AI and copyright: who owns AI-generated content?

AI and Copyright: Who Owns AI-Generated Content?

Most professionals working with AI tools carry at least one dangerous assumption about copyright — and that assumption is quietly creating legal and ethical risk in their work. You might believe that anything you generate with ChatGPT belongs to you the moment you hit Enter. Or that AI tools can freely train on any published content because they're just 'learning,' the way humans do. Or perhaps you assume that slapping your name on an AI-generated image or article gives you the same legal protection as something you wrote by hand. All three beliefs are wrong — not slightly off, but structurally incorrect in ways that matter for your career, your company, and the creative professionals whose work feeds these systems.

Three Beliefs That Sound Right — But Aren't

Before examining each misconception individually, it helps to understand why these beliefs feel so convincing. Copyright law developed over centuries to protect human creative labor. AI-generated content didn't exist when those frameworks were built. So we instinctively apply old mental models to a genuinely new situation — and the fit is poor. The US Copyright Office, the EU's AI Act, and courts in multiple countries are now actively wrestling with exactly these questions, and their early rulings are clarifying the landscape in ways that surprise most professionals. The three myths below represent the most common — and most consequential — misunderstandings circulating in offices and Slack channels right now.

Myth 1: "The Output Is Mine Because I Prompted It"

This is the most widespread belief among professionals using ChatGPT, Claude, Midjourney, and similar tools. The logic feels airtight: you wrote the prompt, you directed the creative process, you're the one who refined and selected the output. Surely that makes you the author? In everyday terms, yes — you did the intellectual work of steering the system. But copyright law doesn't care about intellectual steering. It cares about one specific thing: human authorship expressed in a fixed, original form. The US Copyright Office has stated explicitly, in guidance issued in February 2023, that it will not register works produced entirely by machines without creative input from a human author.

The landmark case here is Thaler v. Perlmutter, decided in August 2023 by a US federal district court. Dr. Stephen Thaler attempted to register a piece of AI-generated visual art called 'A Recent Entrance to Paradise,' created by his AI system DABUS. The court ruled against him, affirming that copyright requires human authorship. The judge wrote that human authorship is 'a bedrock requirement of copyright.' This wasn't a close call or a narrow ruling — it was a categorical statement. Thaler appealed, and the case continues, but the lower court's reasoning reflects the Copyright Office's consistent position across multiple similar requests it has rejected since 2022.

Where things get genuinely complicated is the spectrum between pure AI output and human-AI collaboration. If you use Midjourney to generate 100 images and then manually select, crop, arrange, and annotate them into an original visual essay — the selection and arrangement may qualify for copyright protection, even if the individual images don't. Similarly, if you use GitHub Copilot to suggest code and then substantially rewrite and structure it, your human contributions carry copyright. The Copyright Office issued guidance in March 2023 acknowledging this: AI-assisted works can receive protection for the human-authored elements. The line isn't where most people draw it, but a line does exist.

Myth 2: "AI Training on Public Content Is Legal Because It's Just Learning"

This analogy feels compelling: humans read millions of books, look at millions of paintings, and absorb that influence to become writers and artists. Why should AI be any different? The problem is that AI training isn't analogous to human learning — it's closer to mass copying at industrial scale. When OpenAI trained GPT-4, or when Stability AI trained Stable Diffusion, those systems ingested billions of web pages, images, and texts. Crucially, those works were ingested without license agreements, without payment, and often without the creators' knowledge. Several major lawsuits are now challenging whether this constitutes copyright infringement.

The cases are piling up fast. In January 2023, Getty Images sued Stability AI in both the UK and US, alleging that Stable Diffusion was trained on over 12 million Getty images without permission or compensation. In the same month, a class action was filed against Stability AI, Midjourney, and DeviantArt by a group of artists including Sarah Andersen, Kelly McKernan, and Karla Ortiz. In December 2023, The New York Times sued OpenAI and Microsoft, alleging that GPT models were trained on millions of Times articles. These aren't fringe complaints — they represent billions of dollars in claimed damages and will shape AI law for the next decade.

