Skip to content

Image data

Model and Property Releases for AI Training Images: People, Property, Logos and Artwork

Quick answer

A model release signed for advertising or stock licensing does not automatically cover AI training. Buyers should require release language that names machine-learning training as a permitted secondary use, confirm the release scope matches the recognizable people and private property in frame, and treat logos, trademarked products and artwork as separate rights questions. Where images come from operations rather than photo shoots, employee notice, customer privacy terms and biometric law usually matter more than releases.

By SourceX Editorial · Updated

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

When a model release is triggered in an image dataset

A model release is triggered by recognizability, not just by a visible face. Adobe's contributor rules require a release whenever a person can be identified, including by tattoos, distinctive clothing or their surroundings [1][2]. For training data this matters because the frames that carry the most signal, such as close-ups of hands on a tool, a worker at a named workstation or a customer at a counter, are often the ones where a person is identifiable from context.

Run the recognizability test at the frame level, not the collection level. A warehouse dataset can be mostly faceless conveyor shots with a long tail of break-room frames with identifiable staff. Ask the supplier for a per-image flag such as person_present and person_recognizable, and for the release identifier attached to each recognizable person.

Does an existing model release cover AI training?

Usually not unless it says so. Releases written for stock, editorial or marketing use typically grant rights to "use, reproduce and publish" a likeness in media, and that wording may not extend to ingestion into a model, derivative weights or synthetic outputs. Some release templates now name secondary use including AI and machine-learning training explicitly [3], which signals that the market treats training as a distinct use to be consented to rather than assumed.

Check four things in any release a supplier relies on:

  • Use clause: does it name AI or machine-learning training, model development or "analysis by automated systems"?
  • Sublicensing and transfer: can the photographer or company pass the rights to a third-party licensee such as your team?
  • Duration and revocation: is consent perpetual, time-limited or revocable, and what happens to models already trained if it is withdrawn?
  • Sensitive-use exclusions: many releases bar defamatory or sensitive contexts, which matters if you generate people in new scenes.

A release signed before the subject could have contemplated AI training is a weak basis for a generative model and a moderate one for a narrow classifier. Document which reading you rely on.

Property releases: homes, interiors, artwork and trademarked objects

A property release covers private property the photographer does not own, and some stock release terms extend it to private homes, artwork and trademarked objects [3]. In training sets this most often surfaces in real estate interiors, retail fixtures, branded packaging and vehicles. Contributor forums show that even experienced photographers are unsure whether a release, a copyright license or nothing at all is needed when a branded lookalike, such as a recognizable car make, appears in an AI-related image [4].

Separate three rights that a single "property release" field tends to blur:

  • Ownership or access: permission from the owner of a private building or interior to photograph and use it.
  • Copyright in the depicted work: a painting, mural, sculpture or printed poster in frame is a copyrighted work, and incidental inclusion is assessed differently from a work that is the subject of the shot.
  • Trademark and trade dress: logos, product shapes and packaging are not copyright questions, and risk depends heavily on what your model outputs.

For listing and inspection imagery, see real estate listing photos licensing and property inspection photos linked to findings.

Logos and branded products: recognition versus generation

Whether logos are a problem depends on whether your model recognizes brands or generates images. Recognition models for SKU detection on retail shelves need brands in frame, and the output is a label or bounding box rather than a reproduction of the mark. Generative models add output-side risk: Carlini and colleagues extracted over a thousand training images from diffusion models, including photos of individual people and trademarked logos [5].

That finding makes training-side rights and output-side controls a single question for generative buyers. If a model can regurgitate a logo or a released person's likeness, the release and license terms are tested at inference time, not only at ingestion. State the intended use, recognition or generation, in the license so the supplier's representations match how you will deploy.

Operational photos: when releases give way to notice and privacy law

Images captured in the course of business, such as technician job-site photos, customer-submitted damage photos or quality inspection shots, rarely come with model releases. The relevant consent instead sits in employment notices, customer terms of service and privacy policies. Notice rules vary by state and by type of monitoring: New York, for example, requires prior written notice when employers monitor employee telephone, email or internet use [7]. A buyer should ask how capture was disclosed to workers before relying on workplace imagery.

Biometric law adds a separate layer. Illinois BIPA excludes photographs from the definition of biometric identifier but covers scans of face geometry, with written release, retention-schedule and destruction duties attached [6]. If you or a vendor will run face detection, embedding or clustering on the images, the processing can create biometric data even when the source photos were ordinary. See biometric data in AI training datasets and face data consent, releases or anonymization for the deeper treatment.

