Every post on this site needs an image, and this site carries advertising. Those two facts together are the whole reason this page exists, because they mean the licence on that image is my problem and not the tool’s. Nobody sues a prompt.
So this is not a gallery shoot-out and there is no winner at the bottom. I don’t lean on an image generator in my own work and I have never compared them side by side, and the previous version of this page pretending otherwise — it ended with a box calling Midjourney the most stunning, professional-grade AI art of any tool the site claimed to have tried — is one of the reasons the site had to be rebuilt. What I can give you is the part that actually decides things for a small publisher: what the terms say, how hard they are to read, and where I still reach for a camera instead.
Six of the pages I needed would not open for me
On 20 September 2026 I tried to read the pricing and licensing pages of the tools this article used to recommend. Midjourney’s plans page and its documentation, Canva’s pricing page and its content licence agreement, Ideogram’s pricing page and Adobe’s Firefly FAQ all returned HTTP 403 to my request. They open perfectly well in an ordinary browser; large sites block automated requests as a matter of routine, and that is their business.
I am reporting it anyway, because it changes what you should accept from an article like this one. Almost nobody re-reads six terms-of-service documents before republishing a comparison, and terms for generative image tools have changed repeatedly. When you see a confident sentence about what a tool “allows commercially”, the useful question is not whether the writer is honest. It is on what date, and from which page, they last read it.
My practical rule, which costs about four minutes: before you use a generated image on anything that earns money, open that vendor’s own licence page yourself, in a browser, on the day, and look specifically for three things — whether commercial use is permitted on your tier, whether it is permitted on the free tier, and who is stated to own the output. Not a blog summary. The page.
What the US Copyright Office has actually published
If your audience is in the United States, this is the material that matters more than any tool’s marketing page, and it is free to read. The Copyright Office’s AI initiative page, checked on 20 September 2026, lists a policy statement, “Copyright Registration Guidance for Works Containing AI-Generated Materials”, published in the Federal Register on 16 March 2023, and three parts of a report: Part 1, Digital Replicas, dated 31 July 2024; Part 2, Copyrightability, dated 29 January 2025; and Part 3, Generative AI Training, in a pre-publication version dated 9 May 2025.
I am giving you the titles and the dates rather than quotations because I read the index, not the documents, and I would rather say so than paraphrase a legal text I have skimmed. The reason to point at them at all is that they are the primary source on the question people ask most — whether an image a machine produced can be registered — and they are considerably more specific than the summaries of them.
What I am not going to do is tell you what this means for your logo. I am not a lawyer, the previous version of this page gave out trademark advice it had no business giving, and that section has been deleted rather than corrected.
The three image jobs a site like this actually has
Strip out the artistry and a publisher’s image needs come down to three, and they have completely different risk profiles.
- A featured image, to sit at the top of a post and in a card on a listing page. Decorative. Nobody reads it as evidence. This is the job generated images are genuinely good at and the one with the least to go wrong, as long as the licence is right and the image does not silently drift into looking like a photograph of a real place or person.
- A picture of an interface. Dashboards, admin screens, panels, results. This is where I draw a hard line, and I will come back to it.
- A mark or a logo. Rarely needed, high consequence, and the one case where the output is not the end of the job — you need vector files, a single-colour version, a version that survives being 16 pixels wide, and a name nobody else is using.
Treating those three as one category, which is what “best AI image generator” articles do, is why they are useless. The right tool for a decorative header is nearly irrelevant. The right approach for a logo is mostly not a generator at all.
The rule I will not break: no mockup dressed as a screenshot
If a post needs a picture of a control panel, a search report or a billing screen, it gets a real screenshot that I took, with anything private painted out — or it gets no image at all. Not a generated approximation of a dashboard. Not an illustration of a graph going up. Not a “representative example”.
This is not squeamishness. A generated interface is a claim about what a product looks like and what it showed, and it is a claim you did not verify. It is the visual version of writing “in our testing” about a test that never happened, which is precisely what this site used to do in text and precisely what an ad network’s reviewer is looking for. The full account is in how that went on this site; the short version is that inventing evidence is the one category of error nothing else on a page can compensate for.
The cost of this rule is that some posts have fewer images than they “should”. I have made peace with that.
Where I still use a photograph
Anywhere the image is making a claim about the real world. A place, a person, a physical product, a thing that exists. A generated image of a building in a town is a picture of a building that is not in that town, and readers who know the town notice immediately.
Two working rules I apply across every site I run, and they have nothing to do with which generator is best:
- Every page gets its own image; none is ever reused across pages. The same header photo on six pages tells a reader, and a reviewer, that the six pages came out of one batch.
- I look at the image before I assign it. Obvious, routinely skipped, and the source of most of the genuinely embarrassing mistakes — the wrong subject, six fingers, unreadable invented text along the bottom edge.
Alt text belongs in the same discipline. Google’s own guidance on using generative AI content asks publishers to review the accuracy of automatically generated metadata, image alt text included. An auto-written alt attribute describing something that is not in the picture is a small lie that costs nothing to fix and that nobody checks.
Logos, and what I have removed from this page
An earlier article on this site told readers that a design agency logo runs $500 to $5,000, that AI logo makers charge $20 to $65, that one tool’s brand kit costs “around 96 USD per year”, that vectorising costs “about 10 USD monthly”, and that a freelancer will finish the job for $50 to $150. It also stated that you generally cannot trademark a logo built from a stock element library.
I have bought none of those products, commissioned none of those freelancers, and I am not qualified to tell anyone what is or is not registrable as a trademark. Every one of those numbers is gone, and I have not replaced them with fresher ones I would be just as unable to defend.
What survives is craft advice that does not depend on a test: a logo you can actually use is a package, not a picture. A vector master, transparent exports at several sizes, a one-colour version, a version that reads on a dark background, and a favicon that is still legible at the size of a browser tab. If a tool hands you a single PNG, you have bought an image and you will be buying the logo again the first time you need a banner printed. And before any of that, the name — checking your national trademark database is a free search and the one step where professional advice is genuinely worth paying for.
Before an image goes on a page that carries ads
- Open the vendor’s licence page in a browser today and confirm commercial use is allowed on the tier you are actually on. Free tiers frequently differ from paid ones.
- Confirm the image is not a photograph of somewhere or someone real, and is not being presented as one.
- If it depicts an interface, replace it with your own screenshot or delete it.
- Write the alt text yourself and check it describes what is in the frame.
- Check you have not used this image on another page of the same site.
- Note the tool and the date somewhere you can find again, so that when the terms change you know what you published under the old ones.
What people ask me about licences
Do I have to say an image was AI-generated? No rule I can point you to obliges a blog to caption every decorative header. But the moment an image could be mistaken for evidence — a screenshot, a photograph of a place, a picture of a person — the caption stops being a courtesy. My own line is that anything a reader might treat as a record of something that happened gets labelled or gets cut.
Is the free tier good enough for blog headers? For the quality, almost certainly. For the licence, that is exactly the question you have to answer on the vendor’s page rather than here, because free and paid tiers are where commercial-use terms most often diverge, and because I could not read several of those pages on the day I wrote this.
Which one would I use if I had to pick today? I would pick whichever one lets me read its licence in under five minutes and says plainly that I may use the output commercially on the tier I am paying for. That sounds like a dodge and it is genuinely how I would decide. For a decorative header, how legible the terms are matters more to me than which tool draws the nicer gradient.
If you are assembling the rest of a small publishing setup, the inventory of what I actually run on these two blogs covers the pieces around this one, and the post on free tiers goes into which “free” plans are really free. For the writing side of the same workflow, my notes on AI writing tools are the companion piece to this one.