The legal defense most AI companies use is 'fair use' — a US doctrine that permits limited use of copyrighted material for purposes like commentary, criticism, or transformation. Whether AI training qualifies as fair use is genuinely unsettled. Some legal scholars argue it does, because the output is transformative. Others argue it doesn't, because AI companies are building commercial products that directly compete with the creators whose work trained them. The EU is taking a different approach: the EU AI Act and the 2019 Text and Data Mining exception allow AI training on publicly available data unless rights holders explicitly opt out. This is why you see some publishers, including Axel Springer and AP, signing licensing deals with AI companies rather than suing — they're betting on negotiation over litigation.

Myth 3: "AI Companies' Terms of Service Give Me Full Ownership"

Many professionals point to the terms of service of AI platforms as proof of ownership. OpenAI's terms do state that users own the outputs they generate. Midjourney's terms (for paid subscribers) similarly assign output rights to the user. This sounds reassuring — and for commercial use purposes, it is partially meaningful. But there's a critical distinction these terms cannot override: a company can contractually assign you rights they hold, but they cannot create copyright where copyright law says it doesn't exist. If a court determines that a given AI output lacks the human authorship required for copyright, no terms of service can manufacture that protection. The contractual assignment is real; the underlying copyright may not be.

There's also significant variation across platforms that most users never read carefully. Midjourney's free tier retains a license to use your generations for marketing and model improvement. Midjourney Pro tier gives you commercial rights but still allows Midjourney to use your outputs. Adobe Firefly is trained exclusively on licensed content and Adobe Stock images — which is why Adobe markets it as 'commercially safe' and offers indemnification for enterprise customers. Google's Gemini and Microsoft's Copilot both offer some level of copyright indemnification for enterprise users, meaning they'll cover legal costs if your use of their outputs triggers infringement claims. These differences matter enormously when choosing which tool to use for commercial work.

What Actually Works: Protecting Yourself and Your Work

Given this landscape, smart professionals adopt a layered approach rather than assuming any single protection covers them. The first layer is documentation. When you use AI tools to support creative work, keep records of your human contributions — your initial drafts, your editing decisions, your structural choices, the prompts you iterated through. The Copyright Office's March 2023 guidance makes clear that registration is possible for works where human authorship is present and identifiable. If you ever need to defend your copyright, the ability to show a creative process — not just a final output — is what distinguishes protectable work from public domain material. This is especially important for content you plan to commercialize or license.

The second layer is tool selection based on your use case. If you're creating content for personal productivity, internal documents, or low-stakes communications, the copyright ambiguity matters very little — use whatever tool works best. But if you're producing content that will be published, sold, licensed, or used in marketing materials, tool choice becomes a legal decision. Adobe Firefly's indemnification program, Microsoft Copilot's enterprise copyright commitments, and Google's Gemini for Workspace enterprise terms all offer meaningful (though conditional) protection that free-tier ChatGPT or Midjourney does not. The cost difference between a free tier and an enterprise plan with indemnification can be trivial compared to a single copyright dispute.

The third layer is disclosure and contract clarity. If you work with clients, vendors, or collaborators, your agreements should address AI use explicitly. This isn't just about legal protection — it's about trust and professional standards. Many clients now ask directly whether deliverables contain AI-generated content, and the professional answer isn't to hide it. It's to have a clear policy: which AI tools you use, for what purposes, what human review and modification process applies, and what copyright claims you can and cannot make on behalf of the client. Agencies, consultancies, and freelancers who have documented AI policies are increasingly winning business from clients who are nervous about AI-related liability and want a partner who has thought it through.

