Data and intellectual property exposure

The Shared Photocopier

Somewhere in your client’s business, this afternoon, someone is going to paste a document into an AI tool to get help summarizing it. They aren't being reckless; they’re trying to do their job faster with a tool that’s genuinely good at this. In that moment, two things are set in motion: something of theirs has gone out, and something of someone else’s may be coming back. Think of a photocopier in a shared office that keeps a copy of everything it scans. Occasionally, it also slips a stray page from someone else’s job into your stack. Nobody set out to leak anything, and nobody set out to take anything. Both are properties of the machine rather than anyone’s intent. Where the analogy breaks down is that a photocopier’s stray page is obvious when you find it. AI-generated material carries no marking at all, which is where the real difficulty lies.

Data Going Out: The Prompt Problem

AI exposes data at three main points: what goes into training, what goes into a prompt, and what comes out in the results. While training gets the headlines, the prompt is the route that actually happens every day. When an employee pastes client data, confidential strategy documents, or trade secrets into a consumer AI tool, they are often trying to be productive. This isn't misconduct; it’s well-intentioned behavior creating data leakage.

Shadow AI and Dual Harms

This is shadow AI: use that the organization hasn't sanctioned and cannot see. There are two distinct harms here. First, personal data—violations of privacy through the exposure of individual information. Second, confidentiality and trade secrets—the loss of commercially valuable information whose protection depends on it being kept secret.

Recap: Data Outward

Let's pause to recap. We've seen that AI exposes data at three points, but the prompt is the one that happens every day through ordinary staff behavior. Shadow AI means the exposure an organization cannot see is usually much larger than the exposure it can. Remember: these aren't hacks; it's data leakage driven by productivity.

The IP Tangle: Three Questions

When we use generative AI, we also face the risk of what flows into our business. This raises three separate IP questions that we must keep clearly apart. One: what went into training? Is it legal for AI companies to train their models on copyrighted material? This is being actively litigated and is currently unsettled. Two: who owns the output? Ownership is genuinely unclear. If your client uses AI to write code, that material may not be protectable by copyright at all. Three: does the output infringe? If a model produces something very similar to a specific artist's work, using that output might make you liable.

Contamination and Provenance

This leads to contamination: generated material entering a product, a codebase, or a marketing campaign while carrying unresolved provenance—or a traceable origin. The exposure lands on the user, not just the AI developer. You're bringing material into your business without knowing its legal DNA.

The Litigation Wave

We are currently in a massive wave of generative AI copyright litigation. The scale is significant, but the answers are still shifting. In the US, authors reached a settlement with Anthropic reported at 1.5 billion dollars—the largest known copyright payout in history. The New York Times case against OpenAI remains in discovery through 2026, with a trial expected late that year. Meanwhile, in the UK, Getty largely lost its copyright claims against Stability AI, though there was a limited finding regarding trademarks. Notice the divergence: the same technology is receiving different answers in different countries.

Summary and The Explicit Stop

To wrap up: AI exposes data at three points, but the prompt is the one that happens every day through ordinary staff behavior. Shadow AI means the exposure you can't see is usually much larger than the one you can. And generated material raises three unsettled questions: what it trained on, who owns it, and whether it infringes. We're not going to tell you how any of these cases will end, because nobody knows. And what a business should do about any of it is the next course. What you need is the shape of both exposures, and the habit of asking two questions: what has been typed into these tools, and where did the output come from?