It started with a simple question from my business partner, Claire. We run a small content agency — twelve people, mostly remote, operating out of coffee shops and coworking spaces across three time zones. In January, we started using AI tools to help draft client deliverables. Blog posts, social media copy, email sequences. The usual.
Claire, who handles contracts, looked up from her laptop one Tuesday and asked: "So... who owns the stuff our AI writes?"
I said, "We do. Obviously."
She said, "Are you sure?"
I was not sure. And three weeks later, after reading more legal opinions, court filings, and policy papers than any reasonable person should, I am even less sure. But I learned a lot. And some of it is genuinely alarming for anyone building a business that involves AI-generated content.
The Current Legal Landscape (It Is a Mess)
Let me save you the three weeks of research and summarize where things stand as of early 2026:
The US Copyright Office Position
The US Copyright Office has been clear-ish on one thing: purely AI-generated content cannot be copyrighted. Their reasoning is that copyright requires a human author, and an AI is not a human. This came from the Thaler v. Perlmutter case in 2023, where a court ruled that an AI-generated artwork could not receive copyright protection because "human authorship is a bedrock requirement."
Simple, right? Not even close.
The problem is defining "purely AI-generated." If I type a one-line prompt and the AI produces a 2,000-word article, that is probably not copyrightable. But if I write a detailed 500-word prompt, specify the structure, edit 40% of the output, add my own examples, and restructure the argument? At what point does my human contribution create a copyrightable work?
The Copyright Office says it is a "case-by-case" determination. Which is lawyer-speak for "we have no idea either."
The Kapwing Case (Happening Right Now)
This is where it gets interesting. Kapwing, a video editing platform, recently published a detailed account of their experiment paying artists royalties for AI-generated art. The article, which was trending on Hacker News today with over 100 upvotes and heated debate, reveals the messy reality of trying to do the right thing when the legal framework does not exist yet.
Their approach: they trained their AI model on licensed artwork and paid royalties to the original artists when the AI generated images "inspired by" their style. Sounds fair. Except they quickly ran into questions nobody could answer:
- How do you calculate a fair royalty when the AI combined elements from 47 different artists?
- What if the AI output does not look like any specific artist's work but clearly could not exist without the training data?
- Who arbitrates disputes when an artist says "that looks like my style" and the AI company says "it is a coincidence"?
Kapwing's honest conclusion was essentially: "We tried. It is really hard. We are still figuring it out." Which is more transparency than most AI companies offer, but also not exactly reassuring.
The Copyleft Erosion Problem
Meanwhile, a different angle of this debate is exploding in the open-source community. A piece that went viral this week (417 points on Hacker News, over 450 comments) argues that AI is fundamentally eroding copyleft licenses like the GPL.
The argument goes like this: if you write open-source code under a GPL license, anyone who uses your code must also make their derivative work open-source. That is the whole point of copyleft — it keeps the chain of openness going. But what happens when an AI model is trained on GPL code and then generates "new" code that is functionally identical but technically not a copy?
Is the AI output a derivative work? The AI companies say no. The open-source community is livid. And the courts have not decided.
My developer friend Jake — not a lawyer, but someone who has contributed to open-source projects for fifteen years — put it bluntly: "They are using our work to train their models and then selling the output. If a human did that, we would call it plagiarism. But because a computer did it, suddenly it is innovation."
I am not saying Jake is right. I am saying his frustration is shared by a lot of people who build things for free and are watching companies profit from their generosity.
What This Means for Businesses (Like Mine)
After three weeks of research, here is what I concluded for our agency:
The Risk Is Not Zero
If we deliver AI-generated content to clients and someone later challenges the copyright status of that content, our clients could be exposed. Could a competitor copy our client's AI-generated blog post and legally argue it is not copyrightable? In theory, maybe. Nobody has tested this in court yet for commercial content, but the legal foundation for that argument exists.
The "Substantial Human Contribution" Standard
Based on everything I read, the safest position is to ensure every piece of content has substantial human contribution. Our new process:
- AI generates a first draft based on our detailed brief (human input: the strategy, angle, target audience, key points)
- A human writer rewrites at least 30-40% of the content, adding original examples, opinions, and analysis
- A human editor reviews for quality and adds their own perspective
- We document our human contributions for each piece (yes, this is annoying; no, we cannot skip it)
Is 30-40% human rewriting enough to establish copyrightability? Honestly, I do not know. Nobody knows. But it is a defensible position, and that is the best we can do right now.
