Affiliate Disclosure: Some links on this page are affiliate links. If you purchase through them, we may earn a small commission — at no extra cost to you. These recommendations are independent and based on our own research.
- Midjourney vs. DALL·E 3 2026: Which image AI for which job?
- Midjourney Prompt Parameters 2026: The Complete Cheatsheet
- Stable Diffusion Local Setup 2026: The Beginner's Guide
- DALL-E 4 vs. Midjourney v7 vs. Flux Pro 2026: The Big Comparison
- Midjourney vs. Flux Pro vs. DALL·E 4 2026: Which image AI for which job?
- Recraft vs. Ideogram 2026: Which image AI for logos and typography?
Short answer
Why AI images 2026 need a compliance mindset, not a lawyer in every project
Until late 2024 the rulebook for AI images was a minefield of contradictory decisions out of the US, Germany and the UK. 2025 brought long-awaited clarity. The CJEU decided in C-452/25 that AI outputs without substantial human creative input do not qualify for copyright protection, and the US Copyright Office had reached the same conclusion in its 2023 policy statement on registration practices. That settles the headline question for almost every marketing team. What it does not do is tell a brand owner whether the poster on the next campaign is legally safe, whether the product photo will pass Amazon’s review, or whether the licensed stock shot you commissioned last year now has a new twin generated by a competitor’s prompt.
This article is written for the people who have to make those judgment calls daily: in-house marketers, agency leads, freelance designers and small-business owners who hit “generate” and then ship. The perspective is that of a compliance lead with a lawyer’s ear — not a courtroom argument, but a set of rules you can run past a legal review in twenty minutes and ship with confidence. The focus is on commercial AI images 2026 and the four legal layers that decide whether an asset is safe: tool license, copyright, personality rights and trademark law. On top of all four, the EU AI Act transparency duty will apply to every image your company publishes into the EU market from August 2026 on.
If you take one thing away before we dive in: “commercial use is permitted” and “I own this image” are not the same sentence. Almost every legal problem that hits businesses in 2026 originates in that misunderstanding.
The EU legal landscape for AI-generated image rights in 2026
European copyright is harmonized through the InfoSoc Directive 2001/29/EC and a long line of CJEU case law, most importantly Infopaq, Painer and Cofemel. The common thread is a concept the courts call the author’s “own intellectual creation” — sometimes translated as a creativity threshold, or in Germany as the Schöpfungshöhe required under §2 UrhG. The test is not whether an image looks good. The test is whether a human made creative choices that leave a personal imprint on the work.
A prompt like “cat on a skateboard, dramatic lighting, cinematic” does not clear that bar. The prompter expressed an idea; the model made the pictorial decisions about composition, lighting, texture, pose, color grading and countless other details that would be treated as creative choices if a photographer had made them. The CJEU in C-452/25 reinforces that a text instruction, no matter how detailed, does not turn the generator’s pictorial decisions into the prompter’s own intellectual creation. The decision is narrower than some commentators suggested in the first week — it explicitly leaves room for protection where a human exercises meaningful creative control, for instance through extensive inpainting, manual compositing, detailed mask work or the careful integration of generator output into a larger photographic or illustrative piece. But the default for a one-shot generated image is clear: no copyright, no exclusive right, no injunction against a third party who ships a near-identical AI image tomorrow.
The practical consequence in the EU is that AI-generated image rights sit in a legal gap. The image is legally usable, but not legally defensible. It is not stolen — no one’s rights were violated in making it — and it is not owned either, which means a competitor who generates something similar cannot be forced to take it down. For solo creators this is usually fine. For brands that spend serious money on a visual identity, it changes how assets are built.
The US position is materially the same. Since the Copyright Office’s March 2023 statement, registration applications for purely AI-generated works have been consistently rejected. The Thaler decision confirmed the office’s line, and Zarya of the Dawn showed how a mixed work gets dissected: the human-authored text was registrable, the AI-generated images were not. For companies operating on both sides of the Atlantic, the practical rule is identical. Treat AI output as usable, non-protectable raw material until a human has meaningfully worked on it.
The four legal layers that decide whether commercial AI images 2026 are safe
The compliance question breaks cleanly into four layers. Each answers a different question, each has its own rulebook, and each can bite independently.