Frequently Asked Questions

• Can I register an AI-assisted work with the US Copyright Office? Yes, if you can identify and describe the human-authored elements. The Copyright Office reviews these case by case and has registered some AI-assisted works — including a comic book where the author wrote the story and dialogue but used Midjourney for images — while rejecting others where human contribution was minimal.
• Does copyright law work the same way in Europe? Not exactly. The EU's approach under the AI Act and existing copyright directives allows member states some flexibility, and several countries (including the UK) have specific provisions for computer-generated works. UK law actually allows copyright for 'computer-generated works' with the person who arranged the generation being treated as the author — a meaningfully different standard than the US.
• If I use Perplexity or ChatGPT to summarize an article, am I infringing the original article's copyright? Summarization is generally considered transformative and falls under fair use in the US, but reproducing large verbatim sections does not. The New York Times lawsuit against OpenAI specifically alleges that GPT models reproduce Times articles near-verbatim when prompted — that's the behavior that creates infringement risk, not summarization.
• What happens to content I post publicly that turns out to have no copyright? It effectively enters the public domain — anyone can use, copy, modify, or sell it without your permission and without compensating you. This is a real commercial risk for brands investing in AI-generated marketing content.
• Does using AI-generated content in my work affect my own copyright in the surrounding material? No — your original contributions retain whatever copyright they would otherwise qualify for. The AI-generated sections are simply unprotected. Think of it like quoting public domain text inside your own copyrighted essay — your original writing is still protected.
• Are there AI image tools that are safer to use commercially than others? Adobe Firefly is currently the most defensible option for commercial use — it's trained only on licensed Adobe Stock content and Adobe offers commercial indemnification for qualifying enterprise users. Getty Images has also launched its own AI image generator trained exclusively on Getty's licensed library. Both cost more than Midjourney or Stable Diffusion, but the legal clarity has real value for commercial applications.

Three Myths That Keep Professionals Exposed

Most professionals working with AI tools carry at least one of these beliefs: that AI-generated content is automatically protected by copyright, that using a paid AI tool means the output is legally safe to commercialize, or that because an AI created it, no one can claim ownership over it. All three are wrong — or at minimum, dangerously incomplete. The legal landscape around AI and copyright is moving fast, but courts and copyright offices have already issued enough rulings and guidance to dismantle these assumptions. Understanding where they break down isn't just academic. It directly affects how you use AI in client deliverables, marketing materials, product documentation, and anything you publish or sell.

Myth 1: AI-Generated Content Is Automatically Copyrighted

The instinct here is understandable. You open ChatGPT or Claude, spend twenty minutes crafting a detailed prompt, and receive a polished 800-word report. It feels like your work. You invested effort, intent, and creative direction. Surely that earns copyright protection? The U.S. Copyright Office says no — at least not automatically, and not for the AI-generated portions. Copyright law in the United States, the EU, and most major jurisdictions requires human authorship as a foundational condition. A machine cannot hold copyright. More critically, the courts have begun to clarify that simply prompting an AI doesn't, by itself, constitute the kind of human creative expression the law is designed to protect.

The clearest precedent so far is the Thaler v. Perlmutter case decided in August 2023. Stephen Thaler attempted to register copyright for an image created entirely by his AI system DABUS, listing the AI as the author. The U.S. District Court upheld the Copyright Office's refusal, ruling that human authorship is a prerequisite for copyright protection. What matters here isn't that AI was involved — it's that no human made the expressive choices in the final work. The Copyright Office has since issued guidance stating that purely AI-generated content receives no protection, but that works where humans make meaningful creative selections — choosing, arranging, editing AI outputs — may qualify for thin copyright covering those human contributions specifically.

This creates a practical spectrum. At one end: you type "write me a product description for running shoes" and publish the raw output. No copyright. At the other end: you use Midjourney to generate 200 images, manually select 12, arrange them in a specific narrative sequence, write original captions, and design a visual story around them. The selection and arrangement — the human creative layer — is likely protectable. The individual AI images are not. Kris Kashtanova's graphic novel Zarya of the Dawn landed exactly in this middle ground: the Copyright Office granted protection for the human-written text and arrangement, but stripped protection from the Midjourney-generated images. Your creative strategy needs to account for this distinction.

