86% of creators use generative AI.

Generative design is the default now

Generative AI is no longer a side experiment in creative work. 86% of creators actively use creative generative AI, according to the Adobe Creators' Toolkit report, a survey of more than 16,000 creators in the US, UK, France, Germany, South Korea, Japan, India and Australia, run with The Harris Poll in September 2025. It is a vendor survey, and Adobe sells these tools, but the direction matches what every agency and in-house studio reports.

The same survey carries the warning. 69% of creators worry that their content could be used to train AI without their permission. The people who use generative tools most are also the people whose work trained them.

86% of creators use creative generative AI; 69% worry about training without permission (Adobe and Harris Poll, Oct 2025)

For a business buyer, the question has moved. It is not whether designers will use image and video generators, but under which licence, with which training data behind them, and with what label on the output.

The courts have not settled the rights

Two cases frame the legal uncertainty. In the UK, the High Court ruled on 4 November 2025 in Getty Images v Stability AI. Getty had dropped its primary copyright claims during the trial; the secondary infringement claim was dismissed, and Getty won only on some trademark points that the judge described as extremely limited in scope, as Cleary Gottlieb's case note records. The model maker largely won, but on a narrowed case.

In the US, Disney and Universal sued Midjourney in June 2025 in a California federal court. They allege that Midjourney trained on their characters and reproduces them on request, cite more than 150 works and seek up to $150,000 per infringed work. The complaint describes Midjourney as having 21 million subscribers and $300 million in revenue the year before.

Sources: AP, Cleary Gottlieb, ABC News; dates as reported.

Licensing is the new route

Where courts are slow, contracts move fast. On 11 December 2025 Disney announced a $1 billion investment in OpenAI and a licence for more than 200 animated characters, with costumes and props, for Sora and ChatGPT, ABC News reported. The same company suing one image generator licenses its characters to another video generator. The message to the market is clear: licensed generation is acceptable, unlicensed generation is a legal risk.

For brands this changes the buying checklist. A design tool is only as safe as the rights behind its training data and the indemnity behind its output. Some providers train on licensed or owned libraries and offer indemnities; others train on web-scale data and offer none. Both kinds appear in the Good AI Prism, and the difference is a decision fact, not a footnote.

The data behind the pictures

Many image models trace back to a small number of open datasets. The Good AI Prism Registry records them with their licences. The LAION-5B registry card lists 5.85 billion image-text pairs from 2022 and the cleaned Re-LAION-5B release of August 2024 with about 5.5 billion pairs under Apache 2.0; the publisher does not recommend it for production products. ImageNet's 1.28 million training images are licensed for non-commercial research and education only. Common Crawl's dataset card shows the web source under much of it, at more than 300 billion pages.

That is why the question "what was this model trained on?" belongs in every design-tool tender. A licence on the tool does not clean the data underneath it.

Labels are required almost everywhere

Synthetic images and video now need a label in most major markets. Under Article 50 of the EU AI Act, applying since 2 August 2026, deployers who publish deepfakes must disclose that the content was artificially generated or manipulated, and providers must mark synthetic output in a machine-readable format. Generative systems already on the market before 2 August 2026 have until 2 December 2026, under the Digital Omnibus on AI. From the same date, the Omnibus prohibits AI systems that generate non-consensual intimate imagery of identifiable people or child sexual abuse material.

2 Dec 2026 machine-readable marking deadline in the EU for generative systems already on the market

Other markets moved earlier. China's labelling measures have applied since 1 September 2025, with explicit labels and metadata on images, audio and video. South Korea's AI Basic Act has applied since 22 January 2026. India's amended IT Rules have applied since 20 February 2026, with prominent labels, permanent metadata and a three-hour takedown duty. The ETHORITY analysis of synthetic faces and voices compares the regimes.

The practical consequence is a production question, not a legal footnote. Marking has to survive export, compression and re-editing, so it belongs in the pipeline, not at publication.

Generation moves onto the desk

The hardware shift makes governance harder. Work that needed a data-centre GPU in 2024 now runs close to everyday machines. Apple's M5 chip offers more than four times the peak GPU compute for AI of its predecessor, and Counterpoint expected AI-capable PCs to exceed half of global PC shipments in 2026, as the ETHORITY synthetic media analysis records. When image and video models run locally, the central licence register no longer sees every tool. A designer can install an open model on a laptop in minutes. That is good for speed and privacy, and it means the inventory has to include local models, plug-ins and browser tools, not only the subscriptions that procurement signed.

Synthetic media is also a risk channel

The same building blocks lower the cost of impersonation. Cyber incidents remain the top business risk in the Allianz Risk Barometer 2026 at 42%, and AI has climbed to second place at 32%. Fake video calls and cloned faces of executives are already a fraud pattern. Design and brand teams that produce synthetic presenters should agree consent and verification rules with the security team.

In practice

Use case: licensed characters. In December 2025 Disney licensed more than 200 animated characters to OpenAI for Sora and ChatGPT, while it litigates against Midjourney over unlicensed output. For a brand team the lesson is contractual: character, likeness and brand work goes only to tools whose licence and indemnity cover it, in writing.

Scenario: a campaign before the marking date. An in-house studio plans a December campaign with an image generator bought in 2025. Before 2 December 2026 it exports a test file through its usual compression and editing chain and checks that the machine-readable mark survives. If it does not, the campaign moves to a tool that passes, and the result goes into the tool register.

What to do Monday

1. List every generative tool. Record each image, video and design tool in use, with its licence, training-data statement and indemnity.
2. Check marking for 2 December 2026. Confirm how each tool marks output in a machine-readable way, and whether the mark survives export and editing.
3. Require documented consent. For any synthetic presenter, cloned face or likeness, keep written consent from the person depicted, including staff.
4. Prefer licensed generation. For brand and character work, choose tools trained on licensed or owned data, with an indemnity in the contract.
5. Measure cost and rights. Track cost per asset and rights incidents, the two outcomes the Good AI Prism Global Market Study proposes for this area.

The Media & Entertainment topic area follows the same rights questions into newsrooms and archives.