The practical choice for operational imagery is usually between blurring or cropping people out, which removes the release question, and confirming that notice and terms cover AI training for the recognizable frames that remain. Stripping location and device metadata is a parallel step covered in EXIF metadata in image training data. If you need operational imagery with this review done per dataset, you can send an image data request to SourceX.

Releases address likeness and property, but copyright in the photographs themselves is a separate chain. As of October 2026 the US Copyright Office's Part 3 report on generative AI training remains a pre-publication version released in May 2025 [9], and US litigation over training on copyrighted works is ongoing. Rely on a license from the copyright holder rather than on a fair-use theory for licensed operational imagery.

If you place a general-purpose model on the EU market, Article 53 of the AI Act requires a copyright compliance policy and a public summary of training content; these duties have applied since 2 August 2025 [8]. Release and license metadata kept per image makes both easier to produce. For the license side, see AI data license terms.

Release review checklist for image dataset procurement

Use this checklist when a supplier presents releases or claims they are not needed. It pairs with the broader training data due diligence checklist.

Illustrative example: invented to show structure; it does not describe an available dataset.

CheckWhat to requestRed flag
Recognizability flagsPer-image person_recognizable, minor_present, property_private fieldsOnly a collection-level "people may appear" note
Release linkageRelease ID mapped to each recognizable person and private propertyReleases exist but are not linked to frames
Use scopeRelease text naming AI or machine-learning trainingWording limited to "advertising, promotion and publication"
TransferabilityRight to sublicense or assign to licenseesRelease personal to the original photographer
Revocation handlingProcess for withdrawn consent and affected record IDsNo record of who can withdraw or how
MinorsParent or guardian signature and age recordMinors in frame with adult-only releases
Artwork and logosFlags for artwork_in_frame, logo_in_frame, and intended useBrand-heavy images licensed for generation without review
Operational captureEmployee notice and customer terms covering trainingWorkplace footage with no disclosure record
Biometric processingStatement of whether face templates were createdFace embeddings shipped with images

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "image_id": "img_000418",
  "person_recognizable": true,
  "release_ids": ["MR-2291"],
  "release_scope": ["commercial", "ml_training"],
  "property_private": false,
  "artwork_in_frame": false,
  "logo_in_frame": true,
  "licensed_use": "recognition_only",
  "faces_blurred": false
}

A manifest like this lets counsel filter the training set by release scope and lets engineers exclude logo-bearing frames from a generative run without re-reviewing every file. For delivery formats that carry these fields, see dataset delivery formats and schemas.

Request rights-reviewed image data for AI training

SourceX sources operational datasets, including documents and new recordings of hands-on work, from US companies on request; every release is approved by the supplying company and a request does not guarantee a match. Each dataset is rights-reviewed for ownership and consents, personal details are removed or replaced before delivery with the method recorded, and delivery happens under a license defining records, uses, term and delivery. Browse the image data hub or image and inspection photo licensing, then describe the image data you need.

Sources

  1. Adobe, "Model release (Adobe Stock contributor legal requirements)". https://helpx.adobe.com/stock/contributor/legal/model-release.html
  2. Adobe, "Model release overview". https://helpx.adobe.com/ca/stock/contributor/content-policies-guidelines/model-property-releases/model-release-overview.html
  3. pocstock, "Model release". https://pocstock.com/legal/model-release
  4. Adobe Community (Stock Contributors forum), "About property releases and copyright content for AI images" (2023). https://community.adobe.com/t5/stock-contributors-discussions/about-property-releases-and-copyright-content-for-ai-images/td-p/13766273
  5. USENIX Security 2023 (Carlini et al.), "Extracting Training Data from Diffusion Models" (2023). https://www.usenix.org/conference/usenixsecurity23/presentation/carlini
  6. Illinois General Assembly, "Biometric Information Privacy Act (740 ILCS 14/)". https://www.ilga.gov/legislation/ilcs/ilcs3.asp?ActID=3004
  7. New York Public Law, "N.Y. Civil Rights Law Section 52-C*2". https://newyork.public.law/laws/n.y._civil_rights_law_section_52-c*2
  8. European Commission, AI Act Service Desk, "AI Act Article 53: Obligations for providers of general-purpose AI models". https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-53
  9. Digital Policy Alert, "United States Copyright Office released pre-publication version of Part 3 of its report on Copyright and Artificial Intelligence, titled Generative AI Training" (2025). https://digitalpolicyalert.org/event/29927-united-states-copyright-office-released-pre-publication-version-of-part-3-of-its-report-on-copyright-and-artificial-intelligence-titled-generative-ai-training

Tell us what your models need

Share scope, volume, language, format, timing and licensing requirements.

Request data