The owner of a 22-person services firm has read four news stories about AI and copyright in the last month. One said authors won. One said the AI companies won. One said the UK government had stepped back from a copyright exception. One said the EU was tightening transparency rules. She has not formed a useful opinion and is starting to worry she should have one. Her firm uses ChatGPT and Claude every day. Three of her staff have just asked whether they are still allowed to.
This post is for her. The headline copyright fights are loud, complicated, and mostly happening between large rights holders and large AI labs about how the underlying models were built. Very little of that changes what a 5 to 50 person firm should do on Monday morning. The bits that genuinely apply are narrower and worth knowing precisely, so the owner can give her team a clear answer and get on with the work.
What is the AI copyright debate actually about?
The debate is about whether AI labs needed permission to train their models on copyrighted books, articles, images, lyrics, and code scraped or downloaded from the internet. The New York Times is suing OpenAI on that basis. Authors sued Anthropic. Music publishers sued Anthropic. Getty Images sued Stability AI in the UK. The labs argue training is protected fair use. The rights holders argue the labs took commercial value without paying for it.
The cases so far have produced a split picture. In June 2025, Judge Alsup ruled in Bartz v Anthropic that training on lawfully acquired books was fair use, while keeping pirated copies was not. Anthropic settled the piracy side for $1.5 billion. Judge Chhabria reached the same fair-use conclusion in Kadrey v Meta two days later. In November 2025, the UK High Court ruled in Getty v Stability that the final Stable Diffusion model does not contain copies of the training images and so is not a secondary-infringement article. The Thomson Reuters v ROSS Intelligence decision, the outlier, rejected fair use because ROSS was building a direct competitor to the rights holder’s product.
Why does it matter for your business?
It matters less than the noise suggests, in a precise way. None of the active cases is about whether an end user can use a commercially available AI tool. They are about how the model was built. No court anywhere has held a small business liable for using ChatGPT, Claude or Copilot in its normal work, and the major vendors carry enterprise indemnities for output-side claims.
Where it does matter is on the other side of the workflow. The vendor indemnities from OpenAI, Microsoft, Anthropic, Google and Adobe cover claims that the model’s training data or output infringes a third party’s intellectual property right. They do not cover what you paste in. If your firm puts a client’s confidential document into a free-tier consumer AI tool that uses customer content for model improvement, the vendor’s training-data indemnity does nothing for you. The exposure is your contract with the client and the IP rights of whoever owns the material. That distinction, vendor covers their side, you cover your input, is the load-bearing one for an SME using AI tools every day.
Where will you actually meet it?
You will meet it in four places in the firm. AI-assisted client deliverables, where the question is whether the output resembles existing published work too closely. Third-party content put through AI, where the question is whether you had the right to use the material. Confidential client data pasted into consumer AI tools, where the vendor’s terms may breach your duty of confidence. And contractual disclosure of AI use to clients.
The first two are output-side. The third is input-side and is the most common live exposure in owner-led firms today. The fourth is the cheapest to fix and the most often left undone. A single sentence in the schedule of a services contract usually does the job. Read across to where your data goes when you paste it into a chatbot for the data-flow side, and to who owns the work when AI wrote it for the ownership question, which sits next to copyright but is a distinct issue.
When to ask versus when to ignore?
Ask when AI is touching client confidential information, regulated material, third-party content you do not own, or work the firm warrants as original. In those situations the answer is to use a paid tier with terms that prohibit training on customer content, get the client’s contractual consent, and have a human review the output before delivery. Ignore the temptation to track every news story and tune the firm’s practice to it.
The single durable rule, which holds whichever way the law moves in the US, UK or EU, is this. Use AI tools only on data and content your business owns outright or has a clear documented right to use. Owned client data goes through a paid tier with the right terms. Third-party content goes through AI only if you have permission to use it. Generated output gets reviewed by a human before delivery. That rule sits underneath every published case so far and does not depend on which way the next one goes. It is also short enough to brief the team on in a five-minute conversation.
Related concepts
This sits inside the IP, ownership and disclosure section of a wider cluster on AI risk, trust and governance for owner-led firms. The neighbouring posts cover the questions that sit next to copyright without being the same question. Ownership of AI-assisted deliverables is a separate matter from whether the underlying model was trained legally, and is worth reading alongside this one.
For the ownership question, see who owns the work when AI wrote it. For the contractual disclosure question, disclosing AI use to customers. For the input-side exposure that vendor indemnity does not solve, where your data goes when you paste into a chatbot. For the wider regulatory picture, the EU AI Act for UK and EU SMEs covers the transparency obligations on general-purpose AI providers, and UK AI regulation after the pro-innovation pivot covers the UK position after the March 2026 report. None of these posts is legal advice. Each one is the proportionate version of a question owner-led firms keep being asked, sized for a firm the owner can see across in a single room.
If you have read four contradictory news stories about AI and copyright and want twenty minutes to talk through what actually applies to your firm, book a conversation.
