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

Quick Answer

Training a custom AI model on product images can create copyright risk if you do not control the necessary rights to every image, plus trade-secret, likeness, trademark, privacy, and contract risk. In the United States, training is not categorically lawful or unlawful: copying can implicate a copyright owner's exclusive rights, fair use depends on the facts, and an output is not automatically a derivative work merely because the model encountered an image. Clear the inputs, restrict how the model and vendor may use them, and review outputs before commercial use; other jurisdictions apply different rules.

This is not legal advice

This article gives general information, not legal advice. It focuses on the United States and flags relevant EU rules as of August 19, 2026. Ask a lawyer in each jurisdiction where you train the model or use its outputs to review a valuable campaign, an unreleased product, or a disputed right.

Which rights should I clear before training a custom product-image model?

Owning the product does not necessarily mean you own its photos. In the United States, copyright initially belongs to the author unless employment, work-made-for-hire, assignment, or licensing rules change that result under 17 USC § 201. Confirm that photographer, agency, stock-library, supplier, and employee agreements authorize model training; possession of the image file is not enough.

The legal analysis also depends on where the training occurs. US copyright owners control reproduction and derivative works under 17 USC § 106, subject to defenses such as the four-factor fair-use test in 17 USC § 107. The US Copyright Office's 2025 pre-publication report on generative-AI training says the answer can turn on the works, source, purpose, output controls, and market effect. In the EU, Article 4 of the Digital Single Market Copyright Directive provides for text-and-data-mining exceptions for lawfully accessible material only when rightsholders have not appropriately reserved those rights; national implementation still matters.

Check four additional categories before training:

• People: obtain written permission that specifically covers creating and using a digital replica. New York's Fashion Workers Act, for example, requires clients to obtain prior written consent stating the scope, purpose, pay, and duration under Labor Law § 1037. Other states and countries use different publicity, privacy, labor, and data-protection rules.
• Third-party branding and artwork: remove unlicensed logos, illustrations, packaging, and distinctive trade dress unless they are necessary and cleared. Commercial output that creates likely confusion about origin, sponsorship, or approval can trigger the US Lanham Act even when the training images themselves were lawfully obtained.
• Confidential products: treat unreleased designs, prototypes, and launch plans as confidential. US trade-secret law requires reasonable measures to keep economically valuable information secret under 18 USC § 1839, so an unrestricted upload can undermine the position that the information was protected.
• Personal data: images of identifiable people can create privacy obligations separate from IP. For EU and EEA use, the European Data Protection Board's AI-model opinion addresses lawful basis, anonymisation, and models developed from unlawfully processed personal data.

Does training make every model output a derivative work?

No. US copyright law separates copying during model development from whether a particular output infringes a protected work. An output is not automatically derivative merely because a training image influenced the model; the output must be assessed against the protected expression it allegedly copied. Keep records of the training set, licenses, model version, prompts or settings, and output review so you can investigate a close match instead of relying on the assumption that training permission cleared every output.

What should a custom-model contract cover?

A custom-model contract should define permitted training use, ownership or licensing of any model weights or adapters, and who can access the model. It should also address whether inputs, outputs, and feedback may improve shared models; storage location; security; subprocessors; retention and deletion, including backups; incident notice; export and termination; and responsibility for third-party claims. "Private" or "isolated" marketing copy is not a substitute for those terms.

Does Nightjar train a custom model on my Product photos?

Nightjar does not present Product creation as per-product model training. A Product is a reusable subject that groups the photos defining one item with a factual description and optional physical dimensions, and Nightjar supplies that context to each image generation. A separate feature called a Recipe saves reusable photography direction and output settings without saving the Product itself.

Nightjar's Terms of Service state that it will not use, retain, or access a user's Input or Output to train, fine-tune, or improve general underlying AI systems that serve other users unless the user explicitly opts in. That contractual restriction is the relevant safeguard; it should not be confused with a claim that Nightjar trains a private product model.