The Training Data Question You Can't Dodge
Every major image model was trained on billions of images scraped from the open web. Some of those images were copyrighted. Some were artists' portfolios. Some were medical records that should never have been there. Lawsuits are ongoing — Getty v. Stability, the Andersen class action, Concept Art Association complaints — and the legal answers will not be settled this year, or probably next.
You do not get to wait for the answers. You have a deadline.
Here is the honest framing: using these tools is currently legal in most jurisdictions, but "legal" and "uncontroversial" are different things. Artists whose styles were ingested without consent have a real grievance, and dismissing it makes you sound like a crypto bro in 2021. Acknowledge the mess, then make defensible choices.
The defensible move is this: do not prompt by living artists' names. "In the style of Greg Rutkowski" works because Rutkowski's work was scraped — and he has publicly asked people to stop. Same for Karla Ortiz, Sarah Andersen, Kelly McKernan. Using their names is technically possible and ethically gross. Use movements, mediums, decades, or deceased artists instead: "Art Nouveau poster," "1970s airbrush sci-fi paperback cover," "Mucha-inspired floral border." You get the aesthetic without conscripting a working artist into your moodboard.
Commercial Rights, Tool by Tool
Read the terms. Then read them again when they change, because they will.
As of writing, here is the rough shape:
- Midjourney: Paid subscribers own the images they generate and can use them commercially. Free trial users do not get commercial rights. Images on the public feed are visible to other subscribers — pay for Stealth if your client work cannot leak.
- DALL·E 3 / ChatGPT: OpenAI grants you ownership of outputs and permits commercial use, including resale. Microsoft Copilot images carry similar terms but bind you to Microsoft's content policy.
- Stable Diffusion / SDXL / Flux (open weights): The models ship under permissive licenses (CreativeML Open RAIL-M, Flux's non-commercial vs. dev licenses). Outputs are yours, but the license of the model you ran matters. Flux has a "dev" license that forbids commercial use of the model itself — you can still sell images you made with the commercial-licensed version. Check the specific checkpoint.
- Adobe Firefly: Trained on licensed Adobe Stock plus public domain. Marketed as "commercially safe" with an IP indemnification offer for enterprise customers. If a client demands a clean chain of custody, this is the easiest answer.
Two practical rules. First, save the license terms as a PDF the day you deliver the work — terms shift, and you want a record of what you agreed to. Second, if a client asks "is this AI-generated and are we covered?" the answer is never "probably." Either you can cite the indemnification clause, or you tell them no.
Deepfakes, Likeness, and the Line You Should Not Cross
Generating a recognizable real person — politician, celebrity, your ex, a classmate — in a fabricated scenario is the fastest way to turn a portfolio into a lawsuit or a news story. The legal exposure varies (right of publicity in California, the EU AI Act's deepfake disclosure rules, state-level NCII laws in the US), but the social exposure is uniform: you will be the villain.
Hard rules that have aged well so far:
- No real people in sexual, violent, criminal, or politically inflammatory contexts. Ever.
- No minors, generated or real-likeness, in any ambiguous context.
- No "satire" defense for images that look photographic. If a reasonable viewer might think it is real, treat it as real.
- For client work involving a real person's likeness, get a signed release the same way a photographer would.
Fictional characters belong to their rights holders. A "Pixar-style movie poster" featuring your dog is fun on Instagram and a cease-and-desist on Etsy. Disney, Nintendo, and the major studios are actively scanning marketplaces.
Disclosure: When, Where, How
Disclosure is a judgment call, not a binary. Use this ladder:
Always disclose:
- Journalism and editorial. Reuters, AP, and most newsrooms now require it.
- Academic submissions. Your university almost certainly has a policy — read it. "I used Midjourney for the figures" in a methods section is usually fine; passing AI art off as your own hand illustration is academic misconduct.
- Anything submitted to a contest or grant with rules about original work.
- Government, legal, or regulated industry communications.
Disclose by default:
- Client deliverables. Put it in the SOW: "Visuals produced using generative AI tools (Midjourney, Photoshop Generative Fill)." Surprises later are how you lose clients.
- Book covers, album art, and anything where buyers might reasonably care.
Disclosure optional:
- Personal social posts, memes, thumbnails for your own content, mockups, internal slides. Nobody needs a caption disclosing that your LinkedIn header was AI-assisted. If asked, do not lie.
A clean disclosure line for client-facing work:
Image: Generated with Midjourney v6, composited and retouched in Photoshop.
That sentence has saved more relationships than any lawyer.
A Defensible Workflow
Before you publish or deliver, run this checklist:
- No living artist named in the prompt. Movement, medium, era only.
- No recognizable real person unless you have a release.
- No protected IP (characters, logos, trade dress) in commercial work.
- Tool license matches the use — commercial subscription if commercial output.
- Disclosure decision made, written into the deliverable or the caption.
- Source files saved — prompt, seed, model version, date. If a dispute lands in your inbox in 2027, you want the receipts.
This is not paranoia. It is the same hygiene a photographer keeps with model releases and Lightroom catalogs — boring, fast once it is habit, and the difference between a career and an incident.
If you want the deeper frame — bias, accountability, how to think about AI's social cost beyond the image you are shipping today — work through /courses/ai-ethics-responsible-ai. The questions in this chapter are the surface; that course is the foundation underneath.
Make defensible choices, write them down, and move.