Higgsfield’s new terms make the training question a client-work question
Higgsfield’s revised Terms of Use and Privacy Policy say users retain ownership of their prompts, inputs and outputs and may use outputs commercially. But the company also says standard-user content can be used to improve its models. Deleting an account stops that use going forward. Higgsfield separately says enterprise agreements can provide that client data is not used for training. The new terms apply immediately to people registering from July 26 and on August 27 to existing users unless they accept them earlier.
Why visual creators should care
Ownership and training permission are separate. A photographer can continue owning a client portrait while the service contract still allows the file or resulting generation to contribute to model improvement. That distinction matters for confidential campaigns, unreleased products and identifiable people.
Practical takeaway
Before putting client material into Higgsfield, decide whether standard-account training terms are compatible with the assignment. Record that decision in the project file. Where the client prohibits training or secondary model use, use an agreement that expressly excludes it rather than assuming private visibility means no training.
What to watch next
Watch how Higgsfield implements training controls when the terms become operative for existing users on August 27 and whether non-enterprise users receive any granular opt-out.
Sources
An AI poster followed the rules, won the prize — and caused the rules to change
An AI-generated poster won the Ohio State Fair’s 2026 poster competition and a $1,000 prize. The entrant did not break the rules: the competition had permitted AI since 2024 when its use was disclosed. Following substantial criticism, fair organisers say AI will be prohibited from the poster contest next year.
Why visual creators should care
This is a useful example of disclosure not necessarily resolving the fairness question. The entrant complied with the published rules, yet the institution concluded afterwards that AI-assisted and conventionally created work should no longer compete in the same category.
Practical takeaway
For awards and competitions, save a copy of the eligibility rules as they existed on the submission date and disclose AI exactly as requested. Do not assume that an organisation’s acceptance of AI this year predicts next year’s policy.
What to watch next
Watch whether other photography, illustration and design awards move towards prohibition, separate AI categories or a human-contribution threshold rather than disclosure alone.
Sources
A photography prize draws the line at the lens rather than banning AI outright
The 2026 OD Photo Prize, which closes on August 11, accepts projects that use AI or generative tools as part of the process provided that use is declared. Fully machine-generated imagery without a lens-based foundation falls outside the prize’s scope. Entrants retain ownership while granting limited rights for exhibition, press and promotional use.
Why visual creators should care
This is a more nuanced definition of photographic authorship than either unrestricted AI acceptance or a total ban. It recognises that photographers may use generative processes while still requiring an underlying photographic practice and human authorship.
Practical takeaway
When submitting AI-assisted photography, document which parts were captured and which were generated. Keep the original lens-based files and describe the AI contribution clearly rather than relying on a generic 'AI-assisted' label.
What to watch next
Watch whether other photography prizes adopt a similar lens-based threshold and whether they begin asking for original captures or edit histories as evidence.
Sources
YouTube can search for your synthetic face, but the stolen voice is still your problem
YouTube’s current likeness-detection system lets eligible adult creators enrol using government ID and a short video selfie. YouTube then scans newly uploaded videos for potential matches involving the enrolled creator’s face. Matches can be reviewed and sent through the privacy-removal process. The system currently detects visual likeness only; YouTube says audio likeness detection is planned but not yet available.
Why visual creators should care
Recent creator-impersonation cases show that identity theft can preserve the voice, body, performance and script while altering only the face. A facial detection system therefore addresses only one component of a creator’s commercial identity.
Practical takeaway
Creators who regularly appear on camera should consider enrolling where the feature is available, but should not rely on it as complete monitoring. Preserve original video and audio and continue checking for unauthorised voice or performance reuse separately.
What to watch next
Watch the promised extension into audio detection and whether YouTube eventually treats face, voice and performance as one combined likeness record.
Sources
Magnific turns AI disclosure into part of the user contract
Magnific’s current terms expressly treat users of its AI products as deployers for EU AI Act purposes. Users are responsible for complying with transparency rules and must clearly disclose content that constitutes or could reasonably be perceived as a deepfake. The terms recognise a different disclosure approach for evidently artistic, creative, satirical or fictional work.
Why visual creators should care
The practical significance of Article 50 is now appearing inside the contracts creators click through to use visual tools. Compliance is therefore no longer only something a regulator may expect; it may also be an obligation owed directly to the platform.
Practical takeaway
For realistic generated people, products, places or events, decide on disclosure before publication and preserve the version containing the disclosure. Do not assume that the platform itself will automatically add everything required for your particular use.
What to watch next
Watch how Magnific implements machine-readable provenance and whether its output workflow differentiates clearly between artistic synthetic work and potentially deceptive realistic media.
Sources
Freepik says reference-image permission remains the user’s responsibility
Freepik’s current AI terms require users to have rights or authorisation for uploaded inputs and all necessary permissions from people whose image, voice or identity appears in them. Users remain responsible for infringement or likeness claims involving both input and output. Freepik also says it does not use uploaded images or voices to train its own AI models, while third-party model providers may process those files as service providers under Freepik’s data arrangements.
Why visual creators should care
Multi-model creative suites can make it easy to forget that uploading a portrait is itself a rights decision. A client may have approved publication of a photograph without approving the photograph or model’s likeness being sent through a generative system.
Practical takeaway
Add generative processing to your model and client permissions where identifiable people or confidential material may be uploaded. Record which third-party model was selected and do not treat a platform’s no-training promise as a substitute for consent to processing.
What to watch next
Watch whether Freepik exposes provider-specific data location and retention information more prominently at the moment a creator chooses a model.
Sources
