Generative Fill for Photographers: Rules and Rights

Generative fill lets you add, extend or remove pixels with a short text prompt. This guide explains how it works, what contests and copyright law say about it, and how to disclose it honestly.

Many photographers now meet generative fill inside software they already own. However, a tool that is easy to use is not automatically easy to defend. A fill can change what a picture claims to show, and the rules around that claim differ by context. Below, each point is tied to a primary source, and the notes follow the text.

What generative fill does in a photo workflow

Adobe introduced generative fill in Photoshop on 23 May 2023, first as a beta. You select an area, type a short prompt, and the program adds, extends or removes content. Adobe says the edits live on separate generative layers, so the original image is unaffected.¹ Indeed, that layer protects your file. However, it does not protect your credibility, because a viewer never sees the layer.

In practice, three uses cover most of what photographers do with it. First, removal: a bin, a cable or a stranger in the background. Second, extension: widening the canvas to fit a crop ratio or a text overlay. Third, addition: objects or skies that the camera never recorded. Thus, removal sits closest to old-fashioned retouching. Addition, on the other hand, creates a different kind of picture.

Nevertheless, conventional edits still do most of the work. A well-built stack of adjustment layers and masks, for example, can fix exposure and colour without inventing anything. If you want a refresher, see this guide to layers and masks in Photoshop.

Fill, expand and removal in Photoshop

Adobe’s own tutorial separates the operations. Generative Fill replaces or adds content inside a selection: you make a rough selection larger than the object, type a short prompt and click Generate. Each generation creates a new Generative layer, which the page calls a nondestructive independent layer. Generative Expand extends the canvas. You set the Crop Tool’s fill to Generative Expand and drag the boundary outward, and the prompt can often stay blank².

Removal is presented as another capability of Generative Fill. Adobe’s page does not use the name Generative Remove, and it does not say whether a prompt is needed for removal. It does not mention Content Credentials either³. Consequently, the ethics differ by operation: removal erases evidence, expansion invents context, and fill can do either. I did not read Adobe’s documentation for Lightroom’s tools, so the article does not describe them.

How generative fill works, in plain terms

Adobe does not describe its model in the pages I checked. I therefore explain the technique from the public research. Latent diffusion models, introduced by Rombach and colleagues at CVPR 2022, generate images in a compressed latent space instead of working on raw pixels. The authors reported state-of-the-art results for inpainting, the task of filling a masked region. They also showed that working in latent space lowers the computing cost compared with pixel-space diffusion.⁴

In practice, the model does not recover what stood behind a removed object. Instead, it invents a plausible fill from the surrounding context and your prompt. Because the fill is invented, textures can look too clean, shadows can point the wrong way, and fine detail such as lettering can fail. Therefore, my own working rule is simple. Zoom to 100 percent and check grain, perspective and light direction before accepting any fill.

Content Credentials: a label that travels with the file

Adobe says generative fill supports Content Credentials, which indicate whether content was made by a human, generated by AI or edited by AI.⁵ Adobe's help pages add that credentials are attached in Photoshop, Lightroom, Adobe Stock and Premiere, and that they record who created the content and how it was made.⁶ Creators can also add a verified name and a preference asking supported AI models not to train on their work. Besides, credentials can be added at each editing stage, so a file carries its history.⁷

Still, a label only helps if it survives. Check your export settings, because I could not confirm from Adobe's pages how every platform treats these labels. Finally, keep your master file, with its credentials intact, so that you can show how the picture was made.

What C2PA and IPTC metadata do and do not prove

The C2PA explainer calls a Content Credential a cryptographically bound structure that records an asset’s provenance. It gives no judgment on whether that data is true. Provenance alone cannot tell you whether content is true, accurate or factual, and the explainer says it is not a cure-all for misinformation. Unlike DRM, it does not restrict access or usage. However, the data is tamper-evident, because hashes are linked together and signed with X.509-based credentials⁸.

When an AI system performs an action, the manifest identifies it through the digitalSourceType field⁹. Separately, IPTC added the term Composite with Trained Algorithmic Media on 27 July 2023 for composites made with an AI engine’s inpainting or outpainting. It is a more specific option than the older synthetic composite term¹⁰. I did not verify which applications write that term when you use a fill.

