A synthetic-media platform is making the AI label travel with the image
Twinen updated its Synthetic Media Disclosure Standard on September 10. Realistic AI-generated or materially AI-edited images, video and audio must receive visible disclosure when a reasonable viewer could mistake them for reality. The system distinguishes AI-generated from AI-edited content and can add context such as Face swap, Composite, Dramatization or Restoration. Twinen says the labels persist through its sharing and download tools and that it writes or preserves C2PA Content Credentials where feasible. Private individuals whose realistic face or voice is cloned require verifiable consent.
Why visual creators should care
A disclosure that disappears the moment an asset is downloaded has limited professional value. Twinen is treating synthetic status, transformation type and likeness consent as information that should remain associated with the media beyond the first post.
Practical takeaway
For realistic synthetic-person work, retain the actual consent or release separately from the platform label. Preserve the provenance-bearing export rather than stripping it during client delivery or social republishing.
What to watch next
Watch whether larger visual platforms adopt similarly granular context labels and whether consent assertions eventually become interoperable with C2PA or CAWG identity records.
Sources
An AI photo editor is making permission for the input photograph the user’s responsibility
AI Photo Editor updated both its Terms of Service and Acceptable Use Policy on September 10. Users retain whatever rights they already hold in uploaded photographs and prompts and grant the service a limited licence to send that material to the selected model and store the result. The terms prohibit uploading photographs, artwork, logos or other material the user does not own or license and prohibit using real people in deceptive synthetic scenes without agreement. Crucially, the service says a generated still is not automatically cleared for commercial use.
Why visual creators should care
A platform can let you generate and download an image without clearing the underlying photograph, likeness, logo or artwork. Technical permission to run the tool and legal permission to publish the output remain different things.
Practical takeaway
Keep the source-image licence or model release alongside the generation record. Do not treat successful generation, a paid subscription or a download button as evidence that third-party rights were cleared.
What to watch next
Watch whether smaller image-editing services begin offering model-specific records showing exactly which provider received a reference photograph and whether that provider can retain or train on it.
Sources
The AI indemnity can disappear precisely where a professional workflow becomes complicated
A September 9 Reuters Legal analysis examines contractual indemnification across generative-AI systems and finds that the headline promise of IP protection can be narrowed by exclusions covering customer inputs, modifications, integrations, applications built on foundation models and other downstream activity. The analysis argues that companies need to examine the complete technology stack rather than assuming that an indemnity supplied by the foundation-model provider automatically protects the final use.
Why visual creators should care
That is particularly relevant to visual production, where a designer may upload client photography into one service, call a third-party image model, modify the result elsewhere and finally incorporate it into an advertising campaign. The final asset can sit several contractual layers away from the provider whose indemnity first appeared reassuring.
Practical takeaway
For higher-risk client work, record the full generation and editing chain. Check whether indemnification still applies to uploaded references, third-party models, modifications and the final downstream use rather than recording only that the main platform offers an IP indemnity.
What to watch next
Watch visual-AI platforms for clearer model-by-model indicators showing when enterprise indemnity is active and when switching models or introducing external inputs changes the protection.
Sources
IPTC is starting to measure whether photo provenance and AI opt-outs survive in the real world
IPTC has published the programme and recordings from its September 3 Photo Metadata Conference. Among the substantive updates are forthcoming best-practice guidance for implementing C2PA in news and media, including CAWG Identity Assertions and metadata practices, and a new public monitoring service intended to track how IPTC/XMP metadata, C2PA and AI opt-out signals actually appear in imagery across the global news industry. The conference also presented audits showing that industry-wide compliance with new AI-labelling obligations remains incomplete.
Why visual creators should care
Standards matter only if the information survives actual workflows. Measuring real published files can show whether creator credits, rights fields, AI reservations and provenance records survive cameras, editors, agencies, CMS systems and publishers.
Practical takeaway
Do not assume metadata inserted at capture or export will automatically reach the audience. Test representative files after every important handoff in your own workflow, including DAM ingestion, resizing, CMS publishing and social distribution.
What to watch next
Watch IPTC’s forthcoming C2PA best-practices book and its public metadata-monitoring results for evidence of which publishers and production tools consistently preserve creator and provenance information.
Sources
C2PA is moving to the last stage before live video reaches the audience
IBC opens in Amsterdam on September 11 with Unified Streaming demonstrating Trusted Media, a system that can add C2PA provenance dynamically when live, on-demand or virtual-channel video is packaged for distribution. Rather than depending entirely on provenance embedded in an original master file, the system can attach information such as publisher identity, source attribution, editorial context and responsible-AI disclosures immediately before delivery to a CDN. The receiving player can then validate that information.
Why visual creators should care
Modern video is continually clipped, repackaged, regionalised, combined with graphics and ads and delivered in different formats. Authenticity information needs to survive that reality rather than assuming one untouched file travels from camera to viewer.
Practical takeaway
For professional video provenance, map the distribution layer as carefully as capture and editing. Ask what happens to authenticity records during packaging, ad insertion, clipping, transcoding and CDN delivery.
What to watch next
Watch the IBC demonstrations from September 11–14 for interoperability with other C2PA newsroom and broadcast systems and for evidence that provenance remains verifiable after common live-distribution transformations.
Sources
Pond5 bans generated stock while allowing real photographs and footage to earn money as training data
Pond5’s current contributor rules do not accept AI-generated content for marketplace licensing. The company says it cannot attribute authorship to an individual or ensure that creators represented in model training have been compensated. Separately, Pond5 licenses genuine contributor material in restricted datasets used for machine learning and computer vision, including generative-AI development. Contributors receive a 20 percent revenue share from dataset licences and can opt out of future dataset inclusion through account settings.
Why visual creators should care
The two policies demonstrate why 'does this platform allow AI?' is the wrong question. The platform can reject synthetic images as stock while simultaneously treating conventional photography and video as licensable AI-training material.
Practical takeaway
Stock contributors should check marketplace eligibility and dataset-training settings independently. Keeping your photographs eligible for normal licensing does not necessarily mean they are excluded from machine-learning deals.
What to watch next
Watch whether stock marketplaces move toward image-level rather than account-level dataset choices and whether contributors receive clearer information about which models or customers license their work for training.
Sources
