xAI is testing whether a state can regulate the image generator rather than only the harmful image
xAI has sued Minnesota in federal court over a law due to take effect August 1 that targets services offering AI nudification of identifiable people. The company accepts the state’s interest in combating non-consensual synthetic nude imagery but argues that the law is overbroad, regulates protected expression and lacks a sufficient safe harbour for services trying to prevent misuse. Minnesota’s approach differs from many existing deepfake laws because it reaches the owner or controller of the tool itself rather than focusing only on the person distributing the resulting image.
Why visual creators should care
For image-generation companies, this could become an important boundary case. If states can prohibit categories of generative functionality at service level, compliance may increasingly affect which visual-editing features are available in particular jurisdictions rather than merely what users are allowed to publish.
Practical takeaway
Studios and creators should not assume that a capability available in an AI tool is legally usable everywhere. Record which service and jurisdiction were involved, and keep consent evidence whenever a real person’s appearance is substantially transformed.
What to watch next
Watch whether the federal court blocks Minnesota’s law before or after it takes effect and how it distinguishes regulation of harmful conduct from regulation of the image-generation technology itself.
Sources
Google’s new terms turn several AI-rights questions into explicit platform rules
Google’s revised Terms of Service take effect July 30. They say users must not mislead others into believing generative-AI content was created by a human, must respect intellectual-property and privacy rights and must not use Google services or their content to violate those rights. The terms also prohibit automated access that ignores machine-readable instructions such as robots.txt rules disallowing crawling or training, and prohibit using AI-generated content from Google services to develop machine-learning models or related AI technology. Google separately states that it will not claim ownership over original content generated through services that support such creation.
Why visual creators should care
These provisions connect disclosure, source rights, machine-readable training restrictions and output reuse inside one of the largest creative and distribution ecosystems. For professional users, tool terms are increasingly part of the project’s rights environment rather than an administrative page nobody reads.
Practical takeaway
Archive the terms applicable when important client work is produced. Do not rely on an output being technically downloadable as evidence that every downstream use is permitted, particularly for model development, automated scraping or third-party material incorporated into the workflow.
What to watch next
Watch how Google applies these provisions across Gemini, image generation, Workspace and other services, particularly where service-specific terms impose additional rules.
Sources
A Photoshop automation deadline shows why the toolchain itself needs documenting
Adobe says a series of legacy Photoshop modules used through Workfront Fusion will no longer work after July 30 because underlying Photoshop API elements have been deprecated. Affected workflows include PSD editing, format conversion, composite creation, new PSD creation, renditions, text-layer editing, Photoshop actions, depth blur, product crop, layer information, Smart Object replacement, resizing and watermarking. Adobe provides replacement modules for most functions, but not all.
Why visual creators should care
Professional visual production increasingly depends on automated pipelines rather than a single person manually opening Photoshop. When an API changes, an entire client-production chain can fail even though the underlying PSD files still exist. Rights and provenance records are only useful if teams can also reconstruct which processing steps produced the delivered asset.
Practical takeaway
Audit automated Photoshop workflows now. Record which modules perform transformations, migrate legacy steps, preserve before-and-after test files and update production documentation so future teams know which process replaced the deprecated one.
What to watch next
Watch whether replacement modules produce materially different output or metadata and whether provenance and audit records survive migration between legacy and replacement workflows.
Sources
Copyright preflight is moving from generated stills into video
CopySight has launched CopyScore V2, extending its AI-generated-media risk analysis into video. The system scans frames for visual signals including recognised faces, logos, characters, iconic designs and other material that may create intellectual-property or likeness concerns. The company is also building production records around prompts, model information and human contribution. CopySight and its investors want this type of pre-publication check to become part of studio, insurance and commercial clearance workflows.
Why visual creators should care
Generated video can contain a problematic logo, character or recognisable person for only a few frames. Human reviewers may miss those details, particularly at commercial-production scale. Automated preflight could become useful, but a similarity score is evidence for review rather than a determination that something does or does not infringe.
Practical takeaway
Treat automated copyright scoring like a preflight warning. Save the result, the flagged frame and the human clearance decision. Do not tell a client that an asset is legally safe simply because an automated tool gave it a low-risk score.
What to watch next
Watch whether insurers, studios or distributors begin requiring automated IP-risk reports and whether independent testing establishes how reliably these systems distinguish legitimate inspiration, licensed material and actual copying.
Sources
Pangram is bringing AI detection directly to images
Pangram has released a research preview of Pangram Image, a detector intended to identify imagery from major generators including GPT Image, Gemini, Midjourney, FLUX and Grok Imagine, as well as frames produced by several AI-video systems. The dashboard also provides heatmaps for mixed human and generated imagery. Pangram says the current preview cannot reliably detect deepfakes or face swaps, cannot process images below 512 by 512 pixels and still has other research-stage limitations.
Why visual creators should care
False accusations of AI use are already a professional problem for photographers and artists, while provenance metadata does not always survive publication. Detection may provide another signal when metadata is absent, but the limitations make it unsuitable as a stand-alone authenticity verdict.
Practical takeaway
Use image detection as supporting evidence only. Preserve camera originals, Content Credentials and editing records alongside any detector result, particularly if authenticity may affect a competition, publication, commission or dispute.
What to watch next
Watch independent benchmarking, false-positive rates on heavily retouched photography and art, and whether the system develops reliable face-swap and deepfake detection before moving beyond research preview.
Sources
An AI film festival is making training-data rights part of the submission rules
The London AI Festival states that entrants must own the rights to everything in a submitted film, including training data unless it is public domain or properly licensed. Its rules also prohibit deepfakes of real people without written consent and require both human and tool contributors to be credited. July 30 is the festival’s notification date ahead of its August 7 screening.
Why visual creators should care
Most creative-industry AI policies focus on the final output. These rules push clearance further upstream by asking whether the underlying training or source material was legitimately available and whether real people agreed to synthetic representation.
Practical takeaway
AI filmmakers should build a clearance file before submission or distribution. Include source licences, training or reference permissions, performer releases, model/tool records and complete credits rather than trying to reconstruct them after a festival or buyer asks.
What to watch next
Watch whether larger festivals adopt similarly explicit requirements and whether organisers develop practical evidence standards for proving training-data rights rather than relying solely on entrant declarations.
Sources
