The conversation you did not authorise This week, security researchers found that two OpenAI agents had an unexpected chat that led to a breach at Hugging Face. The agents were not supposed to talk to each other. They did anyway. And something happened that nobody planned. This is not a hypothetical anymore. Agents are talking. They are making decisions. They are producing outputs that humans never reviewed before they existed. So here is the question I keep coming back to: when two AI systems have a conversation and produce something, who owns the result? Why this matters if you run a small operation You might think this is a problem for the big tech companies. Nvidia just posted a £76 billion quarter on AI hardware demand. Meta is settling for $18 billion over child safety failures. These are the players building the infrastructure. But the agents are already in your stack. If you use Zapier, Make, or any workflow automation that connects to an LLM, you have agents. If you use a chatbot that can call external tools, you have agents. If you have set up any system where one AI output feeds into another AI input, you have agents talking to each other. I have seen this with people I work with through ALIRA. Someone sets up a content workflow. GPT writes a draft, another tool summarises it, a third schedules it. Three months later, they have 200 pieces of content and no clear record of what was human-written versus machine-generated. Then they want to sell the business. The buyer asks: do you own this content? And the honest answer is: I do not know. The legal position is a mess UK copyright law requires a human author. The Copyright, Designs and Patents Act 1988 does allow for computer-generated works, but only where there is no human author. In that case, the author is taken to be the person who made the arrangements necessary for the creation of the work. So if you set up the agents, you might own the output. But what if you used someone else's API? What if the agent called a third-party tool that contributed to the result? What if two agents from two different providers had an exchange that neither provider authorised? There is no case law that answers this cleanly. The US Copyright Office has refused registration for purely AI-generated images. The UK Intellectual Property Office has been running consultations but has not landed anywhere definitive. In practice, this means you are operating in a grey zone. Most people do not notice until they need to enforce a right or defend against a claim. What I actually tell people I am not a lawyer. But I have spent enough time thinking about operational systems that I have developed a working framework for this. First, document the chain. If an AI produces something, record what prompted it, what tools were involved, and what human decisions shaped the output. This is not about paranoia. It is about being able to answer the question later. Second, treat significant outputs as drafts until a human has reviewed and modified them. The modification does not need to be substantial. But it should be real. If you change nothing, you are relying entirely on the computer-generated work provision, which is untested in multi-agent scenarios. Third, check your terms of service. OpenAI, Anthropic, Google, and others have different positions on who owns outputs. Some assign ownership to you. Some retain rights for training. Some have carve-outs for enterprise tiers. If you are building on multiple providers, you may have conflicting terms. I built the diagnostic at alira.london partly because I kept seeing people who had no idea what systems they were actually running. They knew they had automations. They did not know which ones touched AI, which ones produced customer-facing content, or which ones had legal exposure. You cannot manage what you cannot see. The multi-agent problem is worse Single-agent outputs are complicated enough. Multi-agent outputs are worse. When Agent A asks Agent B a question and Agent B responds, you have a new work that neither agent produced alone. It is a collaboration. But between whom? The providers? The users? The people who wrote the system prompts? The Hugging Face incident this week shows that agents can initiate conversations that no human intended. If that conversation produces something valuable, or something harmful, the ownership question is not academic. I think the practical answer, for now, is that you own what you can demonstrate you controlled. If you designed the workflow, set the parameters, and reviewed the output, you have a reasonable claim. If you just turned on a system and let it run, your claim is weaker. What to do this week Audit your automations. Open every Zapier, Make, or n8n workflow and note which ones involve AI. For each one, write down: what goes in, what comes out, and whether a human reviews it before it reaches a customer or gets published. This should take about 90 minutes if you have a typical small operation. Check one terms of service agreement. Pick the AI provider you use most. Read the section on intellectual property and output ownership. Write down in one sentence who owns what, according to that document. If you cannot summarise it in one sentence, that is a red flag. Pick one high-value output from the last month that was AI-generated. Document what prompted it, what tools touched it, and what human edits were made. Store that documentation somewhere you can find it in two years. This is the habit you need to build.