Generative editing workflows and professional practice

AI is now part of almost every professional photography workflow, whether or not the photographer thinks of it that way: subject-detection autofocus, AI culling, one-click masks, AI denoising, distraction removal, sky replacement and generative expand all use machine learning. The professional questions are not whether to use AI but which uses are acceptable for which kind of work, what to tell clients and audiences, what contracts, contests and stock agencies require, how to protect client images and people's likenesses, and how to keep quality high. This lesson builds a practical framework: a ladder of AI interventions from assistive to fully synthetic, matched to the expectations of different genres; contract clauses; disclosure wording; contest, editorial and stock policies; privacy and consent; quality control; and the ethical questions of labor and environmental cost. PHOT-202 lesson 12 introduced manipulation ethics generally, PHOT-320 lesson 9 covers photojournalism codes and CRE-410 covers creator platforms; this lesson focuses on running a professional workflow.

A ladder of AI interventions

Not all AI use is equal. It helps to sort tools by how much they change the content of an image.

• Level 0, capture assistance: subject-detection autofocus, computational exposure, scene recognition. The content is what the camera recorded.
• Level 1, workflow assistance: AI culling, keywording, face grouping, search and auto-selection. Nothing in the image changes, though choices are influenced.
• Level 2, global corrections: AI denoising, sharpening, lens correction and automatic tone and color. Content is preserved but rendered differently.
• Level 3, local adjustments: AI masks for sky, subject and background, used to adjust brightness, color or contrast in regions. Content is preserved, but emphasis changes.
• Level 4, removal and cleanup: removing dust, wires, blemishes or bystanders with generative or content-aware tools. Content is subtracted.
• Level 5, generative additions and replacements: expanding the frame, replacing skies, adding or changing objects, altering faces or bodies. Content is invented.
• Level 6, fully synthetic: images generated from prompts, possibly guided by reference photos. No photograph of the scene exists.

Different genres accept different levels. Photojournalism and documentary work generally allow levels 0 to 3 within limits and forbid 4 to 6 apart from dust removal. Commercial and advertising work routinely uses levels 4 and 5, subject to advertising law and client approval. Fine art can use any level, with honest description. Event and portrait clients often accept cleanup but may object to body alterations or composited faces they did not request. The ladder gives a vocabulary for contracts, captions and conversations with clients.

Editorial, contest and stock policies

Publications and contests set explicit rules. Major news agencies, including the Associated Press, prohibit adding, removing or altering content in news photographs and state that they will not use generative AI to alter photos; images generated by AI are not used as news photographs except when the AI image itself is the subject of a story and is clearly labeled (AP standards as published; check current versions). The World Press Photo Contest banned generative AI from all its categories for the 2024 contest after briefly allowing it in an open-format category, and its rules continue to define what counts as a photograph and to require original files for verification (at the time of writing). In 2023, artist Boris Eldagsen won a category at the Sony World Photography Awards with an AI-generated image and declined the prize to provoke debate; many competitions since then have added explicit AI rules or separate AI categories.

Stock agencies differ. Adobe Stock accepts generative AI content but requires contributors to tick a "Created using generative AI tools" box, to hold rights to the tools' output, to supply model or property releases when content depicts a real identifiable person or property, and to avoid artist names, real people, copyrighted works and implied depictions of newsworthy events in prompts and metadata. Its guidelines require labeling when generative AI adds a new primary subject or significantly changes a human subject, but not for background extension, object removal or retouching (Adobe guidelines checked October 2026). Getty Images and Shutterstock have taken more restrictive approaches to contributor uploads of AI-generated images while offering their own generators trained on licensed libraries. Policies change frequently, so read the current contributor terms before submitting.

Contracts and client communication

Contracts should say how AI will and will not be used. Clients increasingly ask, and some have their own policies: a brand may forbid generative content because of uncertainty over copyright in AI output (lesson 11), while another may want generative expand to adapt images for every platform. Useful clauses cover the following points.

• Permitted AI use: which levels of the ladder are included in the editing service, for example "routine AI-assisted culling, denoising, masking and cleanup; no generative additions or alterations of people without written approval".
• Disclosure: whether and how deliverables will be labeled, and whether content credentials will be attached.
• Client-supplied material: whether the photographer may upload client images, logos or products to cloud AI services, which may process data on third-party servers.
• Likeness: no alteration of identifiable people's faces or bodies beyond agreed retouching, and no training of models on client people or products without written consent.
• Ownership and warranties: what the photographer can warrant about generated elements, given uncertainty over copyright in AI output, and whether the tool provider offers indemnity.
• Approvals: client sign-off before delivering images with generative additions.