1. Tool license — the contract with your provider
The tool’s terms of service govern what you are allowed to do with the output. They are a contract, they take effect when you pay your subscription, and they do not grant copyright — they grant a usage right. The 2026 state of the major tools:
| Tool | Commercial Use | Special Limit |
|---|---|---|
| Midjourney | From Basic plan (10 $/mo) | Companies over 1 M $ revenue need Pro plan |
| DALL-E / ChatGPT | Yes, unrestricted | No use in moderation-violating contexts |
| Flux Pro (BFL) | Unrestricted | None (most liberal license) |
| Stable Diffusion (CreativeML Open RAIL-M) | Yes | No “harm” purposes |
The important point: the license grants you a usage right, not a copyright. You can do whatever the contract permits — nothing more, nothing less.
2. Copyright — almost always absent
In the EU (harmonized via the InfoSoc Directive) and the US, only personal intellectual creations with recognizable human individuality are protectable. A short text prompt does not have the creative depth that either jurisdiction requires. The exception applies when you substantially edit the output — manual retouching, combining with other works, precise positioning via inpainting, or integration into a larger human-authored piece. In that case the edit itself can become a protectable work, but protection covers only the edited part, not the raw generator output underneath.
3. Personality rights — applied to generated faces
Publicity rights (US) and §22 of the German KUG apply by analogy to AI-generated images that depict real, recognizable persons. The safe side is always generic subjects without a real-world reference. The risky side covers celebrities and politicians named in the prompt, private individuals whose real photo you fed as an image-to-image reference, and anything that crosses into deepfake territory, which is increasingly its own criminal category across EU jurisdictions.
4. Trademark law — a separate check for logos
An AI image can be registered as a trademark — both USPTO and DPMA have accepted such filings since 2024 — but two problems sit underneath the registration. Copyright protection is typically absent, so you can hold the mark while third parties generate near-identical images freely. Distinctiveness is tested strictly, and generic AI aesthetics are often rejected for being too close to existing marks or too close to stock genre. Priority counts from registration, not from generation. Combining AI ideation with human refinement is the pragmatic route: higher legal safety, real copyright protection on the human contribution, and a better chance of clearing the distinctiveness bar.
Midjourney commercial license: Standard, Pro and the revenue threshold
Midjourney’s terms are often misread, so it is worth unpacking them carefully. The relevant sections in the 2026 terms of service draw three key lines.
The first is the subscription threshold. You need a paid plan to use Midjourney output commercially at all. The free trial is not a commercial license, and neither is use of another user’s assets from the public gallery. From the Basic plan at ten US dollars a month upward, commercial use of your own generations is permitted.
The second is the enterprise threshold. Companies with more than one million US dollars in annual gross revenue are required to take the Pro plan or higher. This is a contractual threshold, not a legal one, but Midjourney enforces it — and “you did not have the right plan” is the kind of breach that turns a later infringement claim into a much uglier conversation. For agencies, the key question is whose revenue counts. The conservative reading, and the one Midjourney has confirmed in community AMAs, is the end client’s revenue, not the agency’s. An agency with 400k revenue working for a 10M client needs the Pro plan for that work.
The third is the asset ownership clause. Under the current terms, the subscriber “owns” the assets they create in a contractual sense — Midjourney transfers all the rights it has. But Midjourney also grants itself a broad license to reuse subscriber outputs for training, promotion and platform operation. Pro and Mega plans add a stealth mode that keeps generations out of the public gallery but does not opt out of training. For sensitive client material, that is a governance question you need to answer before you press generate.
A practical subtlety: the contractual “ownership” Midjourney transfers is only as strong as what Midjourney itself has. Because the raw generator output is generally not copyright-protectable, what you receive is a strong contractual usage right, not an enforceable copyright. This matters in disputes with third parties, where “I own this because Midjourney said so” is not a winning argument if the work itself is not a protectable work.
DALL-E commercial rights and the OpenAI terms
OpenAI’s image terms are the most generous at first glance and the most nuanced on second reading. Under the 2026 OpenAI terms of service, users own the output they generate through ChatGPT or the API, subject to compliance with the usage policies. Commercial use, resale and modification are all explicitly permitted, with no revenue threshold and no subscription gate — even free-tier output is commercially usable.
The nuance sits in two places. First, OpenAI is explicit that it does not claim copyright on output and does not warrant that users receive copyright either. The company effectively tells you what we have already established: the output exists in a legal status that depends on your jurisdiction’s copyright law, and OpenAI’s contract cannot change that. Second, the usage policies carve out a meaningful list of contexts — disinformation campaigns, content sexualizing minors, content targeting specific individuals for harassment, and several more — where output is not permitted regardless of its technical feasibility.
