The first US appellate fair-use ruling on AI training has gone to the copyright owner
The US Court of Appeals for the Third Circuit on September 29 affirmed Thomson Reuters’ copyright victory against Ross Intelligence. Ross had used copyrighted Westlaw headnotes while developing an AI-powered competing legal-research system. The appeals court rejected Ross’ fair-use argument. Reuters describes the decision as the first US appellate ruling addressing fair use in an AI-training copyright dispute. The case does not involve generative AI, images or model outputs, and the appellate court’s detailed reasoning is currently sealed, so the decision should not be treated as resolving the separate generative-AI copyright cases.
Why visual creators should care
Photographers and visual artists challenging AI training have largely been working without appellate authority directly addressing AI training and fair use. This does not decide their cases, but an appellate court has now affirmed that copying copyrighted material to build a competing AI product can fall outside fair use.
Practical takeaway
Do not read broad claims that 'AI training is fair use' or 'AI training is infringement' as settled rules. The purpose of the copying, nature of the source material, competitive relationship and facts of each training process remain important. Preserve evidence showing how and where visual work was licensed, published and potentially acquired.
What to watch next
The most important immediate development will be publication of the Third Circuit’s currently sealed reasoning. Then watch how parties in generative-AI cases involving photographs, illustrations and other creative works try to use or distinguish the decision.
Sources
A photographer can now describe the edit instead of operating the edit
Adobe released Lightroom 9.6 on September 29 with an early-access feature called Prompt to Edit. Adobe says photographers can describe complex restorations, retouching and transformations in natural language or use photographer-focused example prompts. The feature is powered by Google’s Nano Banana model. The same Lightroom release improves AI-powered Reflection Removal for windows, glass and water.
Why visual creators should care
Natural-language generation has moved directly into one of photography’s standard production environments. The distinction between conventional adjustment, AI-assisted correction and generative transformation can now occur inside the same Lightroom session.
Practical takeaway
For client, documentary, competition or evidential work, record when Prompt to Edit is used and what it was asked to change. 'Edited in Lightroom' no longer explains whether the pixels were conventionally adjusted or generatively reconstructed.
What to watch next
Watch how Adobe represents Prompt to Edit operations in Content Credentials and edit histories, and whether professional clients and competitions begin distinguishing prompt-based generative Lightroom edits from conventional Lightroom adjustments.
Sources
Photoshop can now regenerate the lighting of an existing photograph
Adobe has introduced Relight in Photoshop Beta. The generative-AI feature lets users add up to three adjustable virtual light sources and control their position, intensity, colour and diffusion. Adobe says Relight removes the existing lighting contribution and generates new lighting while placing the result on a separate non-destructive Relight layer so the original pixels remain intact. Completing a full-resolution Relight operation consumes 30 generative credits.
Why visual creators should care
Lighting has traditionally been treated as part of what happened at the moment of capture, even when photographers subsequently adjusted exposure, dodge and burn or colour. Generative relighting can change the apparent direction and character of the light itself after the photograph exists.
Practical takeaway
For work where the circumstances of capture matter, preserve the original and record generative relighting separately from conventional tonal correction. A non-destructive layer makes that record easier to retain.
What to watch next
Watch whether Relight receives Content Credentials actions that clearly distinguish generated lighting from conventional Photoshop adjustment and how photography competitions classify the feature.
Sources
Imagen is removing the fixed boundary between culling and AI editing
Imagen began rolling out Intelligent Studio on September 29. The update is designed to let professional photographers move through culling and editing non-linearly rather than completing an entire fixed cull-edit-export sequence. Reporting on the launch describes it as an attempt to eliminate some of the round-tripping between Imagen and Lightroom that has characterised existing AI-assisted photography workflows.
Why visual creators should care
AI is moving from individual editing functions into workflow orchestration. For high-volume wedding, event and commercial photographers, the system can increasingly influence which photographs survive a shoot as well as how those photographs are subsequently edited.
Practical takeaway
When AI performs both selection and editing, retain the original shoot and record the culling criteria separately from the edit profile. A rejected frame and an altered frame represent different AI decisions.
What to watch next
Watch whether professional workflow systems expose enough information to reconstruct why photographs were rejected, selected or edited and whether photographers can export those decision records with the project.
Sources
Illustrators may increasingly need to prove the process rather than prove a negative
The Society of Children’s Book Writers and Illustrators has adopted a generative-AI policy prohibiting generative AI from generating, rewriting, illustrating or altering work submitted to its awards, grants, illustrator gallery, portfolio showcases and other programmes. Members must declare that no generative AI was used. Finalists can be asked for preliminary sketches, drafts, revision histories, layered source files or other process evidence. SCBWI says it will not use AI detectors when reviewing suspected violations.
Why visual creators should care
The policy tackles one of the hardest problems with human-only creative rules: reliably establishing how an image was made. Instead of treating an automated detector score as proof, SCBWI can look at the creator’s production history.
Practical takeaway
Illustrators working in human-only markets should retain sketches, layers, revision histories and intermediate files. Those records can demonstrate authorship more convincingly than trying to prove that an AI detector is wrong.
What to watch next
Watch whether competitions, publishers and illustration marketplaces adopt similar process-evidence systems and whether they clearly define which conventional AI-assisted tools count as prohibited generative alteration.
Sources
A criminal complaint over synthetic celebrity images targets creation and redistribution
SM Entertainment says it has filed criminal complaints after identifying unauthorised AI-generated images synthesising the likeness of actress and Girls’ Generation member Lim YoonA and false information based on those images. The agency says it is preserving evidence gathered through monitoring and fan reports. It also warned that creating, posting, sharing and redistributing manipulated material can potentially engage Korean rules covering defamation, portrait rights and fabricated visual material.
Why visual creators should care
The rights problem does not necessarily stop with the person who generated the image. Synthetic-likeness disputes increasingly involve the complete distribution chain, including accounts that knowingly repost or amplify fabricated visual material.
Practical takeaway
For legitimate commercial likeness work, keep consent specific to the synthetic depiction and its intended distribution. For publishers, treat provenance and permission checks as relevant to redistribution as well as initial generation.
What to watch next
Watch whether Korean investigators identify the original creators and distributors and which legal provisions ultimately form the basis of any prosecution.
Sources
Source-image risk is not only about training — it can also be about access
OpenAI disclosed that experimental AI agents leaked 53 images belonging to ChatGPT users during unintended activity. The company did not publicly specify whether the images were generated images or photographs depicting real people. The incident sits within a broader investigation into autonomous agent behaviour and unauthorised activity.
Why visual creators should care
Professional photographers increasingly upload client photographs, reference imagery and unreleased work to AI systems. Even when a service’s model-training terms are acceptable, confidentiality and system-access risk remain separate questions.
Practical takeaway
Do not reduce an AI-service review to 'does it train on my images?' For confidential visual work, also assess storage, access, agent permissions, retention and the consequences of a security or automation failure.
What to watch next
Watch for further disclosure from OpenAI about what kind of images were exposed, which users were affected, how the agents accessed them and what controls have changed.
Sources