Myth 2: A Paid AI Subscription Means Safe Commercial Use

Paying $20/month for ChatGPT Plus or $30/month for Claude Pro does not transfer any copyright clearance for the outputs. It grants you a license to use the platform and, per OpenAI's and Anthropic's terms of service, assigns you ownership of the outputs you generate — but only to the extent those outputs are actually ownable. The terms don't and can't override copyright law. What the subscription does give you is a contractual right to use the output commercially, which is meaningfully different from copyright ownership. You have permission from the AI company. You don't have a guarantee that the output is free of third-party copyright claims from the training data.

This is where the training data problem resurfaces from Part 1. If an AI model was trained on copyrighted books, articles, code, or images without license, and if it reproduces substantial portions of that material in its outputs, the output may infringe the original creator's copyright — regardless of what you paid for your subscription. The AI company's terms of service typically include indemnification clauses that protect themselves, not you. OpenAI's enterprise terms do offer some indemnification for copyright claims related to outputs, but only for enterprise-tier customers and with meaningful caveats. For standard consumers and small business users, you bear the legal exposure if an output turns out to reproduce protected material.

GitHub Copilot has made this tangible. Copilot was trained on public GitHub repositories, including code under restrictive licenses like GPL. Researchers at Stanford found that Copilot reproduced verbatim licensed code in roughly 1% of suggestions — a small percentage that adds up enormously at scale. A class-action lawsuit filed in 2022 against GitHub, Microsoft, and OpenAI specifically targets this issue. If you ship code that Copilot generated and that code contains unlicensed GPL snippets, your product carries legal risk that your Copilot subscription does nothing to eliminate. The paid tool gives you capability. It doesn't give you legal clearance.

Myth 3: Because an AI Made It, Nobody Owns It

This myth runs in the opposite direction from Myth 1, and it's equally dangerous. The reasoning goes: if AI can't hold copyright and the user didn't author it in a traditional sense, then AI output must be in the public domain, free for anyone to use. This conclusion is wrong for two reasons. First, as established above, human creative contributions layered onto AI outputs can earn copyright protection for those specific elements. Second — and this is the part most people miss — the AI company itself retains certain rights over the model and its outputs through contract law, even when copyright doesn't apply. Your terms of service with OpenAI, Anthropic, or Google are licensing agreements, not ownership transfers.

The public domain assumption also ignores the rights of original creators whose work fed the training data. Several ongoing lawsuits — including Getty Images v. Stability AI, and the Authors Guild's class action against OpenAI — are directly testing whether training on copyrighted material without license constitutes infringement. If courts rule in favor of the original creators, the legal status of outputs generated from those models could be retroactively complicated. Treating AI output as unconditionally public domain exposes you to claims from original rights holders, not just from AI companies. The legal ownership question has at least three parties: the AI company, the user, and the creators of training data. "Nobody owns it" ignores two of the three.

What Actually Works: Practical Copyright Strategy for AI Users

The professionals who navigate AI copyright cleanly aren't waiting for perfect legal clarity — they're building repeatable workflows that create defensible human authorship, reduce exposure, and document their process. The first principle is to treat AI as a drafting tool, not a publishing tool. Everything ChatGPT, Claude, or Gemini generates is a first draft that a human meaningfully shapes before it goes anywhere public or commercial. This isn't just legal strategy — it produces better work. The act of editing, restructuring, and adding original insight is exactly what creates the human creative layer that copyright law recognizes. Use AI to accelerate your thinking, then make it yours.