For businesses, the practical consequence is that DALL-E commercial rights are unusually clean on the contract side and unusually exposed on the protection side. You can ship quickly and cheaply, but you should treat DALL-E outputs the same way you treat Midjourney outputs on the copyright question: usable, not defensible, and worth enriching with human refinement before they carry a brand’s identity.
Flux, ComfyUI and the BlackForestLabs licensing stack
Flux is produced by BlackForestLabs and ships in several tiers with different terms. Flux Pro, the hosted commercial tier, grants the most liberal license of the major providers. Commercial use is unrestricted, including for logos and trademarks, and there is no revenue threshold. Flux Schnell and Flux Dev are the open-weights models; Schnell is released under Apache 2.0 and can be used commercially, while Dev carries a non-commercial research license that prohibits commercial use of its direct outputs unless the user has a separate commercial agreement.
This matters for teams running ComfyUI. ComfyUI itself is a graphical interface licensed under GPL-3.0; the license that governs your commercial risk is the model weights you load into it, not the interface. A ComfyUI workflow running Flux Dev is not a commercial pipeline. The same workflow running Flux Schnell or a commercially licensed Flux Pro endpoint is. Running SDXL or SD 1.5 weights in ComfyUI pulls you into the CreativeML Open RAIL-M license that applies to Stable Diffusion.
The liberal terms on Flux Pro are a genuine asset for brand work. Because the license explicitly permits use in logos and trademarks, the contractual layer is not a blocker — which does not solve the copyright problem, but it removes one compliance concern from the list.
Stable Diffusion copyright and the CreativeML Open RAIL-M license
Stable Diffusion models — SD 1.5, SDXL, SD 3 and derivatives — ship under the CreativeML Open RAIL-M license. It is a “responsible AI license” rather than a classic open-source license, and the distinction matters. The license grants broad use rights, including commercial use, and then attaches a list of use-based restrictions that the user agrees to when they accept the model weights. Those restrictions cover several categories of harmful use — disinformation, unlawful discrimination, exploitation of minors, certain medical and legal contexts without disclaimers, and related categories.
For commercial image generation, the practical restrictions are rarely a blocker. Normal marketing, editorial and product work falls comfortably inside the permitted zone. The license also propagates downstream: if you fine-tune an SD model and distribute the fine-tuned weights, the Open RAIL-M restrictions follow the new weights. This is a concern for companies building internal models and distributing them to partners, not for agencies generating campaign images.
Stable Diffusion copyright questions tend to cluster around a different issue: the training data lawsuits, most notably Getty Images v. Stability AI. These cases are ongoing into 2026, and while they target the trainer rather than the end user, a worst-case outcome — a court ordering specific models withdrawn or retrained — could leave you with a legacy pipeline built on a model you can no longer legally run. The mitigation is documentation, which we return to in the checklist section.
AI image license comparison: which tool fits which use case
Looking at the tools side by side, three patterns emerge for AI images for marketing legal risk profiles.
For straightforward editorial, blog and social content with no brand-critical status, DALL-E is the fastest path. The license is clean, the integration into ChatGPT workflows is tight, and the quality is entirely adequate for everything up to mid-tier campaign use. For higher-end marketing where you want fine-grained control and consistency across a campaign, Midjourney’s Pro tier gives you the license headroom for larger clients and a generator that still leads on aesthetic consistency. For product-facing work — e-commerce hero images, landing page visuals where reuse as a trademark or in a logo is on the table — Flux Pro’s liberal license is the least restrictive starting point, though the copyright problem still applies to its outputs.
For open-weights workflows and any scenario where running on your own hardware matters — strict data governance, offline processing, very high volume — Stable Diffusion variants and Flux Schnell give you the control you need with predictable license terms. The trade-off is that you inherit the compliance work: model documentation, license propagation, and responsibility for verifying that the specific checkpoints you run carry the licenses you think they do.
AI image copyright EU: the AI Act transparency duty from August 2026
Article 50 of the EU AI Act creates a transparency duty that applies to both providers of generative AI systems and to the users who deploy their outputs. From August 2026, commercially published AI images into the EU market must be labeled. Two paths satisfy the duty.
Option A: Automatic C2PA metadata
C2PA — the Coalition for Content Provenance and Authenticity — is the industry standard for embedding provenance information into media files. Midjourney, DALL-E, Flux and Adobe Firefly all write C2PA data into their outputs automatically. Your compliance obligation is therefore not “add metadata”; it is “do not strip the metadata that is already there”. The typical stripping points are aggressive compression in older image-editing software, social media platforms that re-encode uploads, and asset management systems that flatten metadata on ingest. You can verify a file by dragging and dropping it into the official verifier at contentcredentials.org/verify.