What the contests allow: generative fill is usually out

Press and documentary contests set the strictest rules. World Press Photo announced on 20 November 2023 that its Open Format category would ban AI-generated images, including generative fill, matching its other categories.¹¹ Its rules for the 2027 contest go further. They define a photograph as a record made with a camera, disallow generative fill in post-production, and ban AI-powered enlarging and upscaling. Moreover, violations lead to automatic disqualification, and finalists must supply their original camera files.¹²

At the same time, the rules allow limited AI-powered tools. Noise reduction, automatic level, colour and contrast adjustments, and object selection for local edits are acceptable if they do not add, alter or remove captured information.¹³ The line is therefore about information, not software. For instance, a denoise pass works on data the sensor recorded. A generative fill, by contrast, adds pixels the sensor never saw.

A cautionary case from the Sony World Photography Awards

In April 2023, Boris Eldagsen won the Creative Open category of the Sony World Photography Awards with Pseudomnesia: The Electrician. He then declined the prize, saying he had entered to test whether competitions were ready for AI images. He described the work as a mix of prompt engineering, inpainting and outpainting. Those last two techniques are what a fill does to a masked area or an extended canvas. The organisers said he had misled them about the extent of the AI involvement.¹⁴

Generative fill and the EU AI Act

Article 50 of the EU AI Act sets transparency duties that apply from 2 August 2026. Providers of systems that generate synthetic images must ensure that outputs are marked in a machine-readable format and detectable as AI-generated. The duty does not apply where a system only performs an assistive function for standard editing, or does not substantially alter the input data or its semantics¹⁵.

Providers whose systems were already on the market before 2 August 2026 have until 2 December 2026 under the Digital Omnibus¹⁶. Deployers that create deepfakes must disclose that the content was artificially generated or manipulated. For an evidently artistic, creative, satirical or fictional work, the duty shrinks: they need only disclose that the content exists, in a way that does not hamper the display or enjoyment of the work¹⁷.

For a photographer, two questions stay open: whether a given fill makes the image a deepfake, and whether you count as a deployer. I did not find a primary source that answers either, so this is information, not legal advice, and the article takes no position.

Copyright and generative fill: who is the author?

The US Copyright Office addressed this in Part 2 of its report on AI, published in January 2025. It concluded that, with current technology, prompts alone do not make a user the author of the output. Prompts convey ideas, while the system largely determines the expression.¹⁸ However, copyright still protects the human expression in a work that includes AI-generated material. A human's selection, coordination or arrangement of such material can be protected as a whole, and creative modifications can qualify case by case.¹⁹ Using AI as an assistive tool does not by itself limit protection.²⁰

For a photographer, this points to a practical rule. The more of the frame came from the camera and your own edits, the clearer your authorship. Conversely, a fill that replaces a large area with generated content weakens that claim, although the Office decides case by case. Also, this is a US view, so check your own law and your contracts.

Why disclosure matters: viewers cannot reliably spot edits

Research suggests that viewers are poor detectors. In the experiments by Nightingale, Wade and Watson, 707 participants classified 66 percent of photos correctly as original or manipulated, against a chance level of 50 percent. They located the manipulation in 45 percent of cases.²¹ A second experiment, with 659 participants, gave 62 percent and 56 percent. Those photos predate generative fill, so the figures do not measure how well people spot today's edits. Still, they show that "they will notice" is a weak safeguard.

Moreover, Nightingale and Farid reported in 2022 that AI-synthesized faces were indistinguishable from real ones and were rated more trustworthy.²² Disclosure therefore has to come from you.

Disclosure has a cost, though. A 2025 PNAS Nexus study by Wittenberg and colleagues found that labelling AI-generated media online, by process or by harm, significantly lowered viewers’ belief in claims but had little effect on their likelihood to engage with the content²³. However, the study concerned AI-generated content in general, not photographers’ retouching, so the result may not transfer.

A practical checklist for disclosing generative fill

• Keep the originals. Store the RAW file and the layered master. World Press Photo asks finalists for original camera files.²⁴
• Read the current rules. Contest rules changed in 2023 and differ by organiser, so check them before every entry.
• Stay out of news and documentary work. The World Press Photo rules show the standard that press photography sets.
• Tell clients. Say in the brief or contract when AI-generated pixels are present.
• Say it in the caption. A line such as "sky extended with generative fill" costs nothing. Keep Content Credentials on.
• Prefer a conventional edit when it does the job. A grade or a crop needs no disclosure at all.

