A harmless photograph is increasingly enough to create serious likeness abuse
The UK Report Remove service received 420 reports from children during the first six months of 2026 involving images they believed had been manipulated or faked to appear explicit. That already exceeds the 397 reports received during the whole of 2025. The service, operated by the Internet Watch Foundation and NSPCC, says many reports now involve AI or nudification tools. The majority of manipulated-image reports in 2025 and 2026 were serious enough to meet the threshold for child sexual abuse material.
Why visual creators should care
The source material does not need to be intimate. A normal portrait, school photograph or social-media image can provide enough identity information for a generative system to fabricate abusive material. For professional photographers, that changes the downstream risk attached to publishing identifiable images, particularly of children.
Practical takeaway
Photographers working with children should discuss online publication and reuse explicitly with clients rather than treating a model release as the end of the conversation. Schools and organisations should also reconsider how many identifiable high-resolution portraits they make publicly accessible. If manipulated material appears, retain the original photograph, publication date, URLs and screenshots needed to establish the source and timeline.
What to watch next
Watch whether platforms strengthen controls on uploading photographs of minors into image-editing systems and whether provenance or source-image matching becomes part of abuse reporting rather than relying primarily on detection of the final synthetic image.
Sources
Pika's new terms make a digital twin a rights-management problem as well as a creative tool
Pika updated its Terms of Service on August 4. The platform now expressly covers video generation, voice samples, likeness data, AI Selves and access to third-party models. Users retain their inputs, outputs, likeness and voice material, but Pika says that unless a separate written agreement applies it may use anonymised or aggregated inputs, outputs, interactions and usage data to train and improve its models. Personal data supplied specifically to create an AI Self is restricted to developing that user's AI Self rather than other models.
Why visual creators should care
A synthetic version of a creator can contain several separate assets: face, voice, behaviour, personal information and generated performances. Owning those materials does not necessarily mean they never contribute to model development. The distinction becomes particularly important when an agency builds a digital double from somebody else's performance.
Practical takeaway
Treat creation of an AI Self or digital double as a separate permission in performer and client agreements. Record which photographs, voice recordings and behavioural material were supplied and whether the account terms permit secondary model development. For commercial teams, check whether an enterprise agreement provides the restrictions the project requires.
What to watch next
Watch whether Pika introduces self-service no-training controls outside enterprise contracts and how clearly third-party model terms are shown when a user's likeness or client material is routed to another provider.
Sources
AI-generated video has moved from the feed into the television schedule
Roku has added Fairground AI Creator TV, a 24-hour FAST channel devoted to AI-generated programming. Fairground works with participating AI creators and says it compensates partners for generations as well as sharing platform revenue. The channel places machine-generated shorts and narrative material alongside conventional television programming and advertising rather than limiting it to an experimental AI app or social feed.
Why visual creators should care
Synthetic video now has a conventional distribution and monetisation route. That raises the stakes around source material, characters, training provenance, music, performance and model disclosure because the output is no longer simply an experiment posted to an AI community.
Practical takeaway
Creators supplying generated programming should keep a production record comparable to a conventional film deliverable: models used, reference material, voice and likeness permissions, music rights, human contributors and platform licences. Distribution by a major television platform should not be treated as evidence that all underlying rights have been cleared.
What to watch next
Watch whether Roku or Fairground publish formal provenance, model-disclosure or rights-clearance requirements and whether other streaming platforms begin commissioning dedicated AI programming.
Sources
Runway shows what professional no-training protection can look like
Runway Dev offers enterprise teams access to Runway and third-party visual, video, audio and character models through a single production environment. Runway says the service includes contractual commitments not to train on customer data, support for zero-data retention and negotiated data-retention terms for third-party model providers. The platform also returns metadata identifying which model a routing system selected for an individual generation.
Why visual creators should care
This demonstrates that training and retention are negotiable parts of professional AI procurement rather than unavoidable characteristics of generative tools. An agency can demand stronger treatment of confidential campaign imagery than a consumer account provides.
Practical takeaway
When selecting an AI service for client work, ask three separate questions: can the provider train on the material, how long can it retain the material, and can a third-party model receive it? Put the answers into the project agreement rather than relying on a general privacy statement.
What to watch next
Watch whether no-training and zero-retention terms move down from enterprise contracts into ordinary professional subscriptions and whether model-routing platforms make provider and retention information visible per generation.
Sources
The ability to remove a watermark does not create the right to remove it
Runway's current Remove from Video tool expressly lets users instruct the system to remove an object, person, logo or watermark from uploaded footage. The feature can operate on a user's own footage, licensed stock or generated material. The technical ability is straightforward; the rights position is not. A watermark may identify a copyright owner, stock licence, preview copy or provenance system whose removal is restricted by contract or law.
Why visual creators should care
Generative editing makes formerly specialised alterations trivial. That can encourage teams to treat a visible rights or provenance signal as an unwanted graphic element rather than information intentionally attached to the asset.
Practical takeaway
Before removing a watermark, logo or provenance marker, determine why it is there. Preserve the untouched source and record the licence authorising the alteration. Never use removal simply to convert a preview, unlicensed asset or provenance-bearing file into something that appears clean.
What to watch next
Watch whether generative editing services distinguish copyright watermarks from ordinary scene objects and whether provenance systems develop signals that survive visual removal or inpainting.
Sources
Luma separates output ownership from training permission and commercial-use rights
Luma's current Terms of Service say users retain copyright in their inputs and, as between Luma and the customer, own the resulting outputs to the extent permitted by law. However, inputs and outputs provided during paid use may still be used to improve services and train or develop Luma's models. Commercial use of outputs is limited to qualifying paid subscriptions. The terms also prohibit users from removing or obscuring watermarks, Content Credentials or other provenance information Luma may attach.
Why visual creators should care
Three questions that creators often collapse into one are contractually separate here: who owns the file, whether the provider may learn from the file, and whether the user may exploit the result commercially. A client asking 'do we own it?' may therefore be asking the wrong question.
Practical takeaway
Check ownership, training permission and commercial-use rights independently. When identifiable people or confidential products appear in uploaded material, make sure the client has approved the provider's permitted secondary uses as well as the generation itself.
What to watch next
Watch whether Luma introduces explicit no-training tiers or opt-outs and how its provenance restrictions interact with ordinary editing, transcoding and distribution through third-party services.
Sources
