Microsoft’s AI provenance can contain a unique identifier linked to the generation request
Research reported by The Register on August 25 found that Microsoft Paint and Photos can embed a server-issued unique identifier into AI-generated imagery. The identifier is separate from the visible watermark and is associated with the generation request. Microsoft’s own Paint documentation says generated images carry C2PA provenance and that its online safety systems collect prompts together with attributes including device and user identifiers for moderation and abuse prevention.
Why visual creators should care
Provenance is usually discussed as a way to identify how an image was made. This shows that the underlying technical record can potentially carry much more granular information about the generation process. For professional creators, that turns provenance into a privacy and client-confidentiality question as well as an authenticity feature.
Practical takeaway
For confidential visual work, find out what information the generator sends upstream and what identifiers remain attached to the output. Do not assume an invisible watermark contains only a generic 'AI generated' flag.
What to watch next
Watch whether Microsoft explains exactly how long generation identifiers and prompt associations are retained, who can resolve them back to an account, and what information is exposed through C2PA or other verification tools.
Sources
Higgsfield filmmakers keep copyright, but the production history becomes part of the deal
Higgsfield’s current $1 million Global Film Festival rules require AI image and video generation to happen inside a dedicated Higgsfield project and require entrants to retain the underlying assets and generation history for verification. After the submission deadline, eligible projects become openly viewable, including the film, prompts, generation history and retained assets. Entrants keep copyright in their films, but Higgsfield receives a non-exclusive, perpetual, irrevocable, worldwide and royalty-free licence to use compliant entries and excerpts for festival and product marketing, including paid advertising.
Why visual creators should care
An AI filmmaker may retain copyright while still giving away significant control over how the work and its production process can be used. Public generation histories can also expose prompts and working methods that creators might otherwise regard as proprietary.
Practical takeaway
Read competition rights separately from the headline statement that you retain copyright. Check what happens to prompts, project files, rejected generations, marketing rights and public visibility before entering valuable work.
What to watch next
Watch whether creator competitions increasingly require complete generation histories as proof of authorship and whether filmmakers begin treating prompts and production histories as separate licensable material.
Sources
Luma’s current terms make a private-looking AI workflow something creators still need to check contractually
Luma’s current Terms of Service say it applies security measures to protect customer input but also state that input is not covered by the agreement’s ordinary confidentiality obligations. The terms require users to hold the rights and permissions necessary for uploaded material, prohibit deceptive deepfakes and impersonation, prohibit presenting AI output as human-generated, and restrict using Luma services or outputs to train other AI systems.
Why visual creators should care
A project workspace can look private without carrying the same contractual protection as material supplied under a client NDA. For photographers and agencies uploading campaign photography, products, sets or performers as references, that distinction matters.
Practical takeaway
When confidentiality matters, check the actual contract rather than the interface. An upload being inaccessible to other users does not automatically mean it is covered by the confidentiality language you would expect in a client-production agreement.
What to watch next
Watch whether enterprise visual-AI providers begin offering clearer confidentiality, retention and data-location terms for reference images and unreleased client assets.
Sources
Leonardo’s public/private switch can change what happens to the creative material behind an image
Leonardo’s current terms distinguish private content from public content. Private inputs receive stronger restrictions on platform use, while publicly shared material is licensed more broadly for maintaining, promoting and improving the service, including model-related development. Its terms also make subscription status relevant to output rights, particularly for free users.
Why visual creators should care
For designers and AI image creators, publication is not merely a visibility choice. Making a project public can change the platform’s contractual rights in the material, while account tier can affect the rights position around generated output.
Practical takeaway
Record whether a generation was public or private and which subscription tier produced it. Do not assume the same model creates the same rights position across every account setting.
What to watch next
Watch whether creator platforms move these rights differences into clearer in-product warnings before users change a project from private to public.
Sources
Magnific says it can improve the service with your content without using it to train its own models
Magnific’s current AI-product terms say users retain ownership of their inputs and that subscribed users receive the platform’s rights in generated outputs to the extent legally possible. Users must have the necessary copyright, image, voice and personal-data permissions for uploaded material. Magnific says it may host inputs and outputs and use them for security and product improvement, but states that it does not use those inputs or outputs to train its own AI models. Private inputs and outputs are treated as confidential unless the user publishes them.
Why visual creators should care
Terms such as 'product improvement' and 'AI training' are often treated as though they mean the same thing. They do not necessarily describe the same processing, and creators need to know which one a platform is actually promising not to do.
Practical takeaway
When documenting a no-training workflow for a client, preserve the exact platform wording. Do not expand a no-model-training promise into a broader claim that the provider never processes material for safety, hosting or product improvement.
What to watch next
Watch whether Magnific provides equivalent contractual commitments for every third-party model accessible through the platform and whether enterprise users receive stronger audit or retention controls.
Sources
Ideogram’s API has a no-training promise with an important safety exception
Ideogram’s Developer API Agreement says ordinary API user inputs and outputs are not used to train the Ideogram model. It creates an exception for material flagged as violating the company’s usage policies, which may be used to train or improve safety systems. Developers must also clearly identify output as coming from Ideogram, may not upload copyrighted images or images of people without the necessary consent, and may not represent generated output as a real photograph of an actual event.
Why visual creators should care
Agencies increasingly access image models through APIs rather than consumer interfaces. API-specific terms can therefore be more important than the platform’s headline consumer policy, particularly around training, disclosure and uploaded reference photography.
Practical takeaway
If an agency builds Ideogram into a client workflow, save the API terms and document the disclosure shown to end users. Treat the safety exception as part of the data-use record rather than describing the service simply as 'never trains on API content'.
What to watch next
Watch whether API providers standardise no-training terms and disclose more precisely what happens to material retained for abuse detection or safety-model improvement.
Sources
