Copyright and Licensing for AI-Generated Design

A four-factor matrix for designers asking “can I commercially use this?”, grounded in Australian copyright context.

Friday night, 11pm. The deck ships Monday. The hero image is AI-generated. The contract said nothing about AI. Tomorrow morning’s question, asked in the shower: can I actually use this commercially?

The answer depends on four things:

1. Which tool generated it.
2. What subscription tier you used.
3. What jurisdiction the work will be sold or displayed in.
4. Whether the output inadvertently replicates a copyrighted work (the implicit factor that sits with you under most tool TOS).

There is no general rule. This article gives you the matrix.

Most coverage of AI-IP either says “consult a lawyer” or hand-waves. We’ve found neither answers the shower question. So we built a four-factor matrix you can run yourself in five minutes, with one clear escalation gate when stakes warrant it.

Compliance-first isn’t fear-first. Many designers respond to AI-IP uncertainty by either avoiding AI entirely (overcautious) or ignoring the question (under-cautious). The professional-practice frame is neither.

Treat AI-generated assets like stock photography. Know the licence. Document the source. Tell the client if it changes their risk profile. Designers have done this for decades with stock; AI adds new variables to a familiar process.

The Graphic Artists Guild Handbook of Pricing and Ethical Guidelines, 17th edition (MIT Press, November 2025), covers generative AI substantively. Three principles apply directly: disclose to clients how work was made, use tools you can competently operate, and don’t claim ownership you don’t have.

This article is professional practice guidance for designers. It is not legal advice. For high-stakes commercial use (bet-the-company campaigns, rights-managed franchise work, public-figure likenesses) get a lawyer. For everyday client work, the matrix below is enough.

Factor 1: Training data and tool provenance. What a model was trained on directly determines what its output can legally be used for. Models trained on licensed data (Adobe Firefly’s stock library, Getty’s commercially-trained model) carry lower risk. Models trained on web-scraped data without permission (most consumer image generators) carry higher risk. Some split the difference (Stable Diffusion’s open weights with opt-out workflows). Bottom line: the same prompt produces outputs with different commercial clearance depending on which tool you use.

Factor 2: Subscription tier. Almost all major tools gate commercial-use rights behind paid tiers. Free-tier outputs frequently come with non-commercial restrictions or require attribution. Indemnification (the tool’s promise to defend you if a third party sues) is gated higher still, typically on Premium or Enterprise plans only. Adobe Firefly is currently the only major tool offering meaningful IP indemnification at retail subscription levels: Premium ($4.99/mo), Creative Cloud Enterprise tiers, and Firefly Business API users get a published standard liability cap of US$10,000 per output (per Adobe’s Generative AI Product Specific Terms, June 2025). Indemnification covers output-replication claims (third-party alleges your generated output infringes their work) but not training-data infringement claims.

Factor 3: Jurisdiction. Australian fair dealing is narrower than US fair use. The Copyright Act 1968 lists specific exceptions: research and study, criticism or review, parody or satire, news reporting, professional advice, and disability access. Each has defined limits (research allows 10% or one chapter for literary works; criticism requires attribution). US fair use applies a flexible four-factor test that judges apply case-by-case. Australia is strict and prescriptive; the US is flexible. For AI-assisted commercial work in Australia, most everyday client work falls outside fair dealing, so the question becomes one of licence, not exception.

Factor 4: Inadvertent replication. Both Midjourney and Adobe place liability on you if a generated image happens to replicate a copyrighted work (a celebrity likeness, a trademarked logo, a recognisable artist’s style). Most designers skip this factor. Models trained on copyrighted work can reproduce visual patterns from training data closely enough to fail the substantial similarity test US courts use. How to avoid it: reverse-image-search outputs before using them in high-stakes commercial work. Two minutes now beats a takedown notice later.

Tool	Commercial use	Training disclosure	Indemnification	AU notes	Verdict
Midjourney	Paid tiers; subscriber owns	Limited disclosure	None	TOS update 2026-02-24; track Disney/NBCU/DreamWorks/Warner litigation (case 2:25-cv-05275)	conditional
DALL-E / gpt-image family (OpenAI API)	Yes	OpenAI training-data disclosures	None at retail	DALL-E 2/3 retired May 2026; gpt-image-1 / 1.5 / 2 succeed	conditional
Adobe Firefly	Yes	Trained on licensed Adobe Stock + public-domain corpus	Yes, Premium and Enterprise tiers; standard cap $10K USD per output	Strongest position for AU commercial use	yes
Stable Diffusion (open-source variants)	Depends on weights / fine-tune licence	Variable	None; you carry the risk	TBD per variant; SDXL CreativeML OpenRAIL-M differs from Flux Schnell vs Pro licensing	conditional
Runway	Paid tiers	Partial	None at retail	Video-focus; same Midjourney-style risk profile for stills	conditional
Krea	Paid	Limited	None	Newer tool; track TOS evolution	conditional
Ideogram	Paid	Limited	None	Strong for typography (relevant for text-in-image work)	conditional
Magnific	Paid	Built on third-party models	None	Risk profile inherits from underlying model	conditional

The headline reads cleanly: Adobe Firefly is the only mainstream tool that puts its lawyers behind your output. That’s the moat. Everything else asks you to carry the risk yourself.