Option B: Visible labeling
For social posts, print ads, and any distribution channel where C2PA metadata will not survive, visible labeling is the fallback. A short text note in the caption — “AI-generated”, “Image source: Midjourney v7 / AI” or similar — is sufficient. A visible badge on the image edge works for display formats where captions are not available. The icons published by c2pa.org/logos are designed for this purpose. What is not sufficient: alt text, hidden metadata fields in HTML, or disclosures buried in a website’s imprint. The rule of thumb is that a reasonable viewer, seeing the image in its normal context, must be able to perceive the disclosure.
Providers of the generator systems have a separate, stricter obligation to build detection-ready markers into their outputs. That obligation does not transfer to you as a user — your duty is the labeling duty on publication. In practice that means your asset pipeline needs two things: a policy that preserves C2PA where possible, and a template that adds visible disclosure where it does not.
Real-world scenarios for AI images for marketing legal review
The social ad scenario
An agency ships a LinkedIn carousel for a SaaS client. Five slides, all Midjourney v7. The images depict generic people — a woman at a laptop, a team in a meeting room — none referencing real individuals. The license layer is clean, because the client’s revenue is 3M USD and the agency is on the Pro plan. The copyright layer says the images are not protectable, which is fine: the agency and client are not trying to stop anyone else from using a similar image. The personality rights layer is clean because there are no real persons. The trademark layer is not engaged. The AI Act layer requires the carousel to carry a visible AI disclosure or preserved C2PA metadata — LinkedIn re-encodes on upload, so a short caption line does the work. Total compliance cost: one line of caption text and a check that the Midjourney plan matches the client.
The Amazon product photo scenario
A small brand uses Flux Pro to generate a hero shot of its own product on a styled background. The image is legitimate in every legal layer — the product is the brand’s own, the background is generic, the license is unrestricted. The sharper question is Amazon’s platform policy, which in 2026 requires sellers to disclose AI-generated imagery in certain categories and disallows AI-generated imagery that misrepresents the physical product. The legal compliance is easy; the platform compliance requires the seller to read Amazon’s current rules and, where the category demands it, shoot a real photograph as the primary image and use AI-generated imagery only in lifestyle slots.
The editorial use scenario
A newsroom generates illustrations for a long-form article on climate policy. No real persons are depicted, the style is illustrative rather than photorealistic, and the images are captioned “AI illustration” under each figure. Editorial use is among the cleanest cases for AI image work: the asset is ephemeral, the copyright protection question is not critical, and the AI Act labeling duty is satisfied by the visible caption. The only remaining concern is source diligence — if the illustrations are used alongside photographs from a news agency, the editorial team needs a clear style rule that prevents mistaking AI illustrations for documentary photography.
Takedown risk and what happens if a tool provider loses in court
The scenarios that worry compliance leads most in 2026 are the ongoing cases against tool providers. Getty Images v. Stability AI, the class actions against Midjourney and the broader pattern of training-data litigation will produce judgments across 2026 and 2027. The most likely outcomes do not threaten end users directly. The less likely but non-trivial outcomes include injunctions forcing a trainer to withdraw a model, to retrain without the disputed data, or to disable a specific output style.
If that happens to a model you built a pipeline around, the takedown risk is not a legal obligation on you — you acquired a legitimate license when you used the tool — but a practical exposure. A brand that built a three-year identity around Midjourney v7 outputs has a business-continuity problem if Midjourney is forced to retrain and the new outputs do not match. The mitigations are the same mitigations that apply to any creative dependency: diversify the toolset for brand-critical work, maintain prompt and version logs that would let you regenerate equivalents in a successor tool, and keep layered source files (prompts, seeds, post-processing steps) rather than only the final composited asset.
For individual takedown risk — a third party claiming that your specific image infringes their work — the mitigations are simpler. Keep the prompt, keep the tool version, keep generation timestamps, and be able to show that your process did not use a specific reference image or artist’s name. Most of these claims resolve without litigation when the respondent can show clean process documentation.
Protective checklist for businesses before launch
Before a campaign, a product page or a brand asset goes live, the compliance pass below is the minimum. It is designed to fit into a normal marketing review and not to require a legal team for every image.
