An AI image platform is giving paid and free generations very different rights
KOOX AI updated its Terms of Service on September 14 for its image- and video-generation and editing platform. Users retain whatever rights they hold in authorised source material. For output generated with an active paid subscription or paid credits, KOOX assigns to the user whatever rights it holds in the output, to the extent permitted by law. Free output is treated very differently: KOOX retains whatever rights it may possess and grants the user only a personal, non-commercial, non-transferable licence. A later eligible purchase can trigger an assignment of rights in earlier free output. KOOX separately warns that AI output may not qualify for copyright protection, may resemble other generations and may still involve third-party rights.
Why visual creators should care
The familiar phrase 'you own your generations' can hide plan-level exceptions. A creator could generate an image, use it in a commercial pitch and only later discover that the free tier did not grant commercial rights in the platform’s interest in that output.
Practical takeaway
Record the account tier or credit type used when commercially important work is generated. Do not rely on a general ownership statement without checking whether free, trial and paid generations receive the same licence.
What to watch next
Watch whether other visual platforms begin attaching asset-level licence information showing exactly which rights applied when each generation was created.
Sources
Broadcasters are testing AI editing without sending the unpublished footage to the AI editor
A new IBC technical paper from researchers at the European Broadcasting Union, RAI and SRG SSR describes a sovereign architecture for AI-assisted editing of confidential audiovisual material. Raw video and audio remain inside infrastructure controlled by the broadcaster. Local systems analyse the material and expose time-aligned metadata such as transcripts, shot boundaries, face clusters, scene descriptions and acoustic events to editorial AI agents. The agents can therefore reason about the production without necessarily receiving the underlying confidential media.
Why visual creators should care
Professional video teams routinely work with unreleased interviews, embargoed material, confidential client footage and identifiable contributors. Sending all of that to an external multimodal model simply to obtain AI editing assistance creates a very different risk from keeping the source local.
Practical takeaway
For sensitive productions, ask whether an AI editing system needs the original pixels and audio or can operate from locally derived metadata. Data minimisation can be a production-design decision rather than something handled only by a contract.
What to watch next
Watch whether this local-first architecture moves from broadcast research into mainstream NLEs, DAM systems and agency editing environments.
Sources
The BBC is using synthetic faces to hide identities rather than imitate them
BBC Research & Development presented two AI-driven anonymisation workflows at IBC designed to protect contributors while preserving the emotional character of their performance. Traditional face blurring or actor reconstruction can distance viewers from an interview and alter its tone. One BBC approach instead uses a bespoke head-swapping pipeline built around an autoencoder and super-resolution model so the real contributor’s identifiable appearance can be replaced while more of the underlying performance remains visible.
Why visual creators should care
Synthetic likeness technology is usually discussed as an impersonation risk. This is the reverse use case: altering a real face specifically to prevent identification while preserving journalistic meaning.
Practical takeaway
For documentary and sensitive editorial work, treat AI anonymisation as a transformation requiring its own consent and editorial record. Keep the protected original separate and document why the synthetic replacement was made.
What to watch next
Watch how broadcasters communicate this transformation to audiences and whether C2PA or other provenance records are used to distinguish protective anonymisation from deceptive face replacement.
Sources
A product-photography platform says you own the images while its AI can still learn from the product input
Stuv AI’s September 14 terms govern a B2B visual-commerce system that can turn one raw product photograph into studio, catalogue, lifestyle and editorial images as well as product video. The terms say users retain ownership of uploaded content and own generated images and video, with commercial use permitted. At the same time, users grant Stuv a licence to use, store and process their content for providing and improving the service. Stuv separately says its platform learns from product inputs and brand preferences to maintain consistency, while information about brand identity, colours and product styles is not shared with third parties.
Why visual creators should care
Ownership and machine learning are separate questions. A brand can continue to own its product photograph while still granting a service permission to process or learn from that photograph in ways relevant to the AI system.
Practical takeaway
When confidential product photography enters an AI visual-commerce system, check three things separately: ownership, permission to improve or learn from the source, and whether brand-specific information can be shared outside the service.
What to watch next
Watch whether commercial image platforms begin distinguishing account-specific brand learning from general model improvement in clearer contractual language.
Sources
Heidi Klum’s AI younger self shows what an authorised digital replica looks like inside a real campaign
s.Oliver’s current 'Casual Since Forever' campaign places Heidi Klum alongside an AI-generated recreation of her younger 1990s self. The campaign exists within Klum’s commercial relationship with the fashion company and was developed with Jung von Matt, with campaign photography by Rankin. The creative device deliberately uses synthetic reconstruction of the same identifiable person at another age rather than concealing the use of AI.
Why visual creators should care
Recent likeness stories have largely involved unauthorised replicas. This provides the useful opposite example: a recognisable synthetic version of a real model used within an authorised commercial campaign. For photographers and agencies, the difference is permission and scope, not the underlying ability of AI to reproduce the face.
Practical takeaway
A model release for conventional campaign photography should not automatically be treated as permission to manufacture younger, older or otherwise synthetic versions of the model. Put digital-replica and generative transformation rights into the agreement explicitly.
What to watch next
Watch whether agencies begin separating conventional photography usage, AI transformation and digital-replica rights into distinct clauses as campaigns like this become more common.
Sources
Pinterest is moving more visual AI work from raw images to reusable representations of them
Pinterest announced on September 14 that it has built a new multimodal AI foundation with NVIDIA for visual search, content understanding, safety and future AI experiences. A key part of the architecture uses Pinterest’s precomputed visual representations so AI systems can work with image embeddings rather than repeatedly processing the original raw images. Pinterest says the infrastructure supports more than 80 billion monthly searches and allows Pinterest Assistant to use substantially more visual context per request.
Why visual creators should care
At visual-platform scale, AI does not always interact with the photograph as a conventional file. Increasingly it interacts with machine-readable representations extracted from that photograph. For creators, that complicates familiar questions about copying, retention, opt-outs and what exactly constitutes use of an image by an AI system.
Practical takeaway
When evaluating visual-platform AI policies, look beyond whether the original JPEG is retained. Embeddings and other derived representations can remain operationally useful to AI systems even when raw-file processing is reduced.
What to watch next
Watch Pinterest’s policies and technical documentation for clearer explanations of how creator opt-outs, deletion and AI controls apply to derived visual embeddings as its multimodal systems expand.
Sources
