ByteDance and Hollywood have moved from cease-and-desist letters to an AI copyright deal
ByteDance and the Motion Picture Association announced an agreement on August 17 to strengthen intellectual-property safeguards in ByteDance's Seedance video and Seedream image-generation models. The agreement follows a February cease-and-desist letter from the MPA over concerns that Seedance could reproduce protected film and television characters and celebrity likenesses without permission. ByteDance says newer versions already include stronger protections, and both parties say they will continue working together as the technology develops.
Why visual creators should care
This is a significant change in how a major AI copyright dispute is being handled. Instead of moving immediately into prolonged litigation, a rights-holder group and a generator provider are trying to put infringement controls directly into the image and video systems.
Practical takeaway
Do not assume a model's safeguards make an output cleared for commercial use. Character blocking and likeness restrictions reduce some risks but do not replace checking the source material, references, trademarks, underlying rights and client brief.
What to watch next
Watch what the agreement changes in practice inside Seedance, Seedream, TikTok, CapCut and Dreamina, particularly around reference images, named characters, celebrity likenesses and user attempts to work around safeguards.
Sources
California's AI transparency law is already exposing a gap between having a rule and having a working tool
A review reported by KQED on August 17 examined 13 large AI companies after California's AI Transparency Act became operative on August 2. The review found that HeyGen, Midjourney, Mistral, Synthesia and xAI did not have public detection tools available at the time of testing, although Synthesia said it embeds the required metadata and is seeking clarification on other technical obligations. The study also found that most available tools worked best on their own unedited outputs. Only Google and OpenAI were able to identify their own generated material after editing in the tests.
Why visual creators should care
The law looked relatively simple on paper: mark generated media and let people check it. Real production immediately makes that more complicated. Images are resized, cropped, recompressed and moved between applications before anyone tries to verify them.
Practical takeaway
Do not treat the existence of a detection page as proof that provenance will survive your workflow. Test an original generation and the actual delivered or published version, and save the result with the project record.
What to watch next
Watch whether California begins enforcement, whether the missing providers release compliant tools and whether detection improves after Photoshop, video editing, screenshots and social-platform processing.
Sources
Adobe is putting rights checks at the point where a custom image model is trained
Adobe's Firefly Custom Models allow creators and brands to train reusable image generators on their own assets so that generated work follows particular characters, illustration styles or photographic aesthetics. Before training, users are asked to confirm that they have the required copyright, intellectual-property, likeness and privacy permissions. Adobe also says Firefly checks uploaded images for Content Authenticity Initiative credentials and blocks material where the creator has expressed an opt-out from AI training or use.
Why visual creators should care
This moves rights management earlier in the process. Instead of waiting until a generated image causes a dispute, the system checks permission signals and requires the person building the model to make a rights declaration before the training begins.
Practical takeaway
For a client custom model, create a training-asset register. Record the copyright owner, licence, depicted people, model releases and any restrictions for every file used to build the model. A folder full of approved campaign images does not automatically mean every image is approved for model training.
What to watch next
Watch whether other custom-model platforms begin recognising machine-readable training opt-outs and whether rights declarations become independently auditable rather than relying primarily on user certification.
Sources
Europe's image industry is telling customers not to throw away the metadata the law may depend on
CEPIC's guidance for visual-media businesses explains how the EU AI Act's Article 50 transparency rules apply to image companies acting as AI providers or deployers. CEPIC recommends continuing to use IPTC metadata to identify AI-generated or manipulated material and specifically encourages customers to preserve and read that metadata instead of stripping it during processing. It also points to C2PA and digital watermarking as complementary technologies.
Why visual creators should care
A photographer or stock agency can attach the correct disclosure and still lose it further down the production chain. The duty therefore cannot stop with whoever generated or supplied the file. Publishers, agencies and clients need workflows that preserve provenance too.
Practical takeaway
Check whether your DAM, website CMS, image optimiser and client-delivery system retain IPTC and Content Credentials. If they strip metadata, keep another auditable record linking the disclosure to the delivered asset.
What to watch next
Watch how image businesses distinguish ordinary retouching from manipulations that trigger Article 50 and whether metadata-preservation requirements begin appearing in agency and client contracts.
Sources
Invisible watermarking is being connected directly to copyright recovery
IMATAG and ImageRights have launched an integrated copyright-enforcement workflow for professional visual content. IMATAG embeds an invisible identifier into an image that is linked to the asset, owner, licence and recipient. Its system is designed to retain the identifier through cropping, resizing, compression, re-encoding and screenshots. When an apparently unauthorised copy is detected online, the claim can be passed directly to ImageRights for licence-compliance work, fee recovery or legal action.
Why visual creators should care
For photographers and stock agencies, detecting that a familiar-looking photograph exists online is only half the problem. The difficult part is proving which file it was, who licensed it and whether that particular use was actually authorised.
Practical takeaway
If you use watermark-based enforcement, connect the mark to licence data rather than merely your name. Record the asset ID, client, permitted use, territory and licence period so a detected copy can be checked before an infringement claim begins.
What to watch next
Watch whether watermark evidence is tested more frequently in litigation and whether stock platforms begin embedding licence-specific identifiers into contributor images automatically.
Sources
Another provenance system is trying to keep authorship and AI-use information attached to media from the start
The Provenance for Trust platform has opened for media-industry use following a collaboration involving CEPIC, the Journalism Trust Initiative, TrustMyContent, UncovAI, Sciences Po MediaLab and other partners. The platform is intended to help authenticate the origin of media, protect rights information and resolve questions about provenance and AI use across images, video and audio.
Why visual creators should care
The useful question for professional visual work is increasingly not simply whether something is AI-generated. Publishers may need to know who created the original, what happened afterwards and whether the file has a reliable history at all.
Practical takeaway
When evaluating provenance tools, look for systems that preserve authorship and rights information alongside AI disclosure. Knowing that AI touched a file is less useful if the system cannot tell you who supplied the underlying work.
What to watch next
Watch how widely the platform is adopted by news and picture agencies and whether its provenance records survive the image-processing systems used by large publishers and social networks.
Sources
