ElevenLabs’ visual AI does not carry the company’s normal IP indemnity
ElevenLabs’ current Image & Video terms govern its expanding image-generation, video-generation and avatar products. The service can rely on third-party model providers, meaning uploaded material may be processed under those providers’ requirements as well. Most importantly for professional users, ElevenLabs explicitly says that any intellectual-property indemnity contained in a customer’s underlying ElevenLabs agreement does not apply to Image & Video or output generated through it. If users voluntarily share generated work publicly, ElevenLabs also receives a broad licence to use and sublicense that shared output.
Why visual creators should care
A company may have an enterprise relationship with ElevenLabs and still receive materially different legal protection when it moves from voice products into image or video generation. The interface or vendor name alone therefore does not tell an agency what protection attaches to an asset.
Practical takeaway
Before relying on a vendor’s indemnity for client work, check the product-specific terms. Record the actual image or video product and underlying model, particularly when a host platform can swap third-party providers.
What to watch next
Watch whether ElevenLabs introduces separate indemnity for approved image and video models and whether the interface identifies which third-party provider handled each generation.
Sources
One filmmaker release says generic consent is not enough for an AI version of a performer
PYLON’s current Talent and Likeness Release requires filmmakers to describe AI-generated depictions of identifiable people specifically in a separate schedule. It says generic language covering 'all media now known or hereafter devised' is not sufficient where a subject may reasonably object to synthetic depiction. The release separately identifies face swaps, de-aging, age progression, lip-sync to new dialogue, voice cloning, generative extensions of a performance and LoRA or fine-tune training on a performer’s likeness.
Why visual creators should care
Traditional model and performer releases were usually written around photographs and recordings that actually took place. Generative production can manufacture completely new shots and performances, so broad legacy language may tell the subject very little about what they are really authorising.
Practical takeaway
For synthetic people, describe the actual AI technique and scope in the release. If a LoRA, embedding or reusable model is created from a performer, record how it can be used and when it must be deleted.
What to watch next
Watch whether production companies, festivals and talent agencies begin adopting similarly specific AI schedules as standard additions to model and performer releases.
Sources
Canva makes AI-training participation a setting rather than assuming every upload is available
Canva’s Terms of Use, effective August 19, say users retain ownership of their uploaded content and give Canva the rights needed to host and provide the service. Canva says account privacy settings allow users to control whether general usage data and User Content can be used to improve AI-powered features. It also expressly prohibits data mining, extraction or scraping of Canva and its available content for AI, machine-learning or data-science purposes.
Why visual creators should care
The two rules show the distinction between platform-controlled AI improvement and third-party scraping. A creator can make a choice about Canva’s own AI use of uploaded content while Canva separately prohibits outsiders from harvesting material from the service for their models.
Practical takeaway
Check the AI privacy setting on the actual account used for client work and record it with the project. Do not assume that the fact Canva prohibits third-party scraping also means Canva itself never processes content for AI improvement.
What to watch next
Watch whether Canva moves AI-training choices to project or team level and whether enterprise customers receive auditable records showing the setting that applied when an asset was uploaded.
Sources
AI virtual staging shows why a photograph can become a third-party data workflow
Estiq’s Terms of Service took effect August 30 for its AI virtual-staging and property-imaging service. Users must hold the rights to photographs they upload and grant Estiq and its third-party AI processors permission to process those images to provide the service. The terms identify third-party providers including Google Gemini and NanoBanana and warn that AI-staged imagery is illustrative rather than a guarantee of structural accuracy, scale or physical feasibility.
Why visual creators should care
Real-estate photography is an everyday commercial workflow where apparently simple retouching can now send a photographer’s or client’s property images into external AI infrastructure. The generated room may also depict renovations, furniture or dimensions that never existed.
Practical takeaway
For virtual staging, tell the client when photographs are being sent to external model providers and retain the original unstaged photograph. Keep generated staging clearly distinguishable from documentary property imagery where accuracy matters.
What to watch next
Watch whether real-estate portals begin requiring standard disclosure for generatively staged listings and whether photographers’ contracts begin addressing downstream AI processors explicitly.
Sources
Photo Mechanic is preparing to show how normal metadata editing can coexist with camera provenance
IPTC’s Photo Metadata Conference on September 3 will include a live workflow beginning with photographs from a C2PA-enabled Sony camera and ingesting them through Photo Mechanic. Camera Bits will demonstrate its Provenance Locker technology, designed to allow normal IPTC metadata edits while retaining the provenance of the original signed camera photograph as a parent ingredient for later C2PA signing. The workflow will also demonstrate addition of an IMATAG watermark.
Why visual creators should care
Photographers need to caption, keyword, crop and otherwise process files after capture. A provenance system that breaks whenever normal editorial metadata is added is not particularly useful. This demonstration addresses the point where camera authenticity meets the real picture-desk workflow.
Practical takeaway
When adopting C2PA capture, test the whole chain rather than only the camera. Confirm that ingest, captioning, DAM processing and export preserve a trustworthy link back to the signed original.
What to watch next
Watch the September 3 demonstration for how well camera provenance survives everyday Photo Mechanic metadata work and whether similar workflows reach Lightroom, Capture One and major editorial CMS systems.
Sources
AI-video forensics is starting to look at whether motion obeys the real world
Recent research introduces MotionPhys, a detector that analyses physical consistency in motion trajectories rather than relying primarily on visual artefacts or signatures associated with a particular generator. The researchers argue that generated videos can look plausible from frame to frame while failing to preserve properties such as inertia, continuous forces and coherent trajectories across a complete action. Their experiments report generalisation across multiple video-generation systems.
Why visual creators should care
Generator fingerprints can disappear as models change. A method based on physical behaviour offers another possible forensic signal for documentary filmmakers and picture desks faced with increasingly convincing synthetic footage.
Practical takeaway
Treat physical-consistency detection as another forensic clue, not proof. For consequential footage, preserve the earliest file, identify the source and seek provenance and real-world corroboration alongside detector results.
What to watch next
Watch independent testing on current commercial generators and on footage after cropping, compression, frame interpolation and social-platform processing.
Sources
