Microsoft will defend you in court and pay the judgment. Getty Images starts its legal protection at fifty thousand dollars per generated image. Both promises are real, both are in writing, and both answer the same question: what happens if somebody else claims the image you generated was already theirs.
That is a reasonable thing to worry about. It is also, for most small businesses, not the loss that is actually coming. The likelier one is quieter, nobody sells protection against it, and no vendor will bring it up, because it is not a risk that might materialize. It is the default position the second you hit generate. The marketing asset you just made is probably not yours, and there is nothing stopping a competitor from using it too.
That gap between the protection on sale and the exposure in front of you is what this piece is about. Three fixes close most of it, and all three are things an owner can do without a lawyer.
Copyright asks how much of it you made, not what you paid or which tool you used
The rule underneath all of this is short. United States copyright requires a human author, and material generated by a machine without meaningful human contribution is not protectable. The Copyright Office set this out at length in Copyright and Artificial Intelligence, Part 2: Copyrightability, published January 29, 2025, which concluded that prompts alone do not give a user enough control over the output to make that user its author.
That position is no longer just the Office’s view of its own practice. The D.C. Circuit affirmed it in Thaler v. Perlmutter on March 18, 2025, holding that the Copyright Act of 1976 requires a human author because the statute’s provisions on ownership, inheritance, duration and signatures all presuppose one, as Skadden’s analysis of the ruling lays out. The Supreme Court denied review on March 2, 2026, which leaves that ruling as binding precedent. Whatever else is unsettled in AI law right now, this part is not.
Here is the part worth sitting with, because it is the opposite of how the question usually gets framed. The law does not grade your asset. It does not care that the image is beautiful, that it took you three hours of prompting, that you paid for the premium tier, or that the tool was trained on properly licensed material. It asks a single question that has nothing to do with any of that: how much of this did a person make? Human-authored material you contribute stays protectable, and so can your creative selection, coordination and arrangement of AI outputs. The generated pixels themselves do not.
Which means your ownership is decided by your editing process, not by your purchase. That is an uncomfortable finding for anyone who assumed the paid tier settled it, and it is good news for anyone willing to work, because the fix is free.
Fix one: sort the asset before you generate it, not after
Most owners treat every AI output the same way, which is why the ownership problem stays invisible until it matters. The sort takes about ten seconds and it has two columns.
Some assets need to be ownable. Your logo, a mascot, an illustration style you intend to build a brand around, the hero image of a campaign you will run for two years, a character your customers will come to recognize. These are assets whose whole value is that they are yours and nobody else’s.
Everything else only needs to be safe. The background texture behind a Tuesday social post. A generic photo at the top of a blog article. A placeholder illustration in a slide deck. Nobody is going to copy these, and if they did you would not notice or care.
The test that separates them is one question, and you already know the answer for every asset you own: which of these would make you genuinely angry to see a competitor using? That one needs a human author. The rest need an indemnity and nothing more.
Doing this sort before you generate rather than after is the whole point, because the human contribution that creates ownership has to happen during the work. You cannot add it retroactively to a file you accepted on the first try six months ago.
Fix two: keep the record on the ownable ones, then register them properly
For anything in the first column, two things have to be true. A person has to have made a real creative contribution, and you have to be able to describe it.
In practice that means working the way a designer works rather than the way a slot machine works. Generate, then redraw, recompose, combine outputs, add your own elements, make deliberate choices about arrangement. Keep the intermediate files. The point is not to satisfy a form later, it is that the contribution is the thing that creates the asset, and the record is just evidence it happened.
Then register the ones that matter, and register them honestly. Applicants have a duty to disclose AI-generated content in a work submitted for registration, per the Copyright Office guidance published at 88 FR 16190 on March 16, 2023. Two fields carry the work. In the “Author Created” field you describe what the human made, for example the selection, coordination and arrangement of the material. In the “Limitation of Claim” section you exclude the AI-generated portion under “Material Excluded.” The same guidance is blunt about what happens if you skip it: the Office may cancel a registration where essential information was omitted, and a court may disregard the registration in an infringement action if you knowingly gave inaccurate information. A registration you obtained by staying quiet is worth less than no registration at all, because you will find out it is worthless on the day you try to use it.
The cost is not the obstacle here. Electronic filing runs $45 for a Single Application, meaning one work, one author, same claimant, not made for hire, and $65 for a Standard Application. For a logo you plan to use for a decade that is not a real number. The obstacle is that almost nobody does it, and then discovers the gap at the worst possible moment.
Fix three: buy the indemnity for the second column, and actually read its conditions
For everything that only needs to be safe, the protection you want is contractual, and it is genuinely worth having. Two options with published terms, and their prices as listed this week.
Microsoft’s Copilot Copyright Commitment, announced by Brad Smith and Hossein Nowbar on September 7, 2023, says that if a third party sues a commercial customer for copyright infringement over the output, Microsoft will defend the customer and pay the amount of any adverse judgments or settlements. It attaches to paid commercial Copilot products. Microsoft 365 Copilot Business currently lists at $18.00 per user per month paid yearly, a promotional rate against a $21.00 list price and available between July 1, 2026 and September 30, 2026, or $25.20 per user per month billed monthly.
Getty Images takes the other approach and trains on material it already controls. Its generative tool was trained exclusively on licensed creative content with web-scraped material excluded, and the same page states that your AI agreement includes Getty Images’ legal protections and indemnification “starting at $50,000 USD per generated image.” Introductory packages run $49 for 25 generations and $149 for 100, with each generation producing four images and unused generations expiring one year from purchase.