Confirm the tool license matches the project. For Midjourney, verify the plan level against the client’s revenue. For DALL-E, note the platform and usage-policy compliance. For Flux, confirm the tier (Pro for hosted, Schnell for open Apache work, not Dev for commercial use). For Stable Diffusion variants, document the model checkpoint and its license text.
Keep a prompt log for every shipped asset. Store the prompt, the tool version, the seed if available, the generation date, any post-processing tools and steps, and the final file. A simple spreadsheet is enough for most teams. This is the single most valuable piece of documentation in a later dispute, and it takes minutes to maintain.
On persons, stay on the generic side of the line. Use generic descriptors (“a woman, 40s, office”) rather than named references. If the creative brief requires a specific person, treat it as a shoot: written consent, release form, and a record of the consent attached to the asset.
On styles, avoid living-artist names in prompts. Public-domain artists (pre-1956 deaths in most EU jurisdictions, subject to national rules) are safe on the copyright side. Style transfers from living artists are a live litigation question that is worth avoiding for commercial work.
Preserve C2PA metadata through your pipeline. Audit your export, compression and CMS steps once — most teams find one or two points where metadata is silently stripped — and fix them. For channels where metadata will not survive, prepare a visible disclosure template that the content team applies consistently.
Plan the AI labeling duty from August 2026 on. The social templates, print disclaimers and product-page copy need to include the disclosure in the standard form. This is a content-operations task, not a legal one, and it is cheapest to bake into the templates once.
For logos and trademarks, layer in human refinement. Generator output is the ideation stage; a designer’s rework produces both the copyright protection and the distinctiveness a trademark office will accept. The same rule applies to icon sets and any other asset you need to defend against copycats.
Back up original prompts for critical projects. If a tool provider is forced to retrain, your prompt logs are the fastest route to regenerating equivalents in a successor tool. Treat the prompt as a source file, not a temporary artifact.
Put the AI clause into client contracts. A short paragraph covering usage permissions, license passthrough, confirmation that no real persons are depicted without consent, and a statement about AI labeling keeps the legal layer clean between agency and client. Most EU-based legal templates now include this clause by default; if yours does not, adapt one in.
For sensitive client material, choose a tool with a training opt-out. Midjourney Pro and Mega offer stealth mode with documented opt-out on Pro and above, OpenAI offers API-level opt-out, and Flux Pro on the enterprise tier offers similar controls. Consumer tiers often feed input back into training, which is a non-starter for confidential client assets.
What to do when in doubt
Three pragmatic rules carry most of the load.
First, “commercially permitted” is not the same sentence as “I own it”. When an asset carries brand weight — a logo, a repeated visual motif, a packaging illustration — plan for human refinement. The cost of adding a designer to the chain is small; the cost of discovering after launch that the hero image is not defensible is large.
Second, faces are always delicate. For campaigns with real models, shoot real models. For scenarios where a model is not realistic, use generic AI faces without references to real individuals, and document that the prompt contained no real-person reference. The deepfake category is growing criminally across the EU, and the cost of a mistake here is much higher than on any other layer.
Third, document everything. Prompt logs, tool versions, generation dates, consent forms, license receipts. In a dispute, provability is gold. The teams who handle AI image incidents well in 2026 are not the ones with the best lawyers — they are the ones who can answer every question about provenance within an hour.
If your project has more than ten thousand euro in creative spend per individual AI asset, or if the asset is a logo or core brand mark, the ten thousand euro rule of thumb kicks in: get an individual legal review. Five hundred euros of legal consulting at the design stage beats a later injunction or a trademark office rejection by a wide margin.
Which strategy actually carries you through 2026
AI images in 2026 are legally usable, rarely protectable and soon mandatory to label. For solo creators this is mostly a non-event: generate, keep the C2PA metadata alive, add a visible disclosure where the channel strips metadata, ship. For agencies and brands, the work sits in three places — prompt documentation, tool license review and human refinement on brand-critical assets. Those three disciplines are the difference between a clean creative production and a court-ready process record when something goes sideways. The EU AI Act pushes the whole market toward transparency from August 2026 onward, and the teams that treat “AI-generated” as a quality signal rather than a disclaimer will come out ahead. The legal layer is simpler than the headlines suggest; the operational discipline to keep it simple is what separates the organizations that ship confidently from the ones that discover their exposure the hard way.
Sources and further reading
Legal references and licence claims rely on the primary sources: the EU AI Act full text (Regulation 2024/1689) for Art. 50 transparency, Midjourney Terms of Service for the commercial licence and OpenAI Terms of Use for DALL·E outputs.
