What You Need to Know
- Commercial generative AI tools are not covered under negotiated agreements- only provide non-sensitive/public inputs to these tools. Never enter proprietary, export controlled, PII or other controlled data into these tools.
- Generative AI can create novel and unresolved intellectual property issues from "new" images derived closely from their original source, to potential claims that the third party tool has a role in the development of a technology the Lab is working on, to complaints regarding the training set of researcher-created models. Use caution, especially around anything related to the development of intellectual property.
Generative AI Tools and Data Protection
As the advancements in artificial intelligence continue to shape the landscape of scientific research and innovation, it is imperative that we establish clear boundaries and best practices for the responsible deployment of Generative AI tools within our information ecosystem. This document aims to provide Berkeley Lab employees with guidance on the utilization of Generative AI technologies while ensuring the security of sensitive information developed at Berkeley Lab.
Licensing Agreements
Berkeley Lab policy requires that all software licensing agreements with suppliers (and other entities), particularly those software suppliers that will host data generated or owned by Berkeley Lab, the UC or the Department of Energy as part of their services, require the software supplier to comply with a variety of Berkeley Lab data security and privacy practices. These include, but are not limited to, the protection of all Berkeley Lab data against loss, unauthorized access and use, and the indemnification of Berkeley Lab, the University of California and the Department of Energy, and compliance with all software-related requirements. The data security and privacy provisions contained in these agreements are a critical part of Berkeley Lab’s ongoing efforts to ensure the protection of the data and information related to its research, operations, and the personal information of its employees.
However, there is currently no agreement with OpenAI, the developer of ChatGPT, or its licensees that would provide these types of protections with regard to ChatGPT or OpenAI’s programming interface. Consequently, the use of ChatGPT at this time could expose the individual user and Berkeley Lab, the UC and the Department of Energy to the potential loss and/or abuse of highly sensitive data and information. Similarly, Google's Bard AI can be accessed with your LBL identity, but is subject to different protections than the core Google Workspace accounts. There are no agreements with popular image generation tools such as Midjourney, Dall-E (OpenAI), or online providers of Stable Diffusion.
Generative AI Tools use
Do not submit confidential or protected information such as Personally Identifiable Information, Proprietary Information and Export Controlled Information on Generative AI tools like Open AI’s ChatGPT or Google’s Bard. In general, any data that is determined to be protected or controlled should not be used to generate a response using Generative AI tools. You can refer to the Lab’s Controlled and Prohibited Information Categories policy to learn more about each information type.
Please be advised that in the absence of a Berkeley Lab agreement covering use of ChatGPT, your use of ChatGPT constitutes personal use, and obligates you, as an individual, to assume responsibility for compliance with the terms and conditions set forth in OpenAI’s own Terms of Use.
For further guidance on using Google Bard, please review the references listed below that address Google’s Generative AI Terms of Service. This includes Google’s Prohibited Use Policy as well as Privacy Policy and Use Restrictions.
Please note that OpenAI states that how it uses or does not use your information is different when using their API services from the consumer ChatGPT interface. For further guidance on using ChatGPT please review the references listed below that address OpenAI’s own recommendations and policies on research and streaming of ChatGPT, and privacy, among others.
Intellectual Property Concerns
Intellectual Property Infringement: Generative AI tools are capable of producing content that closely resembles existing copyrighted works. Researchers need to ensure that their use of such tools does not infringe upon the intellectual property rights of others. This includes avoiding the creation of derivative works without proper authorization or licensing.
Dataset Licensing: AI models often require large datasets for training, and these datasets may contain copyrighted material. Researchers must ensure that they have the necessary rights or permissions to use the data, especially if it includes protected works such as images, text, or audio. DO NOT DISTRIBUTE A MODEL YOU DEVELOPED WITHOUT CONSULTING WITH IPO.
Authorship and Attribution: Researchers who develop an original research model are generally credited as the authors or inventors of the model. However, if the generative AI model significantly enhances the research model, a question of authorship, ownership and attribution for the improved version may arise. Authorship and citation remains a complex and evolving issue for these tools. University and Lab policy require the highest ethical standards in authorship and citation. Consider an expansive approach to citation of these tools as new norms within scientific communities take form.
Prohibited Use
Please also note that OpenAI explicitly forbids the use of ChatGPT and their other products for certain categories of activity. This list of items can be found in their usage policy document.
Further guidance on Google Bard, ChatGPT and other Generative AI tools will be forthcoming as soon as it is available.
Subscriptions
The Lab allows individual Lab-paid "click through" subscriptions to some of these tools for non-sensitive, exploratory work. If you have a need for an API and/or interactive subscription, email [email protected] and explain your use case and desired tool subscription.
References
https://policies.google.com/terms/generative-ai
https://openai.com/policies/sharing-publication-policy
https://openai.com/policies/usage-policies
https://openai.com/policies/privacy-policy
https://openai.com/policies
Additional Readings
https://www.quantamagazine.org/the-unpredictable-abilities-emerging-from-large-ai-models-20230316/
https://www.theatlantic.com/newsletters/archive/2023/03/dont-be-misled-by-gpt-4s-gift-of-gab/673411/
https://www.ucop.edu/ethics-compliance-audit-services/compliance/uc-ai-working-group-final-report.pdf
https://adminrecords.ucsd.edu/Notices/2023/2023-3-23-2.html
https://ethics.berkeley.edu/privacy/appropriate-use-chatgpt-and-similar-ai-tools
