Europe’s AI-media rules move into the enforcement phase
The EU AI Act becomes broadly applicable on August 2, with the European AI Office and national authorities taking responsibility for implementation, supervision and enforcement. For visual creators and organisations, Article 50 requires providers of relevant generative systems to make generated or manipulated content identifiable through machine-readable marking. Deployers must also disclose the artificial nature of qualifying deepfakes and certain AI-generated public-interest material. Earlier systems covered by the EU’s July 2026 amendments may have until December 2 to meet specific technical marking requirements.
Why visual creators should care
This moves synthetic-media provenance from voluntary good practice towards regulated production infrastructure. Agencies, publishers and creators may have obligations of their own even when the generation platform already embeds a technical signal.
Practical takeaway
Separate the provider’s technical marking from the publisher’s visible disclosure. Test whether credentials or watermarks survive Photoshop, video editing, resizing, compression and social-platform upload. Keep evidence of the label and provenance attached to the final delivered asset.
What to watch next
Watch how regulators interpret realistic product images, generated people, de-aged performers and synthetic environments. Commission guidance indicates that deepfake treatment can extend beyond face swaps to realistic depictions of people, objects, places and events that may appear authentic.
Sources
California makes the generator responsible for carrying provenance into the file
The California AI Transparency Act becomes operative on August 2. Covered providers of public generative-AI systems must offer a free tool that allows users to assess whether image, video or audio was created or altered by the provider’s system and retrieve detected provenance information. The law also requires latent disclosures in generated content that are permanent or extraordinarily difficult to remove where technically feasible. Later stages extend additional obligations to large platforms and capture-device manufacturers.
Why visual creators should care
The law places part of the authentication burden on the company that made the generator rather than leaving photographers, publishers and audiences to determine origin after the file has circulated.
Practical takeaway
Where a covered generator offers a detection service, test the finished client export rather than only the original generation. Save the verification result with the project record. A positive origin result proves neither that the source assets were authorised nor that the depicted content is accurate.
What to watch next
Watch how California defines durable disclosure in practice and what happens when ordinary editing, transcoding or platform processing removes or weakens the signal.
Sources
OpenAI takes provenance beyond the image file and into verification infrastructure
OpenAI updated its provenance programme on July 31. Supported audio generated through ChatGPT and the OpenAI API now includes SynthID watermarking, and OpenAI’s public verification tool can check supported audio files as well as images. The company has also introduced verification API access, allowing developers and organisations to incorporate provenance checks into their own products and production systems. The approach combines C2PA Content Credentials, durable SynthID signals and verification tools.
Why visual creators should care
Verification no longer needs to be a person manually uploading one file to a public website. Newsrooms, agencies, archives and platforms can begin incorporating origin checks into intake, review and publishing pipelines.
Practical takeaway
Treat a provenance API result as an origin signal, not a rights clearance. Save the result alongside consent, copyright and licensing records. Test whether verification still succeeds after the media passes through the tools normally used in your production chain.
What to watch next
Watch which OpenAI image and audio products are supported, how the service handles edited or compressed files and whether verification responses expose sufficient information for professional auditing.
Sources
A court protects alleged deepfake victims from having litigation expose them again
A federal judge in San Jose has ruled that plaintiffs in a proposed class action involving alleged sexualised deepfake imagery generated through Grok may continue using pseudonyms. The plaintiffs argue that disclosure of their identities would cause further humiliation, harassment and exposure connected to the imagery at the centre of their claims. xAI had sought to require the plaintiffs to proceed under their real names.
Why visual creators should care
Victims of non-consensual synthetic imagery face a structural problem: asserting their rights can require repeating or publicising the very association they are trying to stop. Pseudonymity can make legal action more realistic for photographers, performers, models and private individuals targeted by generated intimate material.
Practical takeaway
Anyone documenting likeness abuse should preserve evidence without unnecessarily redistributing the harmful image. Use restricted evidence folders, retain URLs and platform records, and seek legal advice before publishing the material as part of a complaint or public campaign.
What to watch next
Watch whether the case advances towards substantive questions about generator safeguards and provider responsibility. The anonymity ruling protects the claimants’ identities but does not decide whether xAI is liable.
Sources
Microsoft’s image model moves deeper into ordinary presentation and enterprise production
Microsoft’s MAI-Image-2.5 family is available through Microsoft Foundry for text-to-image generation and controlled image-to-image editing. Microsoft says MAI-Image-2.5 is also in production in PowerPoint for image-editing workloads. The Pro and Flash variants give enterprise users different quality, speed and cost options while keeping generation inside Microsoft’s broader productivity and cloud environment.
Why visual creators should care
Synthetic imagery is moving out of specialist image-generation sites and into familiar business software. That makes it easier for generated or materially altered images to enter client decks, pitches, internal approvals and advertising concepts without a distinct handoff to an AI specialist.
Practical takeaway
Add image-generation disclosure and source-record fields to presentation workflows, not only to finished campaign production. Record whether a PowerPoint image was generated, edited or merely placed, and retain the source image when image-to-image editing is used.
What to watch next
Watch Microsoft’s provenance implementation, commercial-use guidance and model-version records as MAI-Image moves across PowerPoint, Foundry and other Microsoft services.
Sources
Bing Video Creator disappears while its old outputs remain only temporarily accessible
Microsoft has retired Bing Video Creator as of July 31. A notice in Bing Image Creator says existing creations will remain available for download through users’ creation histories subject to Microsoft’s storage policy. Microsoft says new image models are coming, but the notice does not present a direct replacement for the retired video feature.
Why visual creators should care
Tool retirement is now a routine part of AI production risk. A creator may retain the exported video while losing the interface, generation history and settings needed to reproduce or revise it.
Practical takeaway
Download any required Bing video outputs and preserve prompts, dates, source images and project notes outside the platform. Do not treat a cloud creation history as a permanent production archive.
What to watch next
Watch whether Microsoft introduces a replacement video product and whether creation histories retain model names, provenance and generation records after the tool itself has disappeared.
Sources
