AI and copyright at work: who owns the output and what you can safely use

Key takeaways

In the US, material generated entirely by AI generally cannot be copyrighted. Your own creative contributions, edits, and arrangements can. That cuts both ways: you may not own it, and neither does anyone else.
Training-data lawsuits are partly decided and partly ongoing, and the exposure from how a model was trained sits mostly with the vendor. Your exposure comes from outputs that reproduce protected work.
Commercial use is mostly about your vendor's terms, your plan tier, and your prompts. Never name living artists, brands, characters, or specific works, and never feed in material you do not have rights to.
Indemnification from major vendors is real but conditional: usually paid or enterprise plans only, safety filters on, no intentional infringement, and vendor control of the defense.
Trademarks, logos, and real people's faces and voices are separate from copyright and are the most common way generated content gets a company into trouble.

The short version

First, in the United States, material generated entirely by AI generally cannot be copyrighted, though your own creative contributions can. Second, the lawsuits over whether training models on copyrighted material is lawful are partly decided, partly ongoing, and mostly a vendor's problem rather than yours. Third, the ways a business actually gets into trouble with generated content are usually not copyright at all: they are trademarks, real people's faces and voices, and outputs that reproduce someone else's work.

None of this is legal advice. It is a working overview so you know which questions to ask and when to ask a lawyer, because the law here is changing quickly and differs by country.

Who owns AI output

The US position: human authorship required

The US Copyright Office has said, in guidance issued in 2023 and a longer report in early 2025, that copyright protects human authorship. Text or images produced from a prompt alone are not eligible, no matter how detailed the prompt, because the person did not control the specific expression that came out. Federal courts have agreed in the cases decided so far.

What can be protected: the human parts. If you substantially edit generated text, select and arrange generated pieces into a larger work, or combine a generated element with your own creative material, your contribution can be protected, and the Copyright Office asks you to identify and disclaim the AI-generated portion when you register. Treat any claim of ownership over lightly edited output as uncertain.

The consequence cuts both ways: you may not be able to stop a competitor from copying a generated image on your website, but nobody else owns that output either, including the vendor.

You, your employer, and the vendor

Three documents decide who holds whatever rights exist in output produced at work: the vendor's terms, which usually assign output to the account holder; your employment agreement, which usually assigns work product to your employer, as a work made for hire or by assignment; and your company's AI policy, which may add its own rules.

Read all three before you assume anything, especially on a personal account used for company work, where the documents may point in different directions.

Other countries differ

The United Kingdom and a few other jurisdictions have provisions for computer-generated works, though how they apply to modern generative tools is unsettled. If you publish or license internationally, ask counsel where you need protection.

Training-data disputes, at a high level

Model vendors trained their systems on enormous amounts of copyrighted text and images. Authors, news publishers, visual artists, music companies, and stock-image companies have sued, arguing that copying that material to train a model infringes their rights. Vendors argue that training is a transformative use, which in the US is a fair-use question.

As of mid-2026 the picture is mixed. Some US trial courts in 2025 found that training on lawfully acquired books was fair use in the circumstances before them, while treating pirated copies as a separate problem, and at least one large settlement followed. Other cases, including ones about outputs that reproduce protected work and ones about news content, are ongoing, and appeals will take years. Outside the US, some countries have text-and-data-mining rules that permit training under conditions.

What this means for you: the exposure from how a model was trained sits mostly with the vendor. Your exposure comes from outputs. A model can occasionally reproduce a passage of a book, a distinctive character, or a near-copy of a photograph it saw during training. Publishing that is your problem, not the vendor's, unless an indemnity covers you. The prompts most likely to produce it are the ones that ask for a specific work, a named living artist's style, a brand, or a character.

Using generated text commercially

For text, the terms of the major assistants generally allow commercial use of output on paid plans and usually on free plans too, with restrictions such as not using output to build a competing model. Check the terms for the plan you actually use, and recheck periodically.

The risks with text are mostly practical: a generated passage that closely matches an existing article is a plagiarism problem for your reputation before it is a copyright problem, and a generated statistic that does not exist is a credibility problem.

The habits that keep you safe: rewrite in your own voice, verify every fact and quotation against a source, and paste a few distinctive sentences into a search engine to make sure they are not lifted from somewhere. Do not ask for text in the style of a specific living author for commercial use, and do not paste in material you lack rights to and ask for a close rework.

Using generated images commercially

Images carry more risk than text, because visual similarity is easier to spot and image models are better at reproducing specific styles and subjects. Three decisions matter.

Which tool. Some image generators are trained only on licensed, stock, or public-domain material and are marketed to businesses on that basis. Adobe Firefly is the best-known example, and several stock-photo companies offer generators trained on their own libraries. These tools give up some range for lower risk, which is usually the right trade for anything that goes on a product, an ad, or a storefront. Check each tool's terms: some grant commercial rights only on paid plans, and some require a higher plan tier for companies above a certain size.

What you prompt. Do not name living artists, brands, logos, characters, celebrities, or specific copyrighted works. Do not upload a photograph you do not own and ask for a variation of it. Describe the subject, mood, composition, and medium instead. The closer your prompt is to "make me a copy of that," the closer the output will be to something you cannot use.

What you do with it. Product photos should be photographs of the product; a generated image of something you sell invites consumer-protection questions. Use generated images for illustration, backgrounds, concepts, and mood, label them where a platform or client requires it, and keep the prompt, tool, plan, and date in a file next to the image.

Indemnification, in words

Several major vendors offer to defend paying customers against third-party copyright claims arising from output, and to pay resulting damages, subject to conditions. Microsoft announced such a commitment for its Copilot products, Google for its Workspace and Cloud generative AI services, OpenAI for its enterprise and API customers, Adobe for Firefly under enterprise agreements, and Anthropic in its commercial terms. The details differ, and the details are the point.

Typical conditions: an eligible paid or enterprise plan, content filters and safety features left on, no knowledge that the output was infringing, no prompting for a specific protected work, use of the product as intended, and prompt notice with the vendor controlling the defense. Many exclude output you substantially modified, trademark and right-of-publicity claims, and any use that violates the terms.

What an indemnity is not: it does not make the output copyrightable, it does not protect your reputation, and it does not exist on most consumer plans. If it matters to your business, and for a marketing or publishing team it should, be on a business plan, read the commercial terms, and have counsel confirm what your plan actually covers.

Trademarks, logos, and real people

Trademarks. Trademark law protects against consumer confusion about the source of goods and services, regardless of who or what drew the design. A generated logo that resembles an existing mark is a problem whether or not the resemblance was intended. Before adopting a generated logo or product name, run the same clearance search you would for a human-designed one. Trademark protection comes from distinctiveness and use in commerce rather than from authorship, so a generated logo can generally function as a mark, but building a brand on one deserves a conversation with counsel.

Likeness and voice. Generating a real person's face or voice for commercial use without written permission is a right-of-publicity problem under state law in much of the US, and a number of states have updated their laws specifically for AI-generated voices and likenesses. This includes employees: a training video that uses a cloned colleague's voice needs a signed release that covers AI reproduction specifically. Deepfake laws add criminal exposure for certain intimate or deceptive uses.

Characters and franchises. A mascot that looks like a well-known animated character, a product shot on a set that resembles a film, a parody of a famous ad: each is a copyright and trademark question at once. Treat any recognizable reference as off limits unless counsel has cleared it.

Worked example: a product launch campaign

Say a marketing team of four at a mid-sized outdoor-gear company is producing a launch campaign for a new backpack: web copy, a hero image, a set of social posts, and a short video with voiceover.

The copy is drafted with a general assistant on the company's business plan, then rewritten by the team's writer, with every product claim checked against the engineering spec. The hero image comes from a licensed-data image tool on a paid plan, prompted with a ridge at dawn, mist, and warm light, no brand or artist names, and the backpack is composited in from a real product photograph. One social concept is rejected because the scene resembles one from a well-known film, another because a background hiker's jacket carries a mark that looks like a competitor's logo. The voiceover uses a licensed synthetic voice under commercial terms rather than a clone of the founder's voice, which would have needed a release and legal review.

The team saves the prompts, tool names, plan tiers, and dates in the campaign folder, notes which elements were generated, and confirms that the company's plan includes indemnification for generated output. The whole rights review adds an afternoon to a three-week project.

A practical checklist

Run this before anything generated goes public.

Am I on a plan whose terms allow commercial use of output, and do I know what its indemnification does and does not cover?
Did the prompt name any artist, brand, character, celebrity, or specific work, or include material I do not have rights to? If yes, regenerate.
Does the output resemble something recognizable: a logo, a character, a famous photograph, a passage from a book? If unsure, search for it.
Does it show a real person's face or voice? If yes, do I have a written release that covers AI reproduction?
Have I edited the text substantially and verified every fact and quotation?
Have I recorded the tool, plan, prompt, and date next to the asset, and added an AI-content label where the platform, client, or jurisdiction requires one?
Would I be comfortable explaining how this was made in a deposition? If not, do not publish it.

