Getty is putting licensed human photography directly inside the AI workflow
Getty Images launched its Model Context Protocol server on August 12. The service gives AI assistants, agents and enterprise applications a standardised way to search, retrieve and use Getty’s licensed creative, editorial, news, sport, entertainment and archival photography and video. Getty says the system complements its existing APIs and can support functions including visual search and content download inside conversational and agentic workflows.
Why visual creators should care
AI does not have to generate an image to change the visual-production workflow. It can now act as the researcher or picture editor that chooses an existing licensed photograph. That creates an important alternative to generation: an AI assistant can answer a visual request with human-created, rights-managed material rather than synthesising a substitute.
Practical takeaway
If an AI agent selects Getty material for a campaign or publication, retain the Getty asset ID, licence, usage restrictions and source record. Do not treat the agent’s successful retrieval of a file as proof that your particular use is licensed.
What to watch next
Watch whether AI design and publishing systems begin presenting licensed stock and generated imagery side by side, and whether agents can reliably explain the licence attached to the visual they select.
Sources
Minnesota’s ban targets the AI tool itself, not only the person misusing the image
A federal court rejected xAI’s request for a temporary restraining order that would have prevented Minnesota’s AI nudification law from taking effect on August 1. The law prohibits operators from providing technology that generates fake nude imagery of real people. xAI argues that the measure is unconstitutionally broad and can restrict consensual or otherwise protected expression. The underlying constitutional challenge continues; the July 31 ruling concerned xAI’s request for emergency relief rather than a final decision on the law.
Why visual creators should care
Most synthetic-likeness rules focus on distributing a harmful finished image. Minnesota is testing a different model by placing duties directly on the provider of the image-generation capability. That distinction could significantly affect developers offering generative editing, face replacement and synthetic-person tools.
Practical takeaway
Do not assume that a user’s consent always resolves the platform-level legal question. Generative-image businesses operating across US states increasingly need to check laws governing the capability itself as well as the eventual output.
What to watch next
Watch the forthcoming proceedings on xAI’s First Amendment challenge and whether other states adopt provider-level restrictions rather than relying primarily on victim takedown rights.
Sources
Creators are sometimes removing AI labels because the label itself changes how their work is treated
New research interviewed 21 experienced AI-content creators and tested 6,000 images from six generation platforms through 16 manipulation workflows reported by creators themselves. The researchers found that creators sometimes remove visible AI labels because they fear reputational damage, platform moderation or reduced reach. They also found significant differences between transparency mechanisms: metadata-based signals such as C2PA and IPTC could disappear through cropping or recompression, while some invisible watermarks proved more resistant.
Why visual creators should care
Transparency rules often assume that creators have an incentive to preserve AI labels. In real commercial and social workflows the opposite may be true: if audiences or algorithms penalise labelled work, creators can have a direct incentive to remove the evidence regulators and platforms expect them to preserve.
Practical takeaway
Keep the original labelled output before resizing, cropping or exporting it elsewhere. If disclosure matters for a client or platform, preserve the evidence separately rather than assuming the label will survive the production chain.
What to watch next
Watch whether platforms separate AI-generated from AI-assisted work more clearly and whether disclosure systems become resistant to ordinary editing without creating new privacy concerns for creators.
Sources
Image provenance may need to prove less in order to protect the photographer
A new C2PA-compatible research proposal uses zero-knowledge proofs to hide sensitive provenance information while still proving selected facts about an image. Examples include demonstrating proximity to a location without revealing the exact coordinates, making privacy-preserving likeness claims using biometric embeddings and using perceptual image fingerprints to help recover provenance after metadata has been stripped.
Why visual creators should care
Detailed provenance can create its own risks. A photographer proving where an image was captured may not want to publish exact location data; a creator proving identity may not want to expose biometric information. Provenance systems therefore need ways to establish trust without turning every useful production fact into public metadata.
Practical takeaway
Treat provenance data as potentially sensitive project information. Do not automatically expose capture location, personal identity or biometric information simply because the file format can carry it.
What to watch next
Watch whether privacy-preserving disclosure methods move from research into C2PA tooling and whether newsrooms, photographers and platforms can verify useful claims without seeing the underlying sensitive data.
Sources
A watermark may tell us more about the ecosystem than about one disputed photograph
New research from Cornell argues that AI watermarks are often asked to perform a forensic job they may not reliably be able to do. A synthetic-content watermark must survive substantial alteration rather than simply indicate that a signed file has changed. The authors propose treating watermarks partly as ecosystem-level measurement tools: useful for understanding how synthetic content moves through platforms and communities even when they cannot conclusively identify every individual file.
Why visual creators should care
Creators, clients and regulators can easily treat a missing watermark as evidence that an image is human-made or a detected watermark as definitive proof of generation. Neither conclusion necessarily follows. A watermark is one signal inside a wider provenance record.
Practical takeaway
Do not clear or reject professional visual work on the basis of one watermark test. Combine it with originals, Content Credentials, generation records, photographer evidence and platform information.
What to watch next
Watch how Article 50 enforcement treats failed or removed machine-readable marks and whether regulators expect perfect individual detection or broader evidence that providers deployed effective marking systems.
Sources
Shrinking a photograph before AI detection may throw away the evidence the detector needs
New research called LHSDet targets high-resolution AI-generated imagery by combining low-level texture information, broader visual features and textual semantic analysis. The researchers argue that many existing detectors first downsample large images to fit their models, potentially discarding subtle texture and edge information useful for distinguishing generated images from camera photographs. In their tests, LHSDet outperformed the second-best comparison method by more than nine percentage points in high-resolution detection scenarios.
Why visual creators should care
Professional photographs often begin as large high-resolution files but are evaluated by tools designed around much smaller inputs. If the detection process destroys useful forensic evidence before analysis begins, photographers can face decisions based on an unnecessarily weak version of their original file.
Practical takeaway
If authenticity is disputed, preserve and analyse the highest-quality original rather than submitting only a social-media JPEG or resized export. Detector performance on the published copy may differ materially from performance on the source file.
What to watch next
Watch independent testing across current commercial image generators, genuine camera files and ordinary Lightroom or Photoshop processing. This is preprint research, not a production-ready authenticity standard.
Sources
