Overview
Midjourney is a text-to-image generator run by Midjourney, Inc., a company founded by David Holz that opened its beta in July 2022 and has never taken outside investment. Reporting puts its annual revenue somewhere near $500 million going into 2026, produced by a staff usually counted in the dozens rather than the hundreds, which makes it one of the most profitable AI products per head in the industry. It began as a bot you talked to inside a Discord server, now runs as a web app, and added a video model in mid-2025. As of September 2026 the plans run $10, $30, $60 and $120 a month, metered in fast GPU hours, with a small free allowance that comes and goes. The people using it are art directors, indie game studios, marketing teams, and a very large number of hobbyists who arrived because it produced better-looking pictures than anything else and asked almost nothing of them in return.
What It Knows About You
It takes your prompts, any images you upload as references, everything it generates for you, plus the usual account layer: username, IP address, contact and billing details, and whatever you type in public channels. The volume is modest next to an ad-funded platform. The disposition is the problem. On the $10 and $30 tiers, every prompt you write and every image you make is published to Midjourney’s public gallery by default, browsable by strangers, and it stays there. Privacy is a paid feature: Stealth Mode ships only with the $60 Pro and $120 Mega plans, and switching it on going forward does not retroactively pull down what you already made. The Terms of Service go further than storage. They grant Midjourney a perpetual, worldwide, irrevocable licence to reproduce, publicly display, distribute and make derivative works from the text and image prompts you submit, and that licence survives you cancelling the subscription. If you uploaded a photograph of a colleague as a style reference, you did not merely send it for processing. You licensed it, permanently, and so did they without being asked.
The Real Risks
The first risk is that the service may not be yours to keep. Disney, NBCUniversal and DreamWorks sued Midjourney in June 2025 over training on and reproducing their characters; Warner Bros. Discovery filed its own suit in the Central District of California on 4 September 2025, alleging systematic and willful infringement involving Batman, Bugs Bunny and Rick and Morty. The studios are not only asking for money. They want an injunction forcing copyright filtering, relief that on its face could suspend the product while it is built. Discovery in the Disney case was scheduled to close around September 2026 with summary judgment briefing behind it, and Midjourney spent mid-2026 fighting over whether the studios must disclose their own AI use. Separately, Midjourney is a defendant in Andersen v. Stability AI, the artists’ class action that is now pointed at what would be the first US jury verdict on image-model training. Nothing is decided. But a production pipeline resting on an active defendant is a planning problem, not a philosophical one.
The second risk is that the legal exposure runs downhill to you. Midjourney’s terms disclaim warranties on what it produces and offer no indemnification; the indemnity clause runs the other way, with users agreeing to hold the company harmless. Adobe indemnifies paying Firefly customers, which is precisely the comparison a commercial buyer should make. There is also a clause worth reading twice before a company adopts it: anyone at a business with more than $1 million in annual revenue has to be on Pro or Mega to own the assets at all.
Third, the outputs carry measured bias and Midjourney publishes no evaluation of it. Rest of World’s 2023 study found the model collapsing whole nationalities into a handful of stereotyped looks. A peer-reviewed analysis of 1,200 generated images of anesthesiologists found 83% depicted as White. An academic stereotype benchmark found skew in 46 of 60 tested prompt pairs, and separate work documented mishandling of cultural symbols in depictions of East Asian women. With a chatbot, a skewed answer gets read once. Here the output is the deliverable, so the skew ships into campaigns, books and games.
Fourth, this is one of the few AI products that substitutes for a specific job rather than a task inside one. The Society of Authors surveyed 787 UK members in January 2024: 26% of illustrators had already lost work to generative AI, 37% said their income had fallen in value, and 78% expected it to keep falling. A 2026 survey of 378 working visual artists reported overwhelmingly negative effects on their workplaces. Commissioning an image is now a decision someone makes against a $30 subscription.
Alternatives
- Adobe Firefly. Trained on a corpus Adobe actually licensed, and it indemnifies paying users. Rated C in this Index. You trade the copyright exposure for Adobe’s subscription gravity.
- Commission the illustrator. For one hero image, a flat-fee commission often costs less than a month of Pro plus the hours spent prompting toward something you can ship. You also get revisions from someone who read the brief.
- Licensed stock. Unglamorous, cheap, and the rights question is already answered in writing.
- A local open-weight model. Stable Diffusion derivatives run on your own hardware, so nothing leaves the machine and no gallery publishes your work. The training provenance is no cleaner, and setup is genuine effort.
- Use a photograph. A lot of published work needs a real picture or a clear chart, not a synthetic illustration in a house style readers have started to recognise on sight.
Our Verdict
D. The scores that drive it are job threat at 9 and the pair of sixes on privacy and bias. Midjourney is the clearest case in this Index of a tool whose commercial appeal is the direct removal of a paid role, and it offers the people whose work trained it neither compensation nor an opt-out, which is where it parts company with Firefly. The privacy score is not about tracking, because Midjourney does not follow you around the web. It is about a product where your work is public unless you pay a 6x premium, and where the licence you grant outlives your account. Lock-in stays at 5 because the exit is real: your outputs are ordinary image files you can download and take anywhere. Autonomy sits at 4 because nothing here decides anything for you.
What would move the grade up: making Stealth the default rather than a $60 upsell, narrowing the perpetual prompt licence to what running the service requires, publishing a bias evaluation of the kind researchers have had to do from the outside, and offering commercial customers indemnification. A settlement or a defence verdict in the studio cases would also matter, because a chunk of this rating is the risk that a court reshapes the product under you. What would move it down: an injunction, or a finding of willfulness. As of September 2026 both remain open.