A prompt for a first-pass rights review

The assistant cannot clear rights, but it can make sure you asked the right questions before you spend money on counsel.

Act as a careful marketing operations lead who prepares rights questions for the company's lawyer. You are not a lawyer and will say so. Review the following planned use of AI-generated content and flag issues under four headings: copyright and ownership, trademark, likeness and voice, and vendor terms.

The asset: [DESCRIBE WHAT WAS GENERATED, FOR EXAMPLE A HERO IMAGE, 800 WORDS OF WEB COPY, OR A 30-SECOND VOICEOVER]
The tool and plan: [TOOL NAME AND PLAN TIER, FOR EXAMPLE A BUSINESS PLAN WITH INDEMNIFICATION]
The prompt used: [PASTE THE PROMPT]
Any inputs I uploaded: [DESCRIBE, OR SAY NONE]
Where it will be used: [WEBSITE, PAID ADS, PRODUCT PACKAGING, OR CLIENT DELIVERABLE, AND IN WHICH COUNTRIES]
Anything recognizable in the output: [DESCRIBE, OR SAY NOTHING I NOTICED]

For each heading, tell me whether you see a likely issue, a possible issue, or no obvious issue, with one sentence of reasoning. Then list the questions and documents I should bring to counsel. Keep the whole answer under 350 words. Ask me up to four clarifying questions first if the description is too thin.

When to call counsel

Call before, not after, when: the asset is a logo, product name, or packaging; a real person's likeness or voice is involved; the content will be licensed to clients or used in paid advertising at scale; you publish in multiple countries; or a generated output resembles something you recognize. A one-hour review is cheaper than a takedown, and far cheaper than a lawsuit.

Next steps

Set the rules for your team with How to write a team AI policy and the AI policy generator, and pair this guide with AI privacy at work.
Learn the terms: training data, text to image, voice cloning, deepfake, and AI watermarking.
Tool notes: Adobe Firefly, Midjourney, Canva, and ElevenLabs.
Role playbooks: AI for marketers, AI for graphic designers, and AI for content writers.

Frequently asked questions

Can I copyright something an AI wrote for me?

In the United States, generally not the AI-generated part itself. The Copyright Office's position is that copyright requires human authorship, and a prompt alone does not make you the author of what comes out. What you can protect is your own contribution: substantial edits, the selection and arrangement of generated pieces, and the original material you combine them with. For anything valuable, talk to counsel.

Can I use AI-generated images in my marketing?

Usually yes, with care. Use a tool and plan whose terms allow commercial use, prefer generators trained on licensed or stock material for anything that goes on a product or an ad, never prompt with artist names, brands, characters, or celebrities, and check the output for anything recognizable. Keep the prompt, tool, plan, and date on file, and label the image where a platform or client requires it.

Will my company get sued for using ChatGPT output?

The lawsuits so far have mostly targeted the model vendors over training, not ordinary business users over outputs. Your realistic risk is publishing an output that reproduces someone else's protected text or image, or that infringes a trademark or a person's likeness. Rewriting in your own voice, verifying facts, avoiding prompts that name specific works or people, and being on a plan with indemnification keep that risk small.

What does AI indemnification actually cover?

Several major vendors offer to defend eligible paying customers against third-party copyright claims arising from output and to pay resulting damages, subject to conditions: an eligible plan, content filters left on, no intentional infringement, use of the product as intended, and prompt notice with the vendor controlling the defense. It typically excludes trademark and likeness claims, heavily modified outputs, and consumer plans. Read your plan's commercial terms.

Who owns the output if I use AI at work, me or my employer?

Three documents decide it: the vendor's terms, which usually assign output rights to the account holder; your employment agreement, which usually assigns work product to your employer; and your company's AI policy. On a company account for company work, the answer is almost always your employer, to the extent anyone can own it. On a personal account used for company work, the documents may conflict, which is one more reason to use the company tool.

AI can profile a reporter from their last twenty stories, draft the pitch, turn a call into a briefing doc, and pull a coverage report in minutes. It cannot know who is real, what your client actually said, or what the FTC now counts as a fake testimonial. Those stay your job.

Job playbook

AI for Lawyers

AI drafts, summarizes, and finds a starting point faster than any associate, and it will hand you a fabricated case with a straight face. Here is how to get the speed without the sanctions or the privilege problem.

Terms in this guide

Training data Generative AI Text-to-image Diffusion model Deepfake AI watermarking Voice cloning