Now run those against the obligation, because this is where owners get caught twice.
The first catch is in the conditions. Microsoft’s commitment requires the customer to use the content filters and safety systems built into the product, to not attempt to generate infringing material, and, the one that actually bites, to not provide input to the service that the customer does not have appropriate rights to use. Uploading a competitor’s product photo or a stock image you never licensed and asking the tool to restyle it is exactly the move that voids the protection you are paying for, and it is a completely ordinary thing for a busy owner to do without thinking.
The second catch is the bigger one. Neither of these protections gives you ownership, and they were never meant to. Getty’s fifty thousand dollars defends you against somebody else’s claim. It does not let you stop a competitor from running your ad. Microsoft’s commitment pays your judgment. It says nothing whatsoever about whether you can enforce anything against anyone. The two exposures point in opposite directions, and the entire market sells protection for only one of them. If you have been feeling covered because you upgraded to a paid tier, this is the sentence to reread.
The exposure none of the three fixes touches
Copyright asks who made the thing. Trademark asks a different question: does it confuse customers about who they are buying from? An AI-generated logo, product name or packaging design can be perfectly clear of every copyright problem discussed above and still collide with somebody’s registered mark, and not one indemnity in this article addresses that.
The check is free and takes minutes. The USPTO retired its old TESS system in November 2023, so searches now run through the USPTO’s Trademark Search at tmsearch.uspto.gov, which covers the full federal register at no cost and offers a basic mode alongside an advanced query builder. Run any name or mark you are about to commit to through it before it goes on a sign, a truck or a package. This is the highest-value few minutes in the entire article, and it is the step people skip because it does not feel like it belongs with the AI question. It does. Generating a logo now takes seconds, which means you are producing candidate marks far faster than you have ever cleared them.
Wider AI obligations are moving separately from all of this, and mostly in a lighter direction than 2026 was expected to bring; we covered what small businesses still owe as the big AI compliance deadlines soften earlier this year. Copyright is not part of that softening. It was settled by a court instead.
What we are not claiming here
Three limits, stated plainly, because the useful version of this argument is the bounded one.
This is about United States copyright only. Other countries answer the authorship question differently, and if you sell into them the analysis changes. It is also about ownership and infringement, not about whether you are allowed to use AI at all. On that separate question the platforms have been clear in writing, and we read the four operative texts in full: none of them penalizes you for using AI to make content, despite the noise about AI slop. And nothing here is a substitute for a lawyer on a specific asset. What it should do is tell you which assets are worth paying one to look at, which is a question you can answer yourself in ten seconds using the sort in fix one.
The through-line in all of it is worth naming, because it cuts against the usual pitch. Every one of these rules rewards the presence of a person. The Copyright Office protects your contribution and not the machine’s. The registration form asks you to describe what you made. Trademark turns on how a human customer perceives a mark. The technology can do the drafting all day; what converts a file into an asset you own is that somebody was there, making choices they could describe afterward. That is not a sentimental point about creativity. It is the actual legal test, and it happens to be the one part of this you fully control.
Frequently Asked Questions
Can I copyright an image or ad that AI generated for my business?
Not the AI-generated part on its own. United States copyright requires a human author, and the Copyright Office concluded in its January 2025 report on copyrightability that prompts alone do not give a user enough control to make them the author of the output. What you can protect is your own contribution: material you created and added, and your creative selection, coordination and arrangement of the outputs. The more of the finished asset a person actually made, the more of it you own.
Do I have to tell the Copyright Office that I used AI when I register?
Yes. Applicants have a duty to disclose AI-generated content in a work submitted for registration. You describe the human contribution in the “Author Created” field and exclude the AI-generated material under “Material Excluded” in the Limitation of Claim section. The Office may cancel a registration where essential information was omitted, and a court may disregard a registration in an infringement action if the applicant knowingly provided inaccurate information, so a quiet registration is worse than none.
Does my paid AI subscription mean I own what it generates?
No, and this is the most common mistake. A paid tier can buy you a contractual indemnity, meaning the vendor defends and pays if a third party claims your output infringes their rights. That is protection against a claim, not ownership. It does not give you the right to stop a competitor from using the same or similar output. Ownership depends on human authorship, which no subscription can supply for you.
What is the fastest way to reduce my risk this week?
Three steps. Sort your assets into the ones that must be ownable, such as a logo or a signature campaign, and the ones that only need to be safe. For the ownable ones, make and document a real human contribution and consider registering them, which costs $45 for a Single Application or $65 for a Standard Application electronically. For everything else, work inside a tool with a written indemnity and follow its conditions, particularly the rule against feeding it material you do not have rights to. Then run any new name or mark through the USPTO’s free Trademark Search before you commit to it.
None of this needs a new budget line or a policy document. What it needs is the decision moved in front of the work instead of behind it. Notice that fixes two and three are things you do to an asset, and can be done late, whereas fix one is a call you make about a job you have not started, and it stops being available the moment the file exists. So take the next thing your business is going to lean on, the logo, the campaign, the character customers will recognize, and put it in a column before you open the generator. If it lands in the ownable column, you already have your instructions: make it the slow way, keep what you made, register it. If it lands in the other one, generate it in nine seconds and let the contract carry it.
