Samsung's 2023 ChatGPT incident, as Troutman Pepper Locke recounts it, shows how a routine debugging prompt becomes a governance failure: engineers uploaded proprietary semiconductor source code to a public model, and the company restricted external generative AI soon after. The Business Standard reported that the uploads happened while employees were seeking help with their work. That is the uncomfortable part. The risk sits with a capable person under deadline using the nearest tool.
UK firms are in the same position, with less excuse. AI is in browsers, office add-ins, developer plugins and meeting recorders. The governance question is no longer whether staff use it. It is whether you can evidence a decision about which tools may touch which information.
The legal test your AI use has to survive
UK trade secret protection depends on conduct: the steps you take to keep information secret.
The Trade Secrets (Enforcement, etc.) Regulations 2018 require the holder to take reasonable steps to keep information secret. Mayer Brown's comparison of UK and US practice sets out the scope: source code, training data, model weights and proprietary algorithms can all qualify, and protection can last indefinitely while secrecy holds. The same analysis notes that there is no UK judicial guidance yet on what reasonable steps means in practice. That gap is not comfort. A court, an investor or an acquirer will read your AI policy, your logs and your training records against a standard that is still forming.
Two routes carry the loss.
- A training leak happens when confidential material is used to train a model and later appears in its output, or when your own model absorbs material you had no right to use.
- A third-party prompt leak happens when an employee sends a file, a code snippet or a strategy note to an external platform whose terms allow the provider to use inputs, or whose outputs reach other users.
IAM's May 2025 guidance on complying with UK trade secret legislation links the second route to jurisdiction: data sent overseas to countries with weaker trade secret safeguards is difficult to square with a duty to take all reasonable measures. It also recommends mapping which trade secrets could interact with AI, cataloguing every internal and third-party AI tool, and reviewing supplier contracts for confidentiality, data reuse and residency terms. That is a governance programme, with legal, security and business owners accountable for each part.
Accidental disclosure is only half the problem. Deliberate extraction now looks like a conversation. Quinn Emanuel's April 2026 client alert describes how an employee can ask an enterprise AI to summarise pricing strategy, product status or a proprietary process, then collate answers across sessions. No file is downloaded. Bulk-download alerts do not fire. Memory export and over-permissioned retrieval make it worse. The forensic question becomes whether a pattern of queries reproduces the secret.
The US case law is moving faster than UK guidance. Troutman Pepper Locke reported that a January 2026 decision from the Northern District of California dismissed a trade secret claim because the plaintiff had developed the alleged secret through ChatGPT, which the court treated as voluntary disclosure to OpenAI. The same firm notes that courts increasingly ask whether a company addressed AI as a specific disclosure risk alongside its general confidentiality policy. UK boards should read that as a direction of travel that will reach them. The UK test is reasonable steps. A general NDA that predates generative AI is weak evidence that you took them.
The exposure runs past a policy refresh. An organisation that wants to enforce a trade secret right generally has to show it took reasonable steps to preserve secrecy, and routine sharing with unauthorised AI platforms weakens that case. Enterprise AI tools reduce the exposure. They leave internal governance to you. That is the correct order. The contract sets the floor. Your operations decide whether you can stand on it.
What to fix before the next prompt
The first deliverable is a written answer to a simple question: which AI tools may touch which classes of information?
Most organisations cannot produce it. Employees get a general confidentiality clause, a browser bar and no approved list. That is not a training problem. It is a control design problem. The fix is to classify information before it reaches a model, then bind each class to a tool and a permission.
A consumer AI account usually gives the provider broad rights over inputs, offers no contractual confidentiality commitment, and gives administrators little visibility. It is suitable for public material and nothing else.
An enterprise AI agreement usually adds confidentiality commitments, administrative controls, audit logs, retention terms and a default that customer data is not used to train public models. It is the minimum for internal material, but only if the contract actually says so and the controls are switched on.
A self-hosted or private model keeps data inside your environment and can be the right home for source code, pricing models or other material that would be fatal to disclose. It also concentrates responsibility on your own access controls, logging and patching.
The contract terms to check are narrow and testable. Does the provider use inputs for model training? How long does it retain them? Which subprocessors can see them? Where are they stored? Can you audit? What happens on deletion? If the answer is unclear, the tool is not approved for confidential material. Enterprise tier is not a magic word. Some enterprise plans still carry broad data usage rights in standard terms.
The operational controls matter more than the label. Klemchuk's analysis of AI chatbot trade secret leaks makes the point that the largest leak is usually internal and accidental, one prompt at a time. The practical response is not a ban. It is a short approved list, a data classification scheme staff can apply in ten seconds, prompt hygiene that removes customer names and deal figures, and logs that show who used which tool with which data class. A ban drives use to personal accounts and removes the logs.
Training needs to be specific. "Do not share confidential information" is not an instruction an employee can follow at the prompt bar. "Never paste source code, customer lists, unpublished pricing, board papers, litigation strategy or pre-filing patent material into a tool that is not on the approved list" is. Give examples of a bad prompt and a sanitised one. Test completion. Record it. A court will not accept a policy that no one can remember.
Board oversight has to move from IT to corporate governance. The source article argues that AI governance belongs in the boardroom because confidentiality, intellectual property, regulatory compliance and risk appetite are board responsibilities. That is right. The board does not need to approve every tool. It needs to approve the risk appetite, receive the audit trail, and ask one question each quarter: can we evidence that our most valuable information did not enter an unapproved model?
The board question is not whether AI is allowed. It is what evidence exists that the rule was followed.
Ask for the approved tool list, the classification scheme, the last audit and the last training completion rate. Ask which tools came off the list in the past year and why. Ask what happens in the first hour after someone reports a paste of source code into a personal account. A board that can answer those questions has oversight. One that cannot has a policy document.
None of this requires slowing anyone down. The firms that keep both the secret and the productivity gain make the safe route the fast route: a sanctioned tool that works better than the free one, a classification scheme that fits on one page, and an exception review that takes minutes rather than weeks.
The question in the title has an uncomfortable answer in most organisations. Employees are not being reckless. Nobody has drawn the line clearly enough for them to stay behind it.
Sources
- How to comply with UK trade secret legislation amid growing risk of unintentional disclosure using generative AI
- Is Your AI Tool Quietly Destroying Your Trade Secrets?
- Client Alert: Preventing, Detecting, and Litigating Trade Secret Theft in the Age of AI
- Protecting AI with IP: Comparing approaches taken in the US and UK
- The Biggest Trade Secret Leak in Your Company May Be Your AI Chatbot
Source: AI, trade secrets and NDAs: Is your organisation ready?, The Business Standard





