Protecting trade secrets when using AI
Trade secret protection depends on reasonable steps to keep the information secret. Feeding confidential material into an AI system without safeguards can remove that condition, and with it the protection.
A trade secret is not a right you acquire and then hold. It exists only while the conditions of section 2 no. 1 of the German Trade Secrets Act are met, and one of them is in the hands of the business itself, namely the reasonable steps taken to keep the information secret. Since 2019 the mere intention to keep something confidential is not enough. This is exactly where the use of AI bites.
Where the line runs
Preserves protection
- A binding list of data categories that must not be entered
- Role-based access on a need-to-know basis
- Contractual limits on the provider, with training on inputs excluded or opted out
- Marking of sensitive documents and training of staff
- Locally hosted or tenant-separated systems where confidentiality matters
Puts protection at risk
- Free consumer accounts whose terms allow inputs to be reused
- No confidentiality agreement with annotation, cloud or development partners
- Inputs without any rule at all, because everyone does it
- A policy in a drawer with no technical measure to match it
- No record of the measures taken, although the holder has to prove them in court
What to do
- Identify sensitive data categories and make the list binding
- Check the terms of the tool in use for how inputs are handled
- Conclude confidentiality agreements with staff and third parties and define the protected subject matter precisely
- Control access by role, encrypt data in transit and at rest
- Write the rules into the AI policy and train on them so that they reach day-to-day work
- Record the measures taken, because the holder has to prove them in a dispute
- Set out a procedure for the case where a secret has leaked
The other side, your own system
Anyone developing a model faces the reverse question. Copyright and patent law often do not reach algorithms, model parameters and datasets, so trade secret protection fills the gap. Offering the system as software as a service keeps technical control in house and limits what can be worked out by observing and testing. Excluding reverse engineering by contract is permissible in that setting. The copyright right to decompile is mandatory and cannot be excluded.
Where the system generates new information, that too can be a trade secret. How it came about is irrelevant. Attribution becomes difficult where provider, user and data supplier all contribute to the same output, which is why it belongs in the contract before the first output exists.
Confidential material in an AI tool?
We review your provider's terms and draft the rules that keep the protection alive.
Get in touchFrequently asked questions
What makes information a trade secret?
Section 2 no. 1 of the German Trade Secrets Act requires three things at once. The information must not be generally known or readily accessible to the relevant circles, and must have commercial value precisely for that reason. It must be the subject of reasonable steps to keep it secret. And there must be a legitimate interest in secrecy. If one of the three falls away, so does the protection.
Does a secret lose protection if we enter it into an AI tool?
Not automatically, but the risk is substantial. Entering confidential material into a cloud system makes it available to the provider. If the provider uses inputs for further training, they can end up in the model parameters and reach third parties through the output. What matters is what the contract allows and which safeguards are in place internally.
Which safeguards count as reasonable?
The standard is a sliding one and depends on the size of the business and the value and sensitivity of the information. The minimum is a confidentiality agreement in every relationship in which confidential material is disclosed, access on a need-to-know basis, and basic technical protection. Where the need is higher, encryption, multi-factor authentication and logging are added.
Does a confidentiality agreement need a contractual penalty?
That is disputed. On the better view a penalty is not required to preserve the character of the secret. What counts is the overall picture of all safeguards. Against providers with strong bargaining power, where a penalty cannot be negotiated, other effective measures can take its place.
Can our own AI model be protected as a trade secret?
Often this is the most practical route, because copyright and patent law frequently do not reach algorithms, model weights and datasets. The conditions of section 2 no. 1 still apply. Offering the system as software as a service keeps technical control with the provider, and reverse engineering can then be excluded by contract. The copyright right to decompile cannot be excluded.