Clear clauses protect both sides. A photographer who delivers a wedding album with a generated sky in one picture may be surprised to find the couple object; a brand that receives product shots with generated props may find it cannot register copyright in those elements. A sentence in the contract avoids both disputes. Tool providers' terms matter too: some enterprise generative tools offer intellectual property indemnity for outputs, while consumer tools usually disclaim responsibility. Read the terms of the tools you use for client work.

Real estate is a good example of why disclosure depends on reliance. A buyer relies on photos to judge a property, so even modest generative changes can mislead. Many real estate boards and listing services have rules against misleading photos, and advertising law in both countries prohibits false or misleading representations. Virtual staging, adding generated furniture to empty rooms, is widely used and generally acceptable when clearly labeled.

Privacy, biometrics and client images

AI workflows often involve uploading client photographs to cloud services and analyzing faces. Face grouping in culling and gallery software creates biometric information: mathematical representations of facial geometry. Privacy law treats such information as sensitive.

In the United States there is no general federal privacy law covering photographers, but state laws apply. Illinois's Biometric Information Privacy Act (BIPA) requires written notice and a written release before collecting biometric identifiers such as face geometry, and it allows individuals to sue, which has led to large class-action settlements; Texas and Washington have biometric laws enforced by their attorneys general, and comprehensive state privacy laws such as California's treat biometric data as sensitive. Photographers using face-recognition gallery tools should explain them in client agreements, obtain written consent where state law requires it, check where the service stores data and allow guests to opt out.

Confidentiality matters beyond biometrics. Unreleased product images, private events and images of children deserve care. Before uploading to an AI service, check whether the provider may use uploaded content to train its models, and choose settings or plans that exclude it where client confidentiality requires. Many providers distinguish consumer terms, under which uploads may be used for improvement, from business terms, under which they are not.

Quality control for AI-assisted work

AI tools make errors that are easy to miss at screen size. A professional quality check should include the following.

• Inspect generative fills at 100 percent for repeated textures, mismatched grain, wrong perspective and edges that do not blend.
• Check lighting consistency: shadows and highlights in generated areas must match the direction and color of the light in the original.
• Check people: no extra fingers, distorted jewelry, changed facial features or skin that looks different from adjacent areas.
• Check text and logos in generated regions, which may be garbled or may introduce real trademarks.
• Match noise and grain: generated areas are often cleaner than the surrounding photo; add matching grain.
• Keep originals and layered files so changes can be reversed and explained.

jobs = 40
hours_saved = [1.7, 1.0, 0.8]                # culling, masking, cleanup per job (assumed)
tool_cost_year = 600 + 15 * jobs             # subscriptions plus per-job fees
rate = 60
saved = sum(hours_saved) * jobs
print(f"hours saved per year: {saved:.0f}")
print(f"value of time saved: ${saved * rate:,.0f}; tool costs: ${tool_cost_year:,.0f}")
print(f"net benefit: ${saved * rate - tool_cost_year:,.0f}")
print(f"extra jobs possible if time is reinvested at 12 h per job: {saved / 12:.1f}")

hours saved per year: 140
value of time saved: $8,400; tool costs: $1,200
net benefit: $7,200
extra jobs possible if time is reinvested at 12 h per job: 11.7

Ethics: labor, consent and the environment

Generative AI raises ethical issues beyond accuracy. Many photographers and illustrators object that models were trained on their work without consent or payment, while competing with them for commissions; using such tools commercially is, for some, a matter of professional solidarity. Some tools offer opt-out mechanisms for training, and some providers pay contributors whose work was used; lesson 11 covers the legal position. Consent is central when people are involved: generating or altering images of identifiable people, especially sexualized or demeaning content, can cause serious harm and is increasingly a crime in both Canada and the United States (lesson 11). Environmental cost is real though often overstated per image: generation consumes electricity and water for data-center cooling, and choosing efficient tools and avoiding wasteful mass generation reduces it. Finally, there is the question of trust: every undisclosed synthetic image presented as a photograph erodes the public's willingness to believe photographs, which is the foundation of documentary photography's value.

A professional AI-use policy, published on a website or included in contracts, turns these values into practice. It might state which ladder levels the studio uses for each service, a commitment to disclose generative changes, how client images are handled, a refusal to create deceptive images of real people, and how the studio responds if a client requests something it will not do. Lesson 12 asks you to write such a policy.