AI, trade secrets and confidentiality

Short answer: Feeding confidential business information into an external (cloud) AI model is legally risky. A trade secret is protected only as long as it is secret and you take reasonable steps to keep it so. Uncontrolled sharing with an external provider — which may reuse or store the input — can undermine that protection and may breach confidentiality or GDPR obligations.

Why input can undermine a trade secret

The protection of trade secrets (Directive (EU) 2016/943) depends on three conditions: the information is secret, has commercial value because it is secret, and is subject to reasonable steps to keep it secret. If that last pillar disappears — for example by placing information without safeguards into an external model that retains the input or uses it for training — the secret can lose its protected status. This is not only about actual leakage, but about the failure to take reasonable steps.

Contractual and confidentiality risks

Confidential information you receive from clients or partners is often covered by a non-disclosure agreement (NDA). Those NDAs typically restrict to whom and how you may disclose the information. An AI provider is a third party; entering information there can breach an NDA, even unintentionally. The same applies to source code, pricing models and unpublished strategy.

The GDPR layer for personal data

If the input contains personal data, the GDPR comes into play. You need a legal basis, a processor agreement with the provider, and clarity on where the data is processed and whether it leaves the EEA. Reuse of input for model training is a separate processing operation that is rarely covered by your original basis.

The difference between offerings

Not every offering is equal. Consumer versions of chat models may reuse input for training by default; business and enterprise variants often contractually guarantee that input is not used for training and not retained. Read those terms carefully: default settings differ from paid tiers, and a free trial environment is rarely covered by the same guarantees. Making this risk manageable begins with choosing an offering with the right contractual safeguards — see also RAG and enterprise AI governance.

Shadow AI as the biggest leak

The real risk often lies not in a deliberate choice but in shadow use: employees who open a free chat model without consultation and paste in a quote, contract or customer file. A single act can then undo an NDA, the GDPR and the secrecy measure all at once. Policy without a usable, approved alternative does not solve this — employees revert to whatever works. So offer a safe route alongside the prohibition.

What to do

• Classify information before input: public, internal, confidential or secret. Secret and NDA-bound information does not belong in an unmanaged model.
• Choose an enterprise offering with a contractual guarantee that input is not used for training and not retained.
• Conclude a processor agreement and check the processing location where personal data is involved.
• Set an input policy in your AI use policy for employees, with clear "do not enter" categories.
• Restrict and log access, so your reasonable steps are demonstrable.