Singapore has opened the AI-training copyright question back up
Singapore’s Ministry of Law and Intellectual Property Office launched a public consultation on August 26 on how copyright and patent rules should evolve around AI. The consultation includes questions about copyrighted material used for AI training, accountability and certainty for rights-holders. Singapore’s current Copyright Act contains a computational-data-analysis exception that can cover machine-learning training where conditions including lawful access are satisfied.
Why visual creators should care
This is particularly important for photographers, illustrators and stock contributors because Singapore already has a relatively broad statutory route for computational analysis. The policy question is now whether creators need stronger transparency, control or other safeguards around that use.
Practical takeaway
If Singapore matters to your licensing or publishing workflow, keep records showing where your images were published, what access conditions applied and whether any machine-readable training reservation accompanied them. A future reform may make those records more important.
What to watch next
Watch the consultation paper and responses for proposals on transparency, licensing, rights reservations and whether Singapore modifies the conditions under which copyrighted visual material may be used for AI training.
Sources
A photographer’s LAION case is about to test what a machine-readable opt-out is really worth
Germany’s Federal Court of Justice is scheduled to hear case I ZR 281/25 on September 3. The plaintiff is a photographer whose image was downloaded while LAION created an image-text dataset that can be used for generative-AI training. Lower courts held that the copying was covered by German text-and-data-mining exceptions. One central issue is whether a website restriction written in ordinary language was sufficient or whether the photographer needed a machine-readable reservation of rights.
Why visual creators should care
This is one of the clearest cases anywhere testing the practical value of AI-training opt-outs for photographers. A right to reserve machine learning use is much weaker in practice if the photographer also has to prove that the reservation was machine-readable when the image was collected.
Practical takeaway
Do not rely only on a sentence in website terms saying images may not be used for AI training. Where the law recognises machine-readable reservations, use technical signals as well and keep evidence showing when they were implemented.
What to watch next
Watch the September 3 hearing for the court’s treatment of machine-readable reservations, scientific-research exceptions and whether creation of an AI-training dataset qualifies for Germany’s text-and-data-mining exceptions.
Sources
Pika says building an AI version of yourself does not put your likeness into its general training pool
Pika’s current AI-self and agent documentation says personal data, training inputs and selfies supplied to create a user-specific AI agent are used for that experience and are not used to train other people’s Pika agents or general-purpose models. Pika also says its terms prohibit creating an account or agent using somebody else’s likeness without explicit permission and provides a reporting route for impersonation.
Why visual creators should care
Products that build reusable synthetic versions of a person create a much more sensitive data relationship than ordinary image generation. A creator may willingly make an AI avatar without intending their face to become training material for unrelated users or models.
Practical takeaway
When creating a synthetic version of yourself or a performer, distinguish the personalised model from general model training. Save the platform’s no-training statement and make sure the person whose likeness is used actually authorised creation of the synthetic identity.
What to watch next
Watch whether Pika introduces more granular controls over who may reuse an agent, how long likeness data is retained and whether generated performances can be exported to other platforms.
Sources
Krea is turning no-training protection into a business-plan feature
Krea’s current pricing and business offering advertises Business Terms of Service that include no-training clauses for customer content. Business users also receive private collaborative workflows, the ability to train LoRAs on large image sets and fine-grained controls over which AI models team members may use. Enterprise customers can negotiate custom terms and receive audit logs.
Why visual creators should care
For an agency, no-training protection is becoming something that can depend on the account tier rather than simply the name of the creative platform. Model-access controls also matter where some models have different contractual, training or client-approval requirements.
Practical takeaway
Record the account tier as part of the AI project record. If a client requires no training, make sure the actual workspace carries that contractual protection rather than assuming every Krea account gets the same terms.
What to watch next
Watch whether no-training protection moves into Krea’s standard plans and whether its model-level controls expose the terms, provider and data treatment attached to each underlying generator.
Sources
Freepik separates private AI processing from model training
Freepik’s current AI terms require users to own or license uploaded material and obtain necessary permissions for any person whose image, voice, identity or personal data appears in it. Freepik says it does not use user inputs or outputs to train its own AI models. Private inputs and outputs are treated as confidential unless the creator chooses to publish them, although processing may still occur where necessary to provide the service, enforce safety or comply with law.
Why visual creators should care
A no-training promise does not mean a provider never processes the file. For client work, creators need to separate training, hosting, safety review, service delivery and public sharing rather than treating all platform processing as one thing.
Practical takeaway
If you tell a client their images will not be used for AI training, quote the platform’s actual promise and keep the project private. Also confirm that releases cover the specific upload and generative use of identifiable people.
What to watch next
Watch whether Freepik extends equivalent guarantees to every third-party model available through its platform and whether creators receive clearer model-by-model data routing information.
Sources
Shutterstock’s strongest AI-image indemnity depends on a human looking at the image
Shutterstock’s current generative-AI offering distinguishes ordinary generated output from images that pass an optional human content review available through eligible enterprise plans. Unreviewed images can carry protection around Shutterstock’s authority to license the generated material, but Shutterstock does not generally indemnify claims arising from what the output depicts, including third-party trademarks, copyrighted material or publicity rights. After successful human review, broader protections can apply under the customer’s agreement.
Why visual creators should care
The generator itself is not the final rights check. Shutterstock’s model treats copyright, trademarks and likeness risk in the finished image as something that may still require human review before a company receives stronger legal protection.
Practical takeaway
If an agency relies on indemnification, keep the review approval with the delivered asset. Do not describe every image created on an indemnified platform as automatically cleared for copyright, trademark and likeness risk.
What to watch next
Watch whether other stock and visual-AI providers adopt human clearance as a separate paid layer and whether clients begin asking for evidence that generated campaign assets were individually reviewed.
Sources