Frequently asked questions about generative fill

Is generative fill allowed in photo contests?

The World Press Photo rules for the 2027 contest disallow generative fill in post-production²⁵. Other contests differ, and I did not retrieve the current rules of Sony, Nikon or Wildlife Photographer of the Year.

Do I have to disclose generative fill?

I found no general rule that I could verify. Contest rules, client contracts and, in the EU, Article 50 for deepfakes may require it. See the EU section above. Disclosure is anyway the safer habit.

Do Content Credentials prove that a photo is real?

No. The C2PA explainer says provenance alone cannot tell you whether content is true, accurate or factual²⁶.

Who owns a picture edited with generative fill?

In the US view, prompts alone do not make a user the author, but human expression and creative selection or arrangement can be protected. See the copyright section above. Other countries may differ.

Sources and limits

However, I did not retrieve several sources, so the article does not cover them: Adobe’s documentation on Lightroom tools, Firefly training data and indemnification; Adobe Stock, Getty and Shutterstock policies; platform labelling rules; AP and Reuters standards; the Sony, Nikon and Wildlife Photographer of the Year rules; UK, EU and Japanese copyright positions; and the Official Journal text of the AI Act and the Digital Omnibus. Moreover, Sidley and the Future of Life Institute are secondary sources. Three other disclosure studies surfaced in my search, but Crossref shows only preprints for them, so I do not cite them.

Related reading: using AI without losing your creative soul, the ethical questions AI raises for photographic truth, the AI-photography manifesto, and the history of photographic manifestos, from pictorialism to AI.

1. Adobe, "Adobe Unveils Future of Creative Cloud With Generative AI as a Creative Co-Pilot in Photoshop", press release, 23 May 2023.
2. Adobe Learn, “Add, remove and expand with Generative Fill”. (Adobe Learn)
3. Same Adobe Learn page. (Adobe Learn)
4. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10684-10695. DOI: 10.48550/arXiv.2112.10752.
5. Adobe, "Adobe Unveils Future of Creative Cloud With Generative AI as a Creative Co-Pilot in Photoshop", press release, 23 May 2023.
6. Adobe Help Center, Content Credentials overview. The page names Photoshop, Lightroom, Adobe Stock and Premiere, and says credentials are applied automatically to content made with Adobe Firefly. It does not name generative fill specifically.
7. Adobe Help Center, Content Credentials overview.
8. C2PA, “C2PA and Content Credentials Explainer”, version 2.4. (C2PA)
9. Same C2PA explainer, section 7.4.4. (C2PA)
10. IPTC news item on the new Digital Source Type term. (IPTC)
11. World Press Photo, "Rules change prohibits generative AI images in all categories", announcement updated 20 November 2023.
12. World Press Photo, Entry rules. The page checked in November 2026 describes the 2027 contest (entries open 1 December, deadline 13 January 2027).
13. World Press Photo, Entry rules, 2027 contest.
14. Silicon Republic, "Artist refuses photography award after revealing photo was made by AI", reported 18 April 2023.
15. Future of Life Institute, “The EU AI Act’s Transparency Rules: A Practical Guide to Article 50”, 14 May 2026. This is a secondary source; I did not read the Official Journal text. (FLI)
16. Sidley described this as a provisional agreement on 24 June 2026; Usercentrics and Cookiebot later report it as adopted on 29 June and in force on 27 July 2026. I did not read the Official Journal text. (Sidley)
17. Same FLI guide. (FLI)
18. U.S. Copyright Office (January 2025). Copyright and Artificial Intelligence, Part 2: Copyrightability, sections II.C to II.F. This is information, not legal advice.
19. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2, section II.F.
20. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2, section II.F.
21. Nightingale, S. J., Wade, K. A., and Watson, D. G. (2017). Can people identify original and manipulated photos of real-world scenes? Cognitive Research: Principles and Implications, 2. Experiment 1: n = 707; experiment 2: n = 659.
22. Nightingale, S. J., and Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences, 119.
23. Wittenberg et al., PNAS Nexus 4(6), 2025, pgaf170; DOI verified in Crossref; I read a summary only. (DOI)
24. World Press Photo, Entry rules, 2027 contest.
25. World Press Photo entry rules, 2027 contest. (World Press Photo)
26. C2PA explainer, version 2.4. (C2PA)