Minnesota’s AI nudification ban survives its first major federal-court challenge
A US federal judge has refused xAI’s request for a preliminary injunction blocking Minnesota’s new law restricting AI-generated fake nude images. The law took effect August 1 and prohibits website operators, software developers and others from allowing users to use AI to 'nudify' images of identifiable people without consent. xAI argues that the statute violates the First Amendment. US District Judge Donovan Frank said the constitutional issues are complex but concluded that xAI had not shown the immediate harm required to stop the law while the case proceeds. xAI says it will appeal.
Why visual creators should care
This moves synthetic-likeness regulation beyond takedowns after publication. The law targets the capability provided by image-generation services themselves, which means visual-AI companies can face obligations around what kinds of transformation they allow users to perform on photographs of real people.
Practical takeaway
For any workflow creating realistic sexualised or intimate synthetic imagery of an identifiable person, consent needs to be treated as a prerequisite rather than a post-production rights check. Platform availability does not establish lawful permission.
What to watch next
Watch xAI’s appeal to the Eighth Circuit and whether other US states adopt comparable laws aimed directly at AI image-generation and nudification services.
Sources
Instagram is again calling genuine photographs AI-generated
Photographers, creators and brands have reported Instagram automatically attaching its visible 'AI Content' label to photographs they say were neither generated nor materially edited with generative AI. The Verge found examples associated with minor operations such as Canva background removal or small blemish fixes, while testing also showed genuinely generated imagery could sometimes pass without a label. Meta had already revised its AI-labelling system after similar problems in 2024.
Why visual creators should care
A false AI label can affect the perceived authorship and authenticity of professional photography. For editorial, commercial and documentary creators, a platform effectively telling viewers that a genuine photograph is synthetic can undermine trust even when the photographer has done nothing deceptive.
Practical takeaway
Keep the original capture and editing history for important images posted to social platforms. If a genuine photograph receives an incorrect AI label, source files, RAW captures and edit records provide much stronger evidence than arguing from the finished JPEG alone.
What to watch next
Watch whether Meta changes its labelling logic again and whether it explains more clearly which C2PA, IPTC or platform-specific signals trigger the visible AI Content badge.
Sources
Fake portraits of real missing people show why synthetic identity needs more than an AI badge
TikTok removed a post containing AI-generated portraits of four Singaporeans who were missing following floods affecting Nepal and Tibet. The fabricated images were presented alongside genuine photographs of three other missing people and had accumulated more than 150,000 views. TikTok said the post violated its rules on edited and AI-generated media because users were misled into believing the synthetic portraits were genuine.
Why visual creators should care
This is not simply a generic misinformation problem. The generator created plausible photographic likenesses of identifiable real people at a moment when accurate visual identification mattered to families, journalists and the public.
Practical takeaway
Do not generate replacement portraits of real people for editorial, missing-person or documentary contexts unless the synthetic nature and purpose are unmistakable and the necessary permission exists. A realistic substitute photograph can easily become mistaken for an authentic record.
What to watch next
Watch whether news and social platforms introduce stronger rules specifically for synthetic representations of victims, missing people and participants in unfolding emergencies.
Sources
New research suggests tracing an AI image back to one training work gets harder as the dataset grows
Research from Zheng Dai and David Gifford at MIT examines what they call 'attribution decay'. Their method removes individual training items, or broader groups of material, and compares what a model would have generated without them. In sufficiently large datasets they found that removing a particular item may sometimes have little measurable effect on the generated result. The work was published in Nature Communications in August and received new visual-arts coverage on September 4.
Why visual creators should care
For artists pursuing training-data or output-infringement claims, technical influence and legal copying are not necessarily the same question. But the research shows how difficult it may become to prove that one particular source image caused a particular output when a model has absorbed a vast visual dataset.
Practical takeaway
Do not assume an output-comparison test will always reveal whether your work entered training. Preserve evidence of collection, dataset membership, licences, opt-outs and model documentation separately from any visual similarity analysis.
What to watch next
Watch how courts and expert witnesses use attribution or data-ablation evidence in visual AI litigation, particularly where plaintiffs challenge training itself rather than only a closely similar output.
Sources
A creator can own their identity while licensing a platform to build and monetise an AI version of it
Fanisin’s current Creator Licensing Agreement says creators retain rights in their original content and persona while granting the company rights to create AI-powered avatars using their likeness, voice and persona. The agreement permits authorised creator material to be used for the creator-specific avatar and allows derivative avatar output. It separately says shared platform models are not trained on Creator Content or avatar interactions without additional consent, while dedicated voice, video-avatar and knowledge-base models are scheduled for deletion after termination.
Why visual creators should care
Synthetic identity agreements increasingly need to distinguish the underlying human identity, the training material, the reusable avatar model and the output generated by that model. Saying simply that the creator 'owns their likeness' no longer explains the whole commercial arrangement.
Practical takeaway
If licensing an AI twin, record separately who owns the original content and likeness, what material can train the avatar, whether other models can learn from those interactions, who controls generated output and what must be deleted when the licence ends.
What to watch next
Watch whether creator-avatar agreements increasingly include independent auditing of model deletion and more granular approval rights over new synthetic-video or advertising uses.
Sources
On YouTube, saying the video is AI-generated does not give you permission to use the person inside it
YouTube’s current likeness-claim system allows an enrolled person or authorised representative to make claims against altered or AI-generated videos matching their face or voice. A likeness claim can block a video. YouTube explicitly says that disclosing altered or synthetic content does not prevent a likeness claim and does not grant permission to use somebody’s identity. Creators disputing a claim can rely on grounds such as explicit permission from the depicted person.
Why visual creators should care
Disclosure and consent solve different problems. An AI label tells the audience how an image or video was made; it does not create the underlying right to synthesize somebody’s face or voice.
Practical takeaway
Keep the likeness licence or release with the project rather than relying on the fact that the synthetic video was correctly labelled. If a platform claim occurs, the useful evidence is permission, not merely disclosure.
What to watch next
Watch whether likeness-claim systems become interoperable across major video and social platforms and whether synthetic-talent licences can be linked directly to provenance credentials.
Sources