A note on the Stable Diffusion row. SDXL, Flux Schnell, and Flux Pro each have distinct licence terms. If you ship open-source-model output on commercial work, read the actual licence for the specific weights you’re using; do not assume one row covers them all.

A note on Midjourney. Disney, NBC Universal, DreamWorks, and Warner Bros. are suing Midjourney over copyright infringement (case 2:25-cv-05275, Central District of California, filed June–September 2025). The case is ongoing with a status conference scheduled for 31 August 2026. Midjourney’s defence rests on fair use. The outcome will likely affect how Factor 1 applies to Midjourney and other tools. If you’re considering Midjourney for high-stakes work, track the case progress.

The Australian Copyright Act 1968 (Cth) governs AI-generated work. As of May 2026, no amendments specifically addressing AI authorship have been enacted; reform runs on two parallel tracks. The Copyright Amendment Bill 2026, which passed the House of Representatives on 31 March 2026 and is proceeding through the Senate, addresses orphan works (Australia’s first statutory framework for works whose copyright owner cannot be identified or located), remote learning provisions, and technical Copyright Act amendments. The Bill does not substantively address AI authorship.

The AI-copyright reform track is the Copyright and Artificial Intelligence Reference Group (CAIRG), established 5 December 2023 by the Attorney-General’s Department as a standing consultative body. CAIRG advises the government on copyright challenges from AI. The government’s October 2025 announcement (26 October) explicitly rejected a text-and-data-mining exception and committed to further consultation through CAIRG; a CAIRG meeting on 28 October 2025 discussed licensing models for copyright material in AI training. Submissions to the consultation closed 11 December 2025.

For designers, watch CAIRG for AI-specific guidance, not the Copyright Amendment Bill. CAIRG recommendations won’t arrive before late 2026. The process is consultative, so treat current AU rules as stable through the rest of this year.

Australian case law on AI-generated work is minimal. Most precedent is American (Thaler v Perlmutter on AI authorship, Andersen v Stability AI on training-data scraping, Kashtanova on registering AI-assisted work). Australian courts haven’t ruled yet. Don’t assume clarity where none exists.

Article 2 of the EU AI Act defines material and territorial scope. The Act applies to any provider placing AI on the EU market regardless of provider location, any deployer of AI located in the EU, and any AI system whose output is used in the EU even if the system was created outside it. Article 50 mandates the transparency obligations: providers and deployers must inform users that they are interacting with AI and label AI-generated content. Article 50 takes effect August 2026, twenty-four months after the Act’s entry into force.

The Act draws a line between providers (companies selling AI systems) and deployers (people using them). An Australian designer using ChatGPT or Midjourney is a deployer; the vendors carry the heavier obligations. But Article 50 transparency requirements apply to you too: if you produce AI-generated content for an EU client, you must label it as AI-generated.

For Australian designers: a campaign produced in Australia but deployed in the EU falls under the Act. Ask clients upfront: “Is this going to the EU?” If yes, you must disclose AI use regardless of where you made the work. Disclosure is straightforward; non-disclosure is not.

This is not a reason to avoid AI for EU work. It’s a reason to disclose.

Pull-out callout: a structured template designers can lift directly. The template formalises the conversation; it does not replace legal advice.

AI-Asset Disclosure to Client: Template

For the [project name] deliverable, the following assets were produced using AI tools:

1. Tool used: [Midjourney v7 / Firefly / gpt-image-1.5 / etc.]
2. Subscription tier: [Free / Pro / Enterprise]
3. Commercial-use clearance: [Yes, under tool TOS / Conditional, see notes / N/A]
4. Indemnification: [Yes, covered by tool / No, designer carries risk]
5. Asset usage: [Hero image / supporting graphic / reference only]
6. Reverse-image-search performed (recommended for commercial work): [Yes, no replication issues found / No, see notes]
7. Jurisdiction of use: [AU only / AU + EU / AU + US / Global]
8. Notes: [any client-specific risk flags]

This disclosure forms part of the project record. Designer warrants reasonable diligence; client retains final responsibility for usage decisions.

The template makes three things automatic: (a) the conversation happens before the work ships, not after; (b) the project record exists if a question arises later; © the responsibility line is drawn before either party can be tempted to redraw it under pressure.

Knowing what you can legally use is a professional fundamental. The four factors are vocabulary. Master them and copyright assessment becomes a five-minute checklist with one clear escalation gate, the same way typography vocabulary or colour theory does.

TGDS Verdict. If a client pays for work, they deserve to know how it was made and what risks you do and don’t carry.

Cert IV in Design covers professional-practice modules including the contractual and IP material this article touches. The Cert IV pathway is the foundational route for designers building the full vocabulary, ethics included.

Design@Work is the applied route for working professionals wanting a refresh on contemporary AI-IP and the disclosure pattern.

Outbound links: Client Disclosure (the conversation: how to bring this up with a client). When AI Gets It Wrong (what to flag when an output has a problem). The Design Advantage (the why behind this article’s what).