Skip to content
PDF

Generative AI Resources Available to NIH Researchers

Generative AI is "a subset of AI that centers on creating new content" from user query -- The CLEAR path: A framework for enhancing information literacy through prompt engineering.

NIH scientists have at their disposal several AI tools free of charge. These include CHIRP and ChatGPT as well as Gemini at the enterprise level. Central to these tools is the chat box, where users can interact with them through natural language and they respond to user questions. For instance, when ChatGPT is asked to "Make a table that summarizes the top running speed (in miles per hour) for alligators, bulldogs, and turtles", it will search for the information and generate a summary table for the user. This contrasts with researchers having to perform key word searches, collating, and summarizing information on their own.

This article will highlight some basic features for each tool and how NIH researchers can start using them.

Note

Scientists will need to be on the NIH network either by being on campus or via VPN to use these generative AI tools. Authentication using NIH credentials is required for access.

CHIRP

Chatbox for the Intramural Research Program or CHIRP is a ChatGPT implementation for NIH intramural work. See https://irp.nih.gov/catalyst/33/2/chirp-a-chatgpt-model-for-the-nih-intramural-community to learn more. To start using CHIRP goto http://chirp.nih.gov/. CHIRP is optimized to assist scientists with NIH specific research. It enables researchers to work with protected sensitive data such as unpublished work and de-identified clinical data. Personally Identifying Information, however, cannot be used on CHIRP.

ChatGPT

ChatGPT is developed by Open AI. The enterprise version is available to NIH scientists. To start using this tool, goto https://go.hhs.gov/chatgpt. See https://tech.hhs.gov/protected/hhsai/chatgpt to learn more. It is important to point out ChatGPT should not be used for the following.

  • Sensitive Personally Identifiable Information (sPII)
  • Protected Health Information (PHI)
  • Classified information
  • Export Controlled Data
  • Confidential Commercial Information or Trade Secret Data

Codex

Codex is another Open AI product. While ChatGPT is a general purpose AI tool, Codex is dedicated to coding and software development and enables tasks such as versioning and pushing code to GitHub. In addition, users can automate tasks such as summarize Git commits for each project as well build custom skills https://developers.openai.com/codex/skills to expand Codex's capabilities. Codex can be used via command line, stand alone desktop application, or popular integrated development environments such as VS Code. See https://nih.sharepoint.com/sites/NCI-Fred/Knowledgebase/AI/SitePages/New-coding-feature-available-in-ChatGPT-Enterprise--Codex.aspx for more details and learn how to get started using Codex.

Google Gemini

Google Gemini is FedRAMP approved and can be used with many data types except for those that are classified. Data that can be used on Gemini include limited sensitive identifiable information, procurement sensitive data, and grant data. See https://tech.hhs.gov/protected/hhsai/gemini to learn about data restrictions for Gemini. To start using Gemini goto https://go.hhs.gov/gemini.

Claude

To start using Claude, visit https://claude.hhs.gov/. Claude is FedRAMP approved so it can be used for many data types except for classified information. See https://tech.hhs.gov/protected/hhsai/claude to learn about Claude enterprise.

Microsoft 365 Copilot

Microsoft 365 Copilot are generative AI tools incorporated into its office suite of tools include Word, PowerPoint, Excel, Outlook, and Teams. Available NIH-wide, Copilot will help streamline work and communication efforts by NIH scientists. These include drafting manuscripts, slide decks, emails, summarize conversations, and facilitate data analysis. To learn more, visit the Copilot section on the NIH AI Hub.

How Can Scientists Use Generative AI

Whether using CHIRP, ChatGPT, Gemini, or Microsoft Copilot NIH scientist can use generative AI to:

  • Upload and analyze documents or research articles.
  • Draft documents and teaching material.
  • Generate summary tables.
  • Analyze tabular data.
  • Help with code as well as generate Quarto Markdown or Jupyter Notebooks.
  • Generate workflow diagrams. As an example, ChatGPT can generate analysis pipeline flow charts by reading public GitHub repositories.

Limitations

Generative AI can make mistakes, so always fact check the results. Some user feed back for NIH generative AI tools are listed below.

  • Chats will become slower as they get longer.
  • CHIRP:
  • May show short term memory loss where replies from a couple of messages back were forgotten.
  • Times out after around 10 minutes, requiring re-authentication.
  • ChatGPT Thinking and Pro mode gives good answers, but these often take a long time to generate.
  • Citations and references:
  • ChatGPT Flagship and Instant modes do not provide references unless specifically asked. However, references are included at the end of the response.
  • Gemini defaults to providing references in-line with response.
  • ChatGPT Thinking and Pro modes gives in-line references with response as default.
  • AI tools can pull duplicate references. For instance, citing an article that was posted on PubMed and PubMed Central.

Education Opportunities

Classes

The NIH Library offers a range of AI classes including those relevant for prompt engineering. See https://www.nihlibrary.nih.gov. To learn about innovative ways in which AI is applied to biomedical research, checkout previous seminars from AI in Biomedical Research NIH Seminar Series on the BTEP Video Archive and check the BTEP calendar for future classes. Additionally, check with CIT (https://nih.sharepoint.com/sites/CIT-TechnologyTrainingProgram) for AI related trainings.

Interest Groups

NIH scientists can check out the following interest groups to learn more about generative AI practices and developments.

Useful References

NIH Resources:

External Resources: