Adobe has paid Stock contributors whose work was considered for Firefly training
Adobe updated its Firefly FAQ for Stock contributors on September 16 and added the 2026 Firefly Contributor bonus to eligible contributor balances on September 17. Adobe says this year’s payment covers eligible photos, videos, vectors, illustrations and generative-AI content considered for Firefly training between June 3, 2025 and June 2, 2026, with the amount also influenced by licensing activity during that period. Adobe explicitly confirms that Stock video is now used in training Firefly. It also says the existing Stock contributor terms permit this training and that there is currently no opt-out option for Adobe Stock content.
Why visual creators should care
For photographers and filmmakers, this makes the training bargain unusually visible. The contributor keeps participating in the Stock marketplace while Adobe’s existing agreement also permits eligible work to enter a generative model, with a discretionary bonus rather than a separate per-asset training licence.
Practical takeaway
Stock contributors should treat marketplace licensing and AI-training permission as separate economic uses even when both sit inside one contributor agreement. Keep records of which libraries can train on submitted work, whether an opt-out exists and what compensation mechanism applies.
What to watch next
Watch whether Adobe changes its no-opt-out position, whether video materially changes future Firefly bonus calculations and whether contributors receive more asset-level information showing which files were actually used in training.
Sources
Meta is being told that saying 'AI' is not enough when a realistic deepfake can cause serious harm
Meta’s Oversight Board published decisions on September 17 concerning realistic AI-generated videos on Facebook. In one case, a synthetic video depicted a Scottish local politician apparently making inflammatory remarks she had not made. The content was reported, but Meta’s systems did not prioritise it for human review and Meta initially kept it online. The Board ultimately required removal under Meta’s Hateful Conduct rules. More broadly, a majority concluded that Meta’s AI-labelling framework needs a lower threshold for its stronger 'High Risk AI' label and recommended additional viewing friction, stronger penalties for repeat distributors and greater transparency about label use.
Why visual creators should care
A generic AI disclosure tells the viewer how material may have been made but says little about the risk attached to a convincing synthetic performance of an identifiable person. Platforms are beginning to distinguish ordinary synthetic media from manipulations that require stronger warnings or reduced distribution.
Practical takeaway
For publishers and agencies, do not treat every synthetic-content label as equivalent. Record whether the asset depicts a real identifiable person, whether the depicted action or statement actually occurred and whether stronger disclosure is appropriate because the synthetic element changes meaning rather than merely appearance.
What to watch next
Meta must respond to the Board’s recommendations. Watch whether it expands High Risk AI labelling beyond crisis and election contexts and whether its reporting systems route more realistic deepfakes to human reviewers.
Sources
A portrait prize is testing what declared AI-assisted art looks like when it sits beside conventional portraiture
The 2026 Brisbane Portrait Prize opens this weekend after two years of argument over its decision to allow declared generative-AI work. Several artists who previously boycotted the competition have returned. Of 49 works in the main competition, The Guardian reports that Jackie Ryan’s The Catch is the only finalist declaring generative-AI use. Ryan began with her own photograph of comedian Alice Fraser and used Krea to generate additional figures, scenery and graphic elements. Prize director Anna Reynolds says generative work is eligible provided it complies with copyright requirements and retains human creative control.
Why visual creators should care
The practical boundary is no longer simply photography versus AI. A single work can contain a conventional portrait photograph, generated supporting figures and extensive human-directed composition. Competitions, clients and publishers increasingly need disclosure capable of describing that mixture.
Practical takeaway
For hybrid visual work, keep the photographic source and generative additions distinguishable in the production record. 'Made with AI' is less useful than recording which elements were photographed, which were generated and what source or reference material entered the model.
What to watch next
Watch whether visual competitions move toward standard disclosure categories for AI-assisted editing, generated additions and fully synthetic imagery rather than a single AI declaration.
Sources
Deepfake checking is moving from specialist verification desks into an ordinary browser tool
F-Secure launched its free Deepfake Detector on September 17. Users can upload an image or video and receive a score indicating whether the file may have been artificially generated or altered. Supported image formats include JPEG, PNG, WEBP and BMP, with MP4 and MOV supported for video. F-Secure presents the result as an assessment rather than proof of authenticity.
Why visual creators should care
Photographers, editors and clients increasingly encounter suspicious imagery outside specialist forensic environments. Easy access to detection can provide a useful first screening step, but statistical detection is fundamentally different from cryptographically verified provenance or an original source record.
Practical takeaway
Use AI detection as one signal, not as an authenticity certificate. For important visual material, combine it with source verification, metadata, Content Credentials where available and contact with the photographer or publisher.
What to watch next
Watch independent testing of false positives and false negatives, particularly on photographs that have been heavily retouched, recompressed, screenshotted or passed through social-media platforms.
Sources
A multi-model AI service now makes downstream clients part of the rights chain
NanoGPT substantially revised its Terms of Service on September 17 for a service providing access to numerous third-party models, including image models. Users retain their input rights and, as between themselves and NanoGPT, own generated output to the extent permitted by law. Generated images, video and audio may be used commercially, but underlying model-provider restrictions still apply. NanoGPT says it does not itself receive permission to train on or sell user content. The new terms go further for commercial integrations: businesses exposing the service to employees, contractors, customers or other downstream users must bind those users to appropriate terms, provide legally required notices and obtain necessary consents. The content provisions explicitly include rights and consent for a person's image, likeness and voice.
Why visual creators should care
An agency can no longer assume the rights check stops between itself and the AI platform. When a multi-model system is incorporated into a client product or production service, provider terms and consent obligations can travel further down the chain.
Practical takeaway
For agency and client integrations, record the front-end platform, underlying model, provider terms and downstream users. Make sure client-facing agreements cover reference-image rights and likeness consent rather than assuming the platform’s output assignment solves those questions.
What to watch next
Watch whether multi-model services begin attaching provider-term snapshots to individual generations so commercial users can prove which model rules governed an asset months later.
Sources
An AI-only stock library is separating commercial licensing from the question of whether copyright exists
Genera Stock’s current terms describe a marketplace where every visual asset is AI-generated. Contributors must have commercialisation rights under the terms of the generator they used and warrant that their prompts and workflow did not target identifiable real people, named living artists, specific copyrighted characters or artworks, brands, logos or product designs. Customers receive a non-exclusive worldwide perpetual commercial licence for paid downloads. Contributors retain ownership only to the extent the law recognises rights in their AI output, while Genera receives the licence needed to market and sublicense accepted assets.
Why visual creators should care
AI stock makes the difference between copyright ownership and contractual permission particularly visible. A customer may have a valid commercial licence to use an image even where the underlying AI output receives limited or no copyright protection.
Practical takeaway
When licensing synthetic stock, record the marketplace licence rather than assuming the asset carries conventional photographic copyright. For contributors, keep the generator terms and workflow records that support your representation that you had the right to commercialise the output.
What to watch next
Watch whether AI-only stock marketplaces begin attaching generator identity, provenance or model-level rights records to individual assets rather than relying primarily on contributor warranties.
Sources