Hub: AI Image Generation 2026 – Market Overview & Workflow. Deeper reads: DALL-E 4 vs. Midjourney v7 vs. Flux Pro – the comparison, Midjourney Prompt Parameters — the Cheatsheet, Stable Diffusion local setup — the beginner’s guide.
Update note (as of 21.04.2026)
This guide is reconciled every 4–6 weeks with new CJEU/BGH rulings, EU AI Act implementing acts and tool ToS changes. Particular attention in 2026: further clarifications around the C2PA standard, German BNetzA enforcement decisions in the image context and possible stock-platform policy changes. Next review: early June 2026.
Related articles
Our central articles on Artificial Intelligence at a glance — sorted chronologically.
practice-use-casesAI for Small Businesses 2026 — 7 Use Cases with Concrete ROI
Which AI applications pay off for small and medium businesses in 2026? Seven proven use cases with time savings, cost and implementation effort.
guides-tutorialsAI Image Generation 2026: Market Overview, Models and Pro Workflow
The complete overview of AI image generation in 2026: Midjourney v7, Stable Diffusion, DALL·E 3, Flux and Ideogram compared — with workflow recommendations and license notes.
guides-tutorialsAI Audio Tools 2026: Speech Synthesis, Transcription and Dubbing Overview
The complete overview of AI audio tools in 2026: speech synthesis (TTS), speech recognition (STT), voice cloning and dubbing — with tool recommendations, pricing and GDPR guidance.
guides-tutorialsPrompt Engineering 2026 – The Complete Guide for Professional AI Use
The practical guide to Prompt Engineering 2026 — techniques, frameworks, examples and templates for ChatGPT, Claude and Gemini.
Frequently Asked Questions
Can I commercially use AI-generated images in 2026?
Yes — all major tools (Midjourney from Basic plan, DALL-E via OpenAI ToS, Flux Pro unrestricted) allow commercial use of their outputs. BUT: 'commercial use' doesn't automatically mean 'copyright ownership'. Two consequences: you can sell, but third parties may reuse your AI image without a license.
Do I own copyright in an AI-generated image?
In the EU and US: mostly no. The CJEU confirmed in C-452/25 that AI output without significant human creative input is not a 'personal intellectual creation' under EU copyright law. The US Copyright Office has said the same since 2023. Output becomes protectable only when you substantially edit it (manual retouching, composition with other elements).
How do licenses differ between Midjourney, DALL-E and Flux Pro?
Midjourney: commercial use from Basic plan (10 $/month), unrestricted for companies under 1 M $ revenue. DALL-E (OpenAI): free commercial use incl. resale, but outputs are explicitly not copyrighted by OpenAI. Flux Pro (BlackForestLabs): unrestricted commercial use, incl. logos and trademarks — the most liberal license.
Can I use AI images as logos or trademarks?
Legally possible — but risky. Problem 1: Without demonstrable human creative input, no copyright protection — your 'logo' can be copied by anyone. Problem 2: Trademark law tests distinctiveness — AI-typical aesthetics can be too similar to existing marks. Recommendation: AI as starting point, then human refinement for protectability and distinctiveness.
What if the AI image shows real people?
AI-generated faces can violate personality rights if they recognizably depict real persons (celebrities, but also private individuals). Publicity rights apply (US) and §22 KUG (Germany). Rule of thumb: deepfakes and recognizable real persons only with written consent. Generic faces ('a woman, 40s, office') are fine.
Must I label images as AI-generated?
Yes — from August 2026 it's mandatory under EU AI Act Art. 50. Commercially used AI images need embedded markers (C2PA standard) OR visible labeling on the image. Midjourney, DALL-E and Flux integrate C2PA metadata automatically — you just need to not remove it. Visible labeling: 'AI-generated' or a symbol in the caption.
Can I prompt in the style of specific artists?
Gray area. 'In the style of Van Gogh' is legally okay (Van Gogh died 1890, public domain). 'In the style of Greg Rutkowski' (living artist) is legally unclear — US courts tend toward 'fair use' on the output, but style copying is potentially actionable. Safest: no living artist names in prompts.
What about AI model training on my data?
When you upload client material to an AI tool (e.g. for Flux ControlNet), check the ToS for whether material is used for training. ElevenLabs, OpenAI (API) and Midjourney (Pro and above) offer opt-out. On consumer tools (Midjourney without Pro), your input becomes part of the training base — problematic for sensitive client assets.
