watermark outputs.
• Thaler v. Perlmutter (US): the Supreme Court denied certiorari on March 2, 2026, leaving the human-authorship requirement settled. A purely AI-generated image is not registrable.
• US Copyright Office: the January 2025 Part 2 report on copyrightability is direct that prompts alone do not yield a copyrightable work, but a brand's separate trademark and trade-dress rights still apply.

A brand using a generative tool to produce product imagery is generally a downstream user and is not on the hook for the provider's training-data exposure unless the output itself is substantially similar to a specific protected work.

How Nightjar handles this

A clarification first, because the question often conflates two different workflows. Nightjar does not fine-tune a per-product model on your catalog. The platform uses a zero-shot workflow: you upload the actual product file as an Asset (Nightjar's term for an image stored in your Team's Library, the shared workspace your account belongs to) and the model conditions on that image at generation time. There is no custom product model to train, leak, or have stolen. That sidesteps the largest category of "trained on my products" risk before the legal questions even start.

Nightjar's Terms of Service (Section 3.c) explicitly restrict using your Inputs and Outputs to train, fine-tune, or improve general models that serve other users without your opt-in. Inputs and Outputs are kept inside your Team's Library and treated as confidential for that purpose.

Where Nightjar does let you teach the system from your reference imagery, the surface is ingredient-level rather than product-level:

• Custom Photography Style: a reusable visual direction for camera, lighting, mood, and color, built from reference Assets you upload. The same Photography Style can guide future Generations (Nightjar's term for a request to create or edit one or more Assets) without ever shipping the source references to other tenants.
• Custom Composition: a reusable arrangement for framing, angle, product placement, and pose, built from one reference Asset.
• Custom Fashion Model: a reusable AI person built from 1 to 5 source Assets, used to wear, hold, or appear with a product. A custom Fashion Model based on a real person should only be created when you hold the right to use that person's likeness.

For scaled production, Nightjar also gives Teams a feature called Recipes: a saved Create-form setup that captures the photography style, composition, model choice, background, custom directions, and output settings, so you can apply the same vetted look to the next product without rebuilding the brief. A Photoshoot Workflow expands one approved input Asset into four cohesive variants without re-prompting from scratch, and Edit Shortcuts (/color, /ratio, /format, plus Try On, Recolor, Product Placement, Reframe, and Change Format) let you iterate on an Asset without exporting to another tool. Each of those flows operates on the Assets already in your Library, which keeps the chain of custody short.

The procurement bar for any vendor in this space is the same: ask where the model is trained, whether your Inputs and Outputs are used for shared-model training, and whether the contract restricts that use without an opt-in. The legal guide to AI product photography walks through the broader vendor due-diligence checklist.