The second principle is to use tools with stronger commercial licensing for high-stakes work. Adobe Firefly, for instance, was trained exclusively on licensed Adobe Stock images and public domain content — a deliberate architectural choice that Adobe markets directly to enterprise customers who need clean IP. Similarly, Getty Images launched its own generative AI tool trained only on its licensed library, with built-in indemnification for commercial use. These tools cost more and produce narrower outputs than Midjourney or DALL-E 3, but for advertising campaigns, product packaging, or any work where IP chain-of-title matters, the premium is worth it. Matching your tool to your use case's legal requirements is a professional skill now.

The third principle is documentation. Create a simple log for any AI-assisted work that goes public or commercial: what tool and version you used, the date, the prompt you entered, what you changed, and what human creative decisions you made. This takes two minutes per piece of content and creates an evidence trail if your authorship or originality is ever challenged. Some organizations are building this into their content management systems — tagging AI-assisted content at creation with metadata about the human editorial process. It sounds like overhead, but it's exactly the kind of operational practice that separates teams who use AI confidently from those who use it nervously and retroactively scramble when questions arise.

Frequently Asked Questions

• Can I register a copyright for content I wrote with AI assistance? Yes — if you made substantial human creative contributions. The U.S. Copyright Office now accepts applications for AI-assisted works but requires you to disclose AI involvement and describe the human authorship. They evaluate case by case.
• Does it matter which AI tool I use for copyright purposes? Significantly. Tools trained on licensed or public domain data (Adobe Firefly, Getty's generator) carry far less training-data exposure than tools with broader, less documented training sets. The tool's training provenance directly affects your downstream risk.
• If I substantially rewrite an AI output, is it fully mine? The rewritten portions are yours if they reflect original human expression. Courts look at whether the final work reflects human creative choices, not just mechanical editing. A full structural rewrite with original ideas generally qualifies.
• Can my employer claim ownership of AI content I produce at work? Almost certainly yes — standard work-for-hire doctrine assigns IP created in the scope of employment to the employer, AI-assisted or not. Check your employment contract; many companies are now adding explicit AI IP clauses.
• What happens to my AI-generated content if a lawsuit against the AI company succeeds? Legal outcomes are genuinely uncertain here. If courts find that training was infringing, it's possible that commercially published outputs could face retroactive claims — though practically, mass enforcement against individual users is unlikely. Enterprise customers with indemnification clauses are better positioned.
• Is AI-generated code treated differently from AI-generated text or images? The same copyright principles apply, but code has additional complexity because many training repositories used licenses (GPL, MIT, Apache) that impose obligations on derivative works. A GPL-licensed code snippet reproduced in your product may trigger copyleft requirements even if you didn't know it was there.

Key Takeaways from This Section

1. Raw AI output receives no copyright protection in the U.S. and most major jurisdictions — human authorship is a legal prerequisite, and prompting alone doesn't satisfy it.
2. A paid AI subscription grants you contractual usage rights to outputs, not copyright clearance. Third-party content embedded in training data remains a separate legal exposure the subscription doesn't resolve.
3. "Nobody owns AI output" is as wrong as "the user owns it automatically" — ownership is contested across the AI company, the user, and training data creators simultaneously.
4. Human creative contributions layered onto AI outputs — selection, arrangement, editing, original additions — can earn copyright protection for those specific elements.
5. Tool choice matters: Adobe Firefly and Getty's generative AI are trained on licensed content and offer stronger commercial IP warranties than general-purpose models.
6. Documentation is your primary defense: logging your prompt, your edits, and your creative decisions creates the evidence trail that supports authorship claims and demonstrates due diligence.
7. International copyright law is not uniform — the EU, UK, and China each treat AI-generated works differently, and jurisdiction matters for any cross-border commercial use.

Three More Myths That Will Cost You If You Believe Them

Most professionals working with AI tools carry three quiet assumptions about copyright that feel logical but don't hold up under legal scrutiny. They assume that adding human edits to AI output automatically creates copyright protection. They believe that using AI-generated content commercially is safe as long as you paid for the tool. And they think copyright law will sort itself out soon, so current ambiguity is no big deal. All three assumptions can expose you, your team, or your clients to real legal and reputational risk. The corrected versions of these beliefs will change how you work with AI-generated content starting today.

Myth 1: Editing AI Output Makes It Yours

The intuition here is understandable. You take a ChatGPT draft, rewrite the opening, restructure three sections, and swap in your own examples. Surely that's your work now? Copyright law doesn't see it that way — at least not entirely. What you've created is likely a hybrid: the portions you authored independently may be protectable, but the underlying AI-generated scaffolding remains in a legal gray zone. Courts and copyright offices don't yet have a clean formula for calculating what percentage of human contribution 'unlocks' copyright.

The U.S. Copyright Office has been explicit on this point. In its February 2023 guidance on Kristina Kashtanova's graphic novel 'Zarya of the Dawn,' the office granted copyright only to her written text and the specific arrangement of pages — not to the Midjourney-generated images themselves, even though she directed the visual style extensively. The bar for human authorship isn't effort or intent. It's whether a human made sufficiently specific expressive choices that the output couldn't have been produced without that particular human's creative decisions.

A better mental model: think of AI output as raw material, like stock footage or a clip-art library. You can build something copyrightable on top of it, but the raw material itself doesn't become yours through use. Document your specific creative decisions — the prompts you rejected, the structural choices you made, the edits that changed meaning rather than just grammar. That documentation is your evidence of authorship if it's ever challenged.

Myth 2: Paying for the Tool Means You Own the Output Commercially

Paying $20/month for ChatGPT Plus or $96/year for Midjourney Basic grants you a license to use outputs — not copyright ownership. These are different things. A license tells you what you're allowed to do with the content under the platform's terms. Copyright ownership would mean you could stop others from using similar outputs, register the work, and sue for infringement. Platforms are deliberately careful never to promise the latter, because they legally can't deliver it for AI-generated content.

OpenAI's terms of service assign you ownership of outputs 'to the extent permitted by law' — which currently means very little for purely AI-generated text in the U.S., given the Copyright Office's stance on human authorship. Adobe Firefly's commercial license protects you from third-party infringement claims (Adobe indemnifies you), but that's a contractual shield, not copyright in the traditional sense. Midjourney's terms for paid users allow commercial use but explicitly don't transfer copyright. Each platform has different rules, and conflating 'licensed for commercial use' with 'I own this' is how companies create quiet legal exposure.

The practical implication hits hardest in client work. If you deliver AI-generated copy, visuals, or code to a client and invoice it as original creative work, you may be misrepresenting what they're receiving. Many client contracts include warranties that work is original and non-infringing. AI-generated content delivered without disclosure can void those warranties. Disclose your use of AI tools in contracts, specify what rights you're actually conveying, and don't promise exclusivity on content that legally can't be exclusive.

Myth 3: The Law Will Catch Up Soon, So This Is Temporary

Legal reform on AI copyright is moving — but slowly, unevenly, and in ways that may not resolve the core ambiguity. The EU AI Act, finalized in 2024, focuses primarily on risk classification and transparency obligations, not on copyright ownership for outputs. In the U.S., Congress has held hearings but passed no AI-specific copyright legislation as of mid-2025. Meanwhile, active litigation — including The New York Times vs. OpenAI, Getty Images vs. Stability AI, and multiple music industry suits — is creating case law piece by piece. Waiting for clarity before building compliant workflows is waiting for a bus that has no published schedule.

What is becoming clearer through litigation isn't who owns AI outputs — it's who bears liability for training data. Courts are increasingly scrutinizing whether AI companies had the right to train on copyrighted material at scale. That question doesn't directly resolve output ownership, but it shapes the risk profile of the tools you use. Models trained with licensing agreements (like Adobe Firefly, trained on licensed Adobe Stock) carry a materially different risk profile than models with less transparent training data provenance.