The Client Communication Problem
We had to update our client contracts. The old language said we "assign all intellectual property rights" in the delivered content. But can we assign IP rights in something that may not be copyrightable?
Our lawyer — who charged us $400 for this advice and somehow managed to say "it depends" fourteen times in a one-hour call — recommended adding a disclosure clause. We now tell clients: "Portions of this content were created with AI assistance. We ensure substantial human editorial contribution, but the copyright status of AI-assisted works is currently evolving under applicable law."
Every client so far has said, "Yeah, fine, whatever." But at least we are covered if the legal landscape shifts.
The EU Is Ahead (Sort Of)
The EU AI Act, which started rolling out in phases in 2025, requires companies to disclose when content is AI-generated. It does not directly address copyright ownership, but it creates a framework for transparency that the US completely lacks.
If you are serving European clients or have European users, you should already be labeling AI-generated content. This is not optional. The penalties for non-compliance are steep — up to 3% of global annual revenue for certain violations.
My take: the EU approach is not perfect, but at least it is an approach. The US is still in the "we will figure it out eventually" phase, which in practice means thousands of businesses are operating in a legal gray zone and hoping nobody sues them. That is not a strategy. That is a prayer.
What About Training Data?
This is the other half of the copyright debate, and arguably the more explosive one. Multiple lawsuits are currently working through the courts:
- The New York Times vs. OpenAI (filed 2023, still ongoing in 2026)
- Authors Guild vs. OpenAI (class action, still ongoing)
- Getty Images vs. Stability AI (ongoing)
- Various music industry lawsuits against AI music generators
None of these have reached final judgment. The outcomes will shape AI copyright law for decades. But if courts rule that training on copyrighted material without permission constitutes infringement, the entire AI industry could face a reckoning. Every major language model and image generator has been trained on copyrighted content. Every single one.
I asked our lawyer about this. He said — and I swear this is a direct quote — "If the courts rule against the AI companies on training data, you will hear the sound of a thousand startups simultaneously drafting their wind-down plans." Then he charged me for another 15 minutes.
Practical Steps for Content Creators and Businesses
Here is what I would recommend based on my three weeks in the rabbit hole:
- Document your human contributions. Keep notes on what the AI generated vs. what humans wrote. This creates an evidence trail if copyright questions ever arise.
- Add disclosure clauses to your contracts. Do not wait for the law to force you. Get ahead of it.
- Do not rely on AI-generated content as your sole competitive advantage. If it turns out that content cannot be copyrighted, your competitors can legally use it. Build moats elsewhere.
- Watch the major lawsuits. NYT v. OpenAI and Authors Guild v. OpenAI will set precedents that affect everyone. Subscribe to updates.
- If you serve EU clients, comply with the AI Act now. Not later. Now. The penalties are real.
- For open-source developers: consider your license choices carefully. If copyleft cannot be enforced against AI-derived works, permissive licenses (MIT, Apache) might be more honest about the reality of how your code will be used.
My Honest Take
Three weeks ago, Claire asked me who owns what AI creates. After 50+ hours of research, my answer is: "Probably nobody, maybe us, possibly everyone, and definitely not the AI."
That is not a satisfying answer. But it is the honest one. The law has not caught up to the technology, and until it does, we are all making educated guesses and hoping for the best.
What I do know is that ignoring this question is not an option. If your business uses AI to create content — and statistically, it probably does — you need to have a plan for what happens when the legal dust settles. Because it will settle. And when it does, some businesses will be ready, and some will be scrambling.
I would rather be in the first group. Even if it means paying my lawyer to say "it depends" a few more times.
I am not a lawyer, and none of this is legal advice. Talk to an actual attorney about your specific situation. Preferably one who charges less than mine.
Related Reading
While copyright law catches up, see how businesses are navigating AI costs in our 6-month AI spending analysis, and explore the broader implications in the 8 levels of agentic engineering. For the open source angle, see why Hacker News banned AI comments.
