Europe now has investigators and confidential reporting tools behind its AI-media rules
The European Union has expanded its AI Office in Brussels with 38 additional staff as enforcement of the AI Act begins. The team will monitor providers ranging from startups to major US and Chinese companies and may request documentation or interview company personnel during investigations. The Commission has also launched confidential whistleblower and compliance tools through which workers and users can report suspected illegal conduct, including harmful deepfakes, sexually explicit synthetic imagery and other AI Act violations.
Why visual creators should care
The transparency regime is no longer merely a set of future technical requirements. It now has personnel, reporting channels and potential market-access consequences behind it. Visual generators, agencies and publishers should expect questions about how synthetic media was marked, distributed and governed.
Practical takeaway
Create an internal route for reporting suspected disclosure or consent failures before they reach a regulator. Preserve model records, labels, exports and approval evidence in a form that can be produced if a client, platform or authority asks how an image was made.
What to watch next
Watch which visual-media cases the AI Office takes up first and whether enforcement initially focuses on provider-level marking, undisclosed deepfakes, non-consensual intimate imagery or public-interest communications.
Sources
The difficult EU question is whether an invented product scene is a deepfake
European retailers have argued that AI-generated advertising should not automatically fall within deepfake disclosure rules when the work is not intended to mislead. Their examples include a generated living-room scene used to display a sofa and AI enhancement of product imagery. The disagreement has become more consequential now that the transparency duties apply, because commercial visual work often creates realistic people, locations and products without claiming to document a real event.
Why visual creators should care
A wide interpretation could bring ordinary advertising, e-commerce imagery, synthetic models and generated campaign environments within visible-labelling requirements. A narrow interpretation could leave audiences unaware that realistic commercial imagery was materially invented.
Practical takeaway
Until the boundary is clearer, record whether the image represents a real person, product, place or event and what an ordinary viewer is likely to believe. Where a synthetic element materially changes that belief, visible disclosure is safer than relying only on embedded metadata.
What to watch next
Watch for Commission or regulator guidance distinguishing artistic and commercial fabrication from deceptive deepfakes, particularly for synthetic fashion models, product visualisation and generated lifestyle photography.
Sources
Canva now treats provenance removal as a terms violation, not an optional workflow choice
Canva’s AI Product Terms, effective June 26, prohibit users from misleading people into believing AI-generated output is human-made. They also prohibit removing, altering or disabling provenance and metadata tags attached to generated material, expressly including C2PA metadata. Users remain responsible for their uploaded inputs and represent that they hold the rights, licences and permissions needed to use them.
Why visual creators should care
Canva sits inside everyday agency, marketing and social-production workflows. A designer who removes metadata during export or passes off generated imagery as entirely human-made may now be breaching the tool’s contract even before copyright or consumer-law questions are considered.
Practical takeaway
Do not deliberately strip Canva provenance from generated assets. Record the source photographs and other references uploaded to Canva, and make sure the person supplying them has the rights required for generative editing.
What to watch next
Watch whether Canva begins enforcing the provenance clause through export controls, account action or automated checks, and whether third-party publishing platforms preserve its C2PA data.
Sources
Midjourney’s new default changes the visual result even when the prompt stays the same
Midjourney made V8.2 its default image model on July 24. The company describes the update as focused on image quality, aesthetics and improved personalisation, with a larger image pool for building personalisation profiles. V8.1, previously the default, remains a distinct model with different speed, prompt interpretation and visual behaviour.
Why visual creators should care
A repeated prompt is not a reproducible production method when the underlying default model changes. Agencies may discover that a previously approved look, character or campaign treatment cannot be recreated simply by reopening the prompt.
Practical takeaway
Record the explicit Midjourney version rather than relying on the platform default. Archive approved outputs, seeds, reference images and personalisation settings, particularly where later campaign extensions or client revisions are likely.
What to watch next
Watch how long V8.1 remains accessible and whether Midjourney provides stronger project-level records showing which model and personalisation profile produced each image.
Sources
IPTC warns newsrooms that Content Credentials do not detect deepfakes on their own
IPTC has published a plain-language C2PA guide for broadcasters, publishers and news agencies. It addresses practical questions including implementation cost, signing tools, publisher identity, what happens when a content-delivery network strips metadata and whether C2PA can detect a deepfake. The guidance distinguishes provenance records from detection and explains the role of CAWG Identity Assertions in connecting media to a verified publisher.
Why visual creators should care
C2PA is increasingly discussed as though it proves that an image is true. In reality, it records claims about origin and editing history. It is strongest when connected to a trusted camera, creator or publisher identity and preserved throughout the production chain.
Practical takeaway
Do not describe Content Credentials as a deepfake detector. Use them to support a wider chain of trust that includes photographer identity, assignment records, camera originals, editorial verification and a known publisher.
What to watch next
Watch whether social networks and content-delivery systems preserve credentials reliably and whether verified publisher identity becomes a normal part of editorial-image distribution.
Sources
Adobe Stock makes contributors responsible for AI imitation, releases and metadata
Adobe Stock’s current intellectual-property guidance says contributors must own or license the copyright in submitted material and obtain necessary model or property releases. It prohibits unauthorised copyrighted works, logos, fictional characters and protected trade dress. The guidance also tells contributors to avoid AI-generated material that imitates known artists or brands and prohibits misleading titles, keywords and attributions. Repeated or serious violations can result in removal, suspension or termination.
Why visual creators should care
An image being accepted by an AI generator does not make it acceptable for stock licensing. The contributor remains responsible for the people, brands, artwork, architecture and visual identity represented in the file.
Practical takeaway
Run a separate rights check before submitting generated work to stock. Review references, artist-style resemblance, recognisable people, logos, packaging, locations and metadata rather than relying on the generator’s safety filter.
What to watch next
Watch how Adobe distinguishes prohibited artist imitation from broad genre or historical influence and whether enforcement becomes more consistent across images, video, vectors and illustrations.
Sources