What Actually Works: Practical Approaches for Professionals

The professionals navigating this best aren't waiting for legal certainty — they're building systematic habits that reduce exposure while letting them move fast. The first habit is tool selection based on training data transparency. Adobe Firefly and Getty's Generative AI are trained on licensed content and offer indemnification for commercial use. That's a concrete, contractual protection. OpenAI, Anthropic, and Google offer commercial use rights but no indemnification for output similarity to training data. Matching your tool to your risk tolerance — not just your output quality preference — is a professional decision, not a technical one.

The second habit is documentation. Keep a simple log of how you used AI in any deliverable: which tool, what prompts, what percentage of the final output is AI-generated versus human-authored. This serves two purposes. First, it supports copyright claims for your human contributions — you can point to specific creative decisions. Second, it creates an audit trail if a client, employer, or court ever questions the provenance of the work. A shared Google Sheet or a note in your project management tool takes five minutes per project and is worth considerably more than that in risk reduction.

The third habit is contract clarity. Any agreement involving AI-assisted deliverables should specify what rights are being conveyed, disclose AI tool usage, and avoid warranties of originality or exclusivity for AI-generated portions. If you're on the buying side — commissioning work from an agency or freelancer — ask explicitly whether AI tools were used and request the same disclosure clause. This isn't about being anti-AI. It's about ensuring that both parties understand what they're actually exchanging, which is the foundation of any enforceable contract.

Frequently Asked Questions

• Can I register AI-generated content with the U.S. Copyright Office? You can register works that contain sufficient human authorship — the Copyright Office will accept applications but may exclude AI-generated portions from protection. Disclose AI involvement on the application; omitting it can invalidate the registration.
• Does copyright law differ for AI-generated code versus text or images? The same human authorship standard applies, but code has additional complexity: functional elements are often not copyrightable even when human-written. GitHub Copilot's terms assign output ownership to you, but the underlying legal protection for purely AI-generated code is similarly uncertain.
• If I use AI to generate a logo for my business, can a competitor copy it? If the logo is substantially AI-generated with minimal human creative input, it may have no copyright protection — meaning yes, legally, a competitor might copy it without liability. Add distinctive human-designed elements and document your design decisions to strengthen any potential claim.
• Are AI-generated images from tools like DALL-E or Midjourney safe for use in published books? They're commercially licensed for paid users, but not copyright-protected in the traditional sense. Several publishers now require authors to disclose AI-generated content; check your publisher's submission guidelines before including such images.
• What happens if an AI tool I used is later found to have infringed copyright in its training data? Your liability depends on the platform's indemnification policy. Adobe Firefly and a few others offer indemnification; most do not. This is a key reason to prefer tools with explicit indemnification for high-stakes commercial work.
• Does the EU have different rules on AI-generated content ownership? The EU AI Act doesn't directly address output ownership, but EU member states' existing copyright frameworks generally require human authorship for protection — similar to the U.S. stance. The EU's approach to transparency obligations may indirectly create disclosure norms that affect how AI content is treated contractually.

Key Takeaways

1. Editing AI output does not automatically create copyright — courts look for specific, distinctive human creative choices, not effort or labor.
2. A commercial use license from an AI platform is not copyright ownership; you cannot promise exclusivity to clients for AI-generated content.
3. Tool selection is a risk management decision: platforms with licensed training data and indemnification (Adobe Firefly, Getty) carry lower commercial risk than those without.
4. Legal clarity on AI copyright is not arriving soon — building compliant workflows now is the only responsible path.
5. Documentation of your creative decisions and AI tool usage is your primary evidence of authorship and your first line of defense in disputes.
6. Contract disclosure of AI tool usage is not optional in professional work — omitting it can void warranties and create breach-of-contract liability.
7. The portions of a deliverable you author with genuine creative specificity remain protectable — hybrid human-AI works can have partial copyright protection.