Mexican Health and Aging Study (MHAS / ENASEM) Last updated: October 2026
Can I Use Large Language Models and Other AI Tools (such as ChatGPT, Google Gemini, Claude, Microsoