How generative AI is impacting trade secret protection

The advent of generative artificial intelligence tools (AI), such as ChatGPT or Microsoft Copilot, has raised new questions about how to best protect intellectual property. As companies consider their intellectual property strategy in light of generative AI, companies should consider its effects on company trade secrets. In particular, companies should consider (1) whether information that is generated by an AI can be protected as a trade secret; and (2) whether inputs sent to an AI can be protected as a trade secret.

SAFEGUARDING AI-GENERATED CONTENT AS A TRADE SECRET

Though the specifics differ in some ways by jurisdiction, to protect information as a trade secret, the information must have value from not being generally known and a company must take reasonable steps to preserve the secrecy of the information. Trade secret law does not include the same sort of author or inventorship requirements as patent or copyright law. Thus, content generated solely by AI, without any human inventor or author, can likely be protected as a trade secret even if it is not protectable by other forms of intellectual property.

However, two practical barriers may complicate trade secret protection. First, generative AIs typically work by predicting the most likely output to a prompt based on the data used to train the AI. If the AI was trained on public data, others may argue the output generated by the AI is "generally known" because it was generated based on the most likely public data. Companies can mitigate this risk by arguing that the data generated by the AI is a trade secret in certain contexts or in certain uses. For example, a company could argue that software code generated by an AI is a trade secret when it is used to solve a particular engineering problem. Alternatively, some generative AIs can be trained using a company's private internal data and then generate content using only that data. Information generated by such an AI is less susceptible to the same arguments regarding whether the information is generally known.

Second, some generative AIs store their own inputs and/or outputs and then are trained again on that data. If such an AI was used to generate information that a company wished to protect as a trade secret, others may argue that company did not take reasonable steps to protect the secrecy of the information because it is now available to others through the AI. However, generative AI's can also be configured to delete or otherwise keep private the output that the AI generated. If a company is considering using AI to generate information it wishes to preserve as a trade secret, it should carefully consider the terms of service or other contractual provisions related to how the AI will treat its inputs and outputs. Because non-disclosure agreements are typically found to be a reasonable step to preserve secrecy, similar contractual provisions for the AI will likely be sufficient to preserve trade secret status.

USING TRADE SECRETS AS INPUTS FOR AI-GENERATED CONTENT

When using AI tools, companies should carefully consider whether company data that is used to train or prompt an AI can still be considered a trade secret. Again, this analysis will likely revolve around whether a company is taking reasonable measures to maintain the secrecy of any data that is provided to an AI. In particular, companies should consider whether data they provide to an AI is used to further train the AI.

As an example, the current terms of service for ChatGPT individual users allows ChatGPT to be further trained on the inputs users provide to ChatGPT. The AI could then potentially output this information to others. Thus, if a user were to provide a trade secret to ChatGPT as an input and ChatGPT absorbed that data to train a future AI, the information may no longer be a trade secret because the information may be generally known or no longer subject to reasonable efforts to maintain its secrecy. In contrast, the current terms of service for ChatGPT business use specify that users retain all ownership rights in their inputs and that ChatGPT does not use the inputs to train future AIs. Providing a trade secret input to an AI under these terms may allow the company to claim it was still taking reasonable steps to maintain the secrecy of the information and that the information did not become generally known. Thus, companies should carefully consider the specific contractual terms for using any AI before providing the AI with any trade secret information.

Much has been written on generative AIs and some courts have addressed generative AIs in various contexts, including patents and copyrights. But few courts have addressed generative AIs in the trade secret context. The law is sure to develop over time, but companies that use generative AI tools should look to trade secrets as an avenue to protect valuable information that may not be protectable under other forms of intellectual property.

Daniel P. Hughes is a partner in the San Diego office of Knobbe Martens. His practice focuses on intellectual property litigation, particularly in the software and medical device space.

Adam B. Powell is a partner in the San Diego office of Knobbe Martens. He is the co-chair of the firm's antitrust practice group and focuses on intellectual property and antitrust litigation, particularly in the medical device and electronics space.