Administrative Staff Policy on Generative AI (GAI) Use
Administrative Staff Policy on Generative AI (GAI) Use
Generative Artificial Intelligence (GAI) is an emerging and imperfect technology that, when used thoughtfully, can offer valuable support to our work. FAS staff are encouraged to experiment and explore the responsible use of GAI as part of our commitment to supporting Harvard's research and educational missions. We also encourage robust discussions about domains of appropriate GAI use, and clarification of mutual expectations among colleagues.
Potential applications include, but are not limited to, leveraging GAI as a thought partner for idea development, refining and editing communications such as emails, memos, job descriptions, and formal letters, and enhancing productivity in administrative tasks.
While GAI tools can be useful and continue to improve, they aren’t perfect. Shortcomings include, for example, false or incomplete information and code with errors. Each FAS staff member bears full responsibility for the accuracy, appropriateness, and quality of their work product, including work product generated or assisted by GAI. Where possible and appropriate, we consider it advisable to acknowledge GAI use in work product.
To protect privacy and enhance cybersecurity, any Harvard-related use of Generative AI should occur exclusively through the HUIT-approved sandbox environment (HUIT AI sandbox) or an OpenAI Enterprise account linked to your Harvard Key. Using tools like AI-facilitated meeting transcription on Zoom should also only happen through Harvard-managed accounts. Any data entered into a GAI tool may not exceed the approved data classification level for that tool. GAI must be used in accordance with all applicable Harvard policies, including those governing cybersecurity and the Guidelines for Using ChatGPT and other Generative AI Tools at Harvard.
Frequently Asked Questions
Harvard-approved GAI tools, such as the AI Sandbox and ChatGPT Edu, are approved for use with up to level 3 (medium-risk) confidential data (learn more about security and privacy). This includes:
- Non-public financial statements
- University contracts
- Research administration records not otherwise classified
- Most Harvard source code
- Non-security related technical specifications
- Sensitive administrative survey data (e.g., performance reviews or course feedback)
- Course materials not otherwise classified
- FERPA-defined non-directory education record data not containing L4 data
Do NOT upload level 4 (high-risk) data, including:
- Social Security Numbers, passport numbers, driver's licenses
- Financial account information (bank accounts, credit/debit card numbers)
- Individually identifiable health or medical information
- System credentials (passwords, PINs, encryption keys)
- Security controls, system procedures and architectures
Data entered into Harvard-approved GAI tools will not be used to train external models.
While we don't currently have a formal GAI curriculum for staff, be on the lookout for events on campus that offer training and discussions on generative AI use. These events are announced through university communications channels and can provide valuable hands-on learning opportunities.
Currently, Harvard has specific restrictions on AI meeting assistants:
- Approved tool: Zoom AI Companion is the only approved AI meeting assistant, available through a limited HUIT pilot program for administrative use. Faculty and staff must be approved by their School/Unit leadership to participate.
- Not approved: Third-party AI transcription tools like Otter.ai, Google Meet transcription, Microsoft Teams transcription, or other non-Harvard contracted services should not be used for Harvard meetings.
- Alternative FAS-approved approach: If you don't have access to Zoom AI Companion, you can:
- Record your meeting in Zoom (with participants' consent)
- Download the standard Zoom transcript
- Upload the transcript to the HUIT AI Sandbox or ChatGPT Edu to generate a summary
This approach is suitable for meetings with content up to Level 3 (medium-risk) data, as both the HUIT AI Sandbox and ChatGPT Edu are approved for Level 3 data.
When using any AI tools for meeting content:
- Obtain consent: Always inform all meeting participants that you're recording the meeting and may use AI to help summarize it. Get their consent before beginning.
- Consider meeting content: Do not use AI tools for meetings discussing Level 4 (high-risk) data.
- Review output: Always review AI-generated summaries for accuracy before sharing them.
- Acknowledge AI use: Include a notice when sharing summaries [see below under [Acknowledging the use of GAI in my work”]
- Manage retention: Delete meeting recordings, transcripts and summaries when no longer needed, following Harvard's records retention guidelines.
To request access to the Zoom AI Companion pilot program, you need approval from your unit leadership and submit a request here. You must also Please be sure to visit the HUIT AI Meeting Assistant Guidelines for complete information.
Currently, there are no established standards or norms for acknowledging GAI use in administrative settings. These norms are still evolving. That said, we recommend following two guiding principles:
- Helpful Colleague Principle: If you would acknowledge the contributions of a colleague in producing work, consider acknowledging significant GAI contributions.
- Deception Principle: If someone might feel deceived upon learning that content was primarily AI-generated, acknowledgment is appropriate.
Examples:
- Minor edits (spelling/grammar checking) from generative AI models, similar to what traditional tools like Word or Grammarly provide, typically don't require acknowledgment
- AI-generated images should be acknowledged, based on the helpful colleague principle (e.g., "Image generated with AI using DALL-E")
- Meeting transcripts or summaries generated with AI tools, like Zoom AI Companion, should be acknowledged to maintain transparency about their source and potential limitations, and where to send inaccuracies (e.g., “Transcript automatically created by Zoom AI on [date]. Please note that AI transcription may contain errors."; "Meeting summary created with assistance from Zoom AI and edited by [your name]. Please submit any corrections to the meeting host.")
Remember that regardless of acknowledgment, you remain fully responsible for the accuracy, appropriateness, and quality of all work products, including those assisted by GAI.