Can You Use AI-Generated Images Commercially in 2026?

The download button is not a legal clearance certificate. It feels like one: you type a prompt, a picture appears, an arrow invites you to take it, and thirty seconds later it's on your product page. But three separate questions hide behind that click, and most guides answer only the one that flatters whichever tool wrote them.

So: can you use AI-generated images commercially in 2026? Almost always, yes. Do you own what you made? Often not, and that surprises people. Can you still get sued? Only over a short, specific list of things, and we'll name every one of them. Three questions, three answers, no megaphone.

Yes, you can use AI-generated images commercially (almost always)

The direct answer first. You can put AI images in ads, on packaging, in client work and on product listings if three things are true:

Your tool's terms allow it. The paid tier usually does; the free tier often doesn't.
The image is clean. No real brands, no real faces, no franchise characters.
The destination allows it. Etsy, Amazon KDP and the big ad platforms accept AI images, with disclosure rules attached.

That's the whole test. The rest of this piece explains why each check exists, where people trip, and which paperwork actually matters, in the order the questions come up: use, own, get sued.

A licence is not a copyright

A licence is a private contract: the tool's terms of service decide whether you may use and sell what you generate. Copyright is public law: it decides whether you can stop other people copying it. You can hold a full commercial licence to an image that carries zero copyright, which means you can sell it, and so can a competitor who generates something nearly identical tomorrow morning.

Think of it as renting a flat versus holding the deed. The tenancy agreement (the licence) lets you live there, host dinner, hang pictures. Only the deed (copyright) lets you stop the landlord offering an identical flat to someone else. Midjourney's own terms admit the cap, granting ownership of your assets "to the fullest extent possible under current law". They grant what they can. The law decides the rest.

The practical comfort: for the everyday commercial uses that make up most real workloads (social posts, site graphics, ads, listings), ownership never comes up. You need a valid licence and a clean image, nothing more. Copyright only matters when you need exclusivity: a mascot, a book cover, a logo you'd defend in court. Use freely. Just don't assume nobody else can make something similar.

Related: budgeting for those paid tiers? See our guide to AI image API pricing.

A split-screen digital illustration set against a solid black background, featuring a vibrant color palette of deep black, hot pink, magenta, and soft warm orange. The left panel shows a close-up interior view of a hand reaching in from the left edge, wearing a dark sleeve, inserting a glowing neon pink key into a door lock. The door is dark wood, equipped with a peephole and handle. Through the open doorway or in the background is a dark urban scene with apartment buildings featuring illuminated windows glowing orange, and a potted plant in the lower right corner. The right panel displays a transparent glass display case resting on a dark rectangular pedestal. Inside the case stands a cream-colored document with visible lines of text, an icon at the top, and a red circular seal with ribbons at the bottom right. A dark padlock is latched onto the right side of the case. Both panels are framed by thin, glowing pink and orange lines with futuristic geometric UI elements, dotted sequences, and small stylized planets orbiting with rings in the upper left and lower right corners. The rendering style is clean 3D vector art with smooth gradients, sharp edges, and strong neon lighting effects casting a futuristic, high-tech mood.

What US law says now: no human author, no copyright

The US position stopped wobbling this year. On 2 March 2026 the Supreme Court denied certiorari in Thaler v. Perlmutter (No. 25-449), leaving standing the D.C. Circuit's March 2025 ruling that a work generated autonomously by a machine, with no human author, cannot be registered. That question is closed.

The more useful question, how much human input is enough, got its framework from the US Copyright Office's Part 2 report on 29 January 2025: "prompts alone do not provide sufficient human control", but human selection, arrangement and editing can be protected. The registrations bear that out in both directions. Kristina Kashtanova's Zarya of the Dawn kept copyright in its text and in the arrangement of its images, while the Midjourney panels themselves were disclaimed. And the Office has already registered at least one AI-assisted image, "A Single Piece of American Cheese", on 30 January 2025, on the strength of human selection, coordination and arrangement.

The live test case is Allen v. Perlmutter in Colorado. Jason Allen made "Théâtre D'opéra Spatial" with more than 624 Midjourney prompts plus Photoshop and upscaling work, and the summary-judgment motions were fully briefed by February 2026 and still pending as of August. Nobody, including us, knows exactly where the line will land. Allen will draw it.

Two footnotes for anyone selling abroad. China's Beijing Internet Court went the other way in Li v. Liu (27 November 2023), holding a Stable Diffusion image copyrightable because of the user's prompt and parameter choices. And the UK still has a statute protecting computer-generated works with no human author, though the government's March 2026 copyright report proposes removing it. Same image, three legal systems, three answers.

The working rule: the more you steer, the more you own.

Will the training-data lawsuits land on you?

This is the question behind the question. People read a headline about a model trained on scraped art and assume the liability trickles down to whoever pressed generate. As of August 2026, it doesn't. Look at the three biggest fights:

Getty Images v. Stability AI (UK). Judgment on 4 November 2025: Getty dropped its main training claims because the training happened outside the UK, the secondary copyright claim failed because the model's weights never stored Getty's images, and the only win was a trademark finding over Getty watermarks in some outputs, which Mrs Justice Joanna Smith called "both historic and extremely limited in scope". She put that one on Stability, not the user. One caveat: Getty was granted permission to appeal the secondary-infringement point on 16 December 2025, so this isn't the last word.
Bartz v. Anthropic. A $1.5 billion settlement, finally approved on 20 July 2026, covering 482,460 certified works at roughly $3,000 each. Read what it was for: pirating books as training inputs. Nothing in it touches people using model outputs.
Disney, Universal and Warner Bros. v. Midjourney. Filed 11 June 2025, consolidated that November, and the one to watch, because it targets outputs: images of Elsa, Darth Vader, the Minions and Homer Simpson, plus a demand for output filtering. If any case reshapes user risk, it's this one. It hasn't yet, and we'd rather flag it as live than pretend it's settled.

The summary: no court has held an ordinary user liable simply for using an AI output. Your risk is what you publish, not what the model ate.

The content that actually gets people sued

Copyright anxiety points at the wrong target. The claims that reach real users come from three places, and none of them involves training data.

Trademarks. An output showing a real logo or brand, used prominently in your own advertising, can infringe. Incidental background use is lower risk; building a campaign around it is asking for a letter.

Real people. Likeness law grew teeth. Tennessee's ELVIS Act, in force since 1 July 2024, protects a person's name, image, likeness and voice against AI simulation, and reaches the tools as well as the users. California added AB 1836 for digital replicas of deceased performers (from 1 January 2025) and AB 2602 for performer digital-replica contracts. Generating a recognisable person for commercial use is the fastest route to a demand.

Characters and named styles. Prompting a franchise character, or "in the style of" a living artist, produces exactly the outputs the Disney suit is built on.

Do one thing today: strike brand names, celebrity names and franchise characters out of your saved prompts. Boring prompts, quiet inbox.

Tool-by-tool terms (mind the free-tier trap)

Terms differ by tool and, more sneakily, by tier. Free tiers are the supermarket cheese cubes: you're welcome to taste, you're not welcome to resell. Here's the version we'd tape above the desk.

Tool	Commercial use	Do you own the output?	The catch	Indemnification
OpenAI (GPT Image / DALL·E)	Yes	Yes, per the terms	Copyright Shield covers API and Enterprise only, with carve-outs; free and Plus tiers get no cover	Enterprise and API
Google Gemini / Imagen	Yes	Google claims no ownership	Every output carries an invisible SynthID watermark; a visible watermark sits on the free web tier	Vertex AI, enterprise, conditions apply
Midjourney	Paid plans only	"to the fullest extent possible under current law"	Companies over $1M gross annual revenue need Pro or Mega; images are public unless you pay for Stealth Mode; free and trial images are non-commercial	None
Recraft	Paid plans only	Paid: yes, with privacy	Free-tier assets are owned by Recraft, sit in a public feed, and carry no commercial licence	None
Leonardo	Paid plans only	Paid: yes	Free-tier creations are public and licensed CC BY-NC, so non-commercial	None
Black Forest Labs FLUX	Depends on the model	Varies by licence	[schnell] is Apache-2.0; [dev] is non-commercial for self-hosting; the Pro API is commercial; a June 2025 licence revision muddied the output question	None
Adobe Firefly	Yes	Yes	Trained on licensed and stock content; indemnity excludes beta features	Qualifying paid and enterprise plans
Canva	Yes	Mostly	Not where the output modifies licensed content; stripping C2PA metadata is forbidden	None

Three rows deserve a second read. Midjourney's $1 million revenue rule catches growing companies by surprise, and public-by-default galleries catch everyone else. Recraft's free tier is the textbook trap: the platform owns the asset and you have no commercial rights at all. FLUX splits by model file, and its June 2025 licence revision left self-hosters arguing in the forums about whether their outputs were commercially usable.

Indemnification is the quiet dividing line for client work. Adobe, Microsoft, Google's Vertex AI and OpenAI's enterprise tiers will defend qualifying customers against IP claims over outputs, conditions attached; consumer tiers, Midjourney and Stability won't. If a client's procurement team asks who pays when an image bites, that single question shortlists your tools for you. It's also why we built JammyJar to route across multiple models rather than marry one: the right tool changes with the stakes.

Related: the free tiers and their catches get a full article of their own in our guide to free AI image generators.

A close-up product photograph features a modern dark-finished wooden mousetrap angled diagonally across a solid pitch-black background in a 16:9 landscape composition. The rectangular wooden base of the trap displays a visible dark brown wood grain texture with subtle purple undertones. Mounted on the base is a metallic mechanism including a tight coil spring and a swinging kill bar made of polished dark metal reflecting faint pink light. Resting squarely on the sensitive metal bait pedal on the right side is a small, square gift box wrapped in a vibrant glowing raspberry-pink paper and tied with a neat matching ribbon and bow on top. The gift box emits a strong neon-like pink luminescence that casts a bright glow onto the surrounding wooden surface and the adjacent metal components. The lighting is dramatic and directional, creating sharp highlights along the metallic edges and casting soft shadows across the wooden platform. The overall mood is sleek, editorial, and stark, contrasting the dark moody background with the vivid electric pink focal point.

Selling on Etsy, Amazon and print-on-demand

Marketplace policy, not copyright law, is what actually takes listings down. The current rules:

Amazon KDP requires you to disclose AI-generated text, images and translations (cover art included) via a checkbox at upload. AI-assisted work, where you made it and used AI to refine it, needs no disclosure. Enforcement tightened through 2025 and 2026, including retroactive reviews and removals.
Etsy wants the AI use disclosed in the listing and the human designer credited. It has removed non-compliant listings in bulk, and compliant sellers report reinstatement within days on appeal.
Adobe Stock accepts AI images with disclosure and bans prompts that name artists, real people or copyrighted characters.
Print-on-demand platforms like Redbubble, Printful and Merch by Amazon allow AI designs, increasingly scan metadata, and still hold you responsible for third-party IP.

The rule we'd print on the wall: three green lights before anything goes on sale. Your tool's licence permits it. The image is clean of other people's IP and faces. The destination allows AI content, disclosed the way it asks. Green, green, green, publish.

An editorial digital illustration rendered in a clean dark mode style set against a solid deep black background. On the left, a rounded rectangular user interface card displays a large preview image of a hooded figure standing on a wet futuristic street beneath towering skyscrapers and a pink moon. Above the preview, three glowing circular status indicators appear in green, amber, and red. Below the preview, stylized line placeholders and pink action buttons are visible, alongside a row of four smaller thumbnail previews. On the right, a matching storefront booth structure is built from dark matte materials outlined with glowing pink neon strips along its edges, awnings, and base. The booth features an arched roof topped with a glowing pink icon, and a vertical sign hangs from its side with the same icon. Inside and on the counter of the booth are several framed digital art prints depicting fantastical and sci-fi scenes, illuminated by warm overhead spotlights. The lower front panel of the booth also incorporates a horizontal panel with three glowing status lights matching the ones on the interface card. Small stylized star and planet shapes float in the dark void surrounding both the card and the booth, connected by thin dotted lines. The color palette consists primarily of deep black, hot pink, magenta, glowing green, and warm amber orange, creating a sleek, futuristic, and nocturnal mood.

The EU AI Act is live, and it asks less of you than you'd think

The AI Act's Article 50 transparency duties became enforceable on 2 August 2026, and no, the Digital Omnibus didn't delay them. Regulation (EU) 2026/1744 pushed only the high-risk deadlines, leaving one carve-out for images: machine-readable marking for systems already on the market before 2 August is deferred to 2 December 2026.

What lands on you depends on your role. Providers (the model companies) must mark synthetic outputs in a machine-readable way. A small business that merely uses AI images is a deployer, and a deployer's disclosure duty is narrow. It covers deepfakes and certain public-interest content; routine marketing images sit outside it. One wrinkle: if you embed a third-party image tool in your own product, the transparency notice has to appear on your surface.

Provenance is becoming plumbing either way. Google reports that SynthID has invisibly watermarked over 100 billion images and videos (their figure, so hold it loosely), C2PA Content Credentials became an ISO standard (ISO/IEC 22144), and OpenAI joined C2PA and added SynthID in May 2026. Removing the visible watermark from your own Gemini image is generally permitted under Google's terms. Stripping C2PA metadata is a different act: Canva's terms forbid it outright, and it can collide with the EU marking rules and marketplace policies. Leave the plumbing alone.

How to reduce your risk in ten minutes

None of this needs a lawyer on retainer. It needs six habits:

Steer and edit. Iterate prompts, select, arrange, retouch. Human authorship is what makes any of it yours.
Keep the receipts. Save prompts, versions and reference images. A dated trail answers "how was this made?" in seconds, whether the asker is a client, a marketplace or a regulator.
Clean the prompt library. No brands, no celebrities, no franchise characters, no living artists' names.
Check the hero assets. Run a reverse-image search and a quick trademark check before an image becomes a logo, a cover or a campaign face.
Match the tier to the stakes. Free tier for drafts, paid tier for anything sold, indemnified enterprise tools when a client contract demands cover.
Disclose where asked. Tick the KDP box, label the Etsy listing, follow the ad platform's rules.

That second habit is the one we built JammyJar around: every prompt, variation and reference image stays organised in one jar, so your evidence trail assembles itself while you work. If anyone ever asks how a visual was made, you can show them. Try JammyJar and keep the whole trail in one place.

FAQ

Can I copyright an AI image?

Not if it's purely AI-generated. The US Supreme Court's cert denial in Thaler on 2 March 2026 left the human-authorship requirement standing. But AI-assisted works are registrable where human creativity shows, as with the selection and arrangement in Zarya of the Dawn. The more you direct, edit and combine, the stronger your claim.

Can I sell AI art on Etsy?

Yes. Etsy allows AI-generated items if you disclose the AI use in the listing and credit the human designer. Non-compliant listings have been removed in bulk, and compliant sellers report reinstatement within days on appeal. The design itself must still avoid third-party trademarks, franchise characters and real people's likenesses.

Do I have to disclose AI images?

The obligation comes from platforms rather than one general law. Amazon KDP requires disclosure of AI-generated content at upload, Etsy requires it in the listing, and Adobe Stock at submission. Under the EU AI Act, an ordinary business using AI images must disclose deepfakes and certain public-interest content, not routine marketing visuals.

Is Midjourney allowed for commercial use?

On any paid plan, yes: commercial rights are included. Two conditions matter. Companies earning over $1 million in gross annual revenue must be on the Pro or Mega plan, and images are public by default unless you pay for Stealth Mode. Free and trial images carry a non-commercial Creative Commons licence.

Can I use AI images in paid ads?

Generally yes, provided your tool's licence permits commercial use and the image avoids real brands, real people and copyrighted characters. Platforms add their own layer: Meta, Google and TikTok require disclosure for political or realistic AI content in paid placements, and FTC endorsement rules still apply to whatever the ad claims.

Start with the button

Back to that download button. It still isn't a clearance certificate, but now you know what is: a licence that covers your tier, an image free of other people's brands, faces and characters, and a destination whose rules you've read. Use is a contract question. Ownership is a law question. Getting sued is a content question.

Tonight's first move takes ten minutes: open the terms page for the tool you actually use, find your tier, and read the commercial-use clause. If it isn't there in plain words, that's an answer too.