Is Midjourney Safe for Work? What Its Terms Actually Allow
Key takeaways
- Midjourney is not safe for confidential work on default settings: its terms let it train on submitted prompts and images.
- Generated images are public and remixable by default. Stealth Mode, which needs a paid plan, hides outputs from the gallery but does not end the training licence.
- Our directory records Midjourney as training on business data with no training opt-out, the least favourable of the four image tools we track.
- Approve it narrowly for non-confidential concept imagery, write the prohibited data types into your register, and re-check every six months.
Midjourney is not safe for confidential work material on its default settings. Its terms let the company use the prompts you type and the images it generates to train its models. Generated images are public and remixable unless you pay for Stealth Mode, and even Stealth Mode only keeps outputs out of the public gallery. It does not stop training. For a small team that makes Midjourney a reasonable tool for generic concept art, mood boards and non-confidential illustration, and a poor one for client material, unreleased product designs, staff details or anything a customer sent you in confidence. This guide sets out what the terms say, how to run the same checks on any image generator, and how to write the decision into a register so the answer holds up the next time someone asks.
What the terms say by default
Midjourney's Terms of Service state that content you create is publicly viewable and remixable by default. The same terms grant the company a licence to reproduce, adapt and distribute the prompts and images you submit. Third-party monitoring services summarise that licence as permission to train models on your content, worldwide, without payment, and one such summary describes it as not revocable. We have not verified that wording against the official text, which is why the official document is the one to read before you rely on it. Our own Midjourney entry in the tool directory records the company as training on personal data and on business data, with no training opt-out. Of the four image tools we track, that is the least favourable combination.
Stealth Mode is a visibility control, not a data control
Midjourney offers Stealth Mode on its paid Pro and Mega plans, and it keeps your generations out of the public gallery. Read closely, that is all it does. Secondary sources agree that Midjourney keeps the right to use your inputs to improve its service after a generation is hidden, so the protection is about who can browse your outputs, not about whether the company can learn from them. A team tempted to rely on Stealth Mode for client work should treat it as a publication setting and document it as such. It should not appear in the register as the reason the work is safe.
Retention and deletion are not spelled out
Midjourney's privacy policy says it keeps personal data only as long as needed for the purposes it lists, and it does not publish a fixed retention period. The directory reaches the same conclusion: no defined retention or deletion timeline was found for prompts or generated content. Requests to delete personal data go through the privacy contact named in the policy. For a compliance file, the accurate entry is that retention is unspecified. Writing that it is short would be an assumption, and an auditor will ask where the figure came from.
Five checks for any image generator
We ask the same five questions of every AI tool in the directory, and they carry over to image generators without change. First, does the tool train on what you submit by default, on personal accounts and on business accounts? Second, is there a documented opt-out, and does it cover both? Third, is there a data processing agreement you can sign, and does it name the subprocessors? Fourth, who owns the output, and what commercial use does your plan allow? Fifth, how long is data kept, and how do you ask for it to be deleted? A tool that fails the first question on a business plan is not an approved tool for confidential work, however good its output looks. The same reasoning applies to text assistants, which we cover in is ChatGPT safe for work.
Worked example: four image tools in our directory
The table below compares the four image generators in our directory on the first three checks. Every figure comes from the directory's own records, not from vendor marketing, so it shows what we have verified rather than what each company says about itself.
What the directory shows across all 60 tools
Across the full directory of 60 tools, 40 list a data processing agreement, 31 offer a training opt-out, and two are recorded as training on business data by default. Midjourney is one of those two; the other is a video tool. Adobe Firefly is the only image tool recorded as not training on either personal or business data, with a DPA listed. The directory snapshot was generated on 7 July 2026, so check the vendor pages again before any purchase decision. The figures are a starting point for your own review, not a substitute for it.
What this means for a team
Midjourney is a workable tool for a narrow set of jobs: generic concept imagery, internal mood boards, and illustration that contains no client names, no real people, no unreleased product and no reference images you do not own. Write that down. A rule is easier to follow when it names the data types rather than the tool. Something along these lines works: staff may generate non-confidential concept art in Midjourney, may not type client, patient, staff or unreleased product details into a prompt, and may not upload a customer's logo or photograph as a reference image.
Commercial use is a separate question
Whether you can use a Midjourney image in marketing depends on the plan you pay for and on the licence terms in force when the image was made. A 2026 commercial-use guide describes plan-based ownership and revenue-related conditions, but that is a secondary source and the current terms should be checked directly. Commercial rights are not the same as data protection. An image can be licensed for your website and still have been used to train a model. Record both answers separately in your register, and have someone check the current plan terms before a campaign goes live rather than after.
Recording the decision in the register
A register entry needs five fields: the tool and plan, the approved data types, the prohibited data types, a named reviewer, and the date of the next review. For Midjourney, a defensible entry reads: approved for non-confidential concept imagery on a named plan; prohibited for client, patient, staff and unreleased product information, and for customer-supplied reference images; reviewed by the operations lead; re-checked every six months against the vendor's current terms. The generative AI policy guide explains how to fit image generators into the wider policy. If staff have already signed up on personal accounts, the shadow AI guide covers how to bring that use into view without a blame exercise.
Where to start this week
Find out whether anyone on your team already uses Midjourney, on which account and for what. If the answer is a personal account used for client work, move that work onto an approved route first, then write the register entry. The tool directory lists the training, retention and contract facts for each tool we track, and how to vet an AI tool before rollout sets out the wider process. If you need the policy itself, the AI usage policy generator gives you a draft to adapt to your own tools and data.
| Tool | Trains on personal data | Trains on business data | Training opt-out | DPA listed |
|---|---|---|---|---|
| Midjourney | Yes | Yes | No | No |
| Adobe Firefly | No | No | Not applicable | Yes |
| Ideogram | Opt-out available | No | Yes | No |
| Leonardo AI | Opt-out available | No | Yes | Yes |
Related guides
- What Is an AI Policy? A Plain-English Guide
- ISO/IEC 23894: AI Risk Guidance vs ISO 42001
- How to Mitigate AI Risk: A Step-by-Step Framework
Check a specific tool in the AI Tool Risk Directory.