Cara’s scrape shows how little a creator opt-out can mean once the image is public
WIRED reports that Cara, the portfolio platform co-founded by photographer Jingna Zhang and designed partly for artists opposed to unauthorized AI training, has suffered three major scraping incidents this month. The first scraper obtained roughly 12 million publicly accessible works and published the archive online. A second collected about 8.5 million links plus usernames, titles and tags, while a third obtained around 123,000 images together with posts and user information. The person behind the first scrape later apologised, deleted his dataset and began collaborating with Cara on Lantern, an open-source system designed to fingerprint artists’ images and alert them if the work appears in publicly available AI datasets.
Why visual creators should care
The incident exposes the practical gap between saying that artwork must not be used for AI training and technically preventing somebody from collecting it. Cara already prohibited this kind of use and offered tools such as Glaze, yet public accessibility still made mass collection possible.
Practical takeaway
Treat a training prohibition as evidence of your rights position, not a guarantee that scraping cannot occur. Preserve original files, publication dates, licence terms and stable identifiers. Dataset-monitoring tools may become increasingly useful for discovering unauthorized use after collection has happened.
What to watch next
Watch the development of Lantern, Cara’s legal response and whether dataset hosts become more responsive to rights-holder removal requests involving links rather than copies of the underlying artwork.
Sources
A documentary is creating footage of a living person who never filmed it
The documentary 1971: Billy Joel’s Lost Interview uses genuine audio from a 1971 radio interview but AI-generated imagery to recreate a 22-year-old Billy Joel, his manager and interviewer. Joel is not involved in the film. Reporting published August 29 says director Adam Ripp is defending the synthetic recreation as expressive documentary filmmaking despite criticism over Joel’s lack of participation. The 20-minute film is scheduled to premiere September 25 at the Long Island Music & Entertainment Hall of Fame’s documentary festival.
Why visual creators should care
Documentaries traditionally use archives, reconstruction and actors to depict events that were not filmed. Generative AI adds a new option: making realistic-looking footage of the actual person. That creates a difficult boundary between editorial expression and control over a living person’s identity.
Practical takeaway
If a documentary uses a synthetic version of a real living person, record why the depiction is being made, whether consent was sought, how clearly the audience will understand that it is reconstructed and whether promotional or commercial uses create additional likeness risk.
What to watch next
Watch whether Joel or his representatives challenge the film or its promotion and whether documentary festivals begin adopting specific disclosure requirements for synthetic historical reconstruction.
Sources
A new lawsuit asks an even more fundamental question about what entered an image model’s training data
A proposed class action filed in federal court in California alleges that xAI used known child-sexual-abuse imagery, including material depicting one of the plaintiffs, in data used to develop Grok’s image and video capabilities. The complaint also alleges that synthetic outputs subsequently posted publicly through X can feed back into training and improvement pipelines. xAI has not been found liable and the training-data allegations remain claims that must be proven.
Why visual creators should care
The case illustrates why training-data provenance is not only a copyright issue. Image datasets can contain material whose possession, processing and reuse may itself be unlawful, creating legal questions far beyond whether a copyright licence existed.
Practical takeaway
For organisations commissioning or deploying visual models, training-data governance should include source legality, consent and exclusion systems as well as copyright licensing. A dataset being technically obtainable does not make every file suitable for training.
What to watch next
Watch xAI’s response, what evidence the plaintiffs produce about the alleged dataset and whether the court permits the proposed class claims to proceed.
Sources
Val Kilmer’s AI performance shows what the same technology looks like when the likeness is licensed
As Deep as the Grave uses generative AI to recreate the late Val Kilmer for a substantial posthumous performance. Recent coverage has shown extended footage from the production. Unlike many synthetic-likeness disputes, the filmmakers say Kilmer’s children and estate approved and collaborated on the reconstruction, supplied archival material and were compensated. The filmmakers say they followed SAG-AFTRA principles around consent, compensation and collaboration.
Why visual creators should care
Placed beside the Billy Joel documentary, the difference is useful. Both productions create footage of a recognisable real person that was never photographed. One begins with estate approval and collaboration; the other involves a living subject who did not participate.
Practical takeaway
For posthumous or synthetic performer work, make the rights chain visible in the production record: who authorised the likeness, what source material was supplied, who approved the performance and how the estate or performer is compensated.
What to watch next
Watch whether distributors and insurers begin requiring evidence of digital-replica consent as part of standard delivery documentation.
Sources
Owning a Midjourney image does not mean the production stays private
Midjourney’s current Terms of Service say creators own the assets they generate to the fullest extent possible under applicable law, subject to several exceptions, including account requirements for companies above a specified revenue threshold. Separately, users grant Midjourney a perpetual, worldwide, sublicensable and irrevocable licence over material they upload and assets they create. Content is public and remixable by default, while Stealth Mode is limited to qualifying paid plans and still does not conceal generations made in shared public spaces.
Why visual creators should care
For an agency, output ownership and client confidentiality are separate questions. An image can belong to the creator while the reference photographs, prompts and generations remain publicly visible or broadly licensed to the platform.
Practical takeaway
Do not put confidential client material into a default-public workflow. Record the subscription tier, privacy setting and generation location, particularly where reference photographs or unreleased designs are involved.
What to watch next
Watch whether Midjourney introduces stronger business confidentiality controls or narrows the licence granted over private professional content.
Sources
HeyGen makes consent an ongoing obligation, not a one-time avatar upload
HeyGen’s current moderation rules require explicit consent from the person represented when a user creates a custom avatar. Creating avatars of real people, including public figures, without that consent is prohibited. HeyGen also says the depicted actor can request removal of their likeness from the service and that users are responsible for honouring removal requests for generated content they have published elsewhere.
Why visual creators should care
This is stronger than treating a selfie upload as permanent permission. It recognises that a synthetic likeness can continue producing new material long after the original recording session, so consent may need to remain connected to later uses.
Practical takeaway
For client avatars or synthetic presenters, store the original consent and define how withdrawal works. A project should be able to answer not only whether the actor consented on day one but what happens if they later revoke permission.
What to watch next
Watch whether avatar platforms build revocation into provenance systems so previously generated assets can be distinguished from new generations made after consent has been withdrawn.
Sources
