Description:This session focuses on the early stages of working with data, when a dataset is large, complex, poorly documented, or unfamiliar. Participants will use AI to interpret data dictionaries, surface undocumented assumptions, and assess whether a dataset can answer their research question. They will also learn how to determine which AI tools they are permitted to use with a given dataset, since research data frequently carries terms of use that restrict AI tools. Participant exercises use a public-domain dataset and U-M's own AI services.- ICPSR policy on the use of large language models: https://www.icpsr.umich.edu/sites/icpsr/about/policies/large-language-models-and-ai
- ICPSR redistribution policy: https://www.icpsr.umich.edu/sites/icpsr/about/policies/redistribution
- Using AI with MIDUS, NSHAP, Add Health and other NACDA-hosted data: https://www.icpsr.umich.edu/sites/icpsr/news/using-ai-with-midus-nshap-add-health-and-other-nacda-hosted-data
- U-M guidance on AI and U-M data: https://safecomputing.umich.edu/protect-the-u/safely-use-sensitive-data/AI-and-UM-Data
- U-M generative AI services: https://genai.umich.edu/
- ACS PUMS documentation: https://www.census.gov/programs-surveys/acs/microdata/documentation.htmlPossible hands-on activities:- Determine which tier of AI tool you may use with a dataset you are currently working with, and where to check
- Translate a cryptic codebook into plain language, with the model flagging what it cannot determine instead of guessing
- Extract documentation into a structured format that records, field by field, what the model could not determine from the documentation alone
- Generate a fitness-for-purpose checklist against your own research question, and draft specific questions to send a data provider
Before the session:1. All attendees please bring a laptop.2. Sign in at genai.umich.edu to make sure your U-M GPT access is working. We'll use it for all hands-on exercises. 3. Come with a research question you're working on, and the name of a dataset you're using or considering for it. If the dataset has a data use agreement or terms of use page, have the link handy. Please don't plan to upload your own data during the session. We'll use a public dataset for the exercises, and part of the session covers how to check which AI tools you're permitted to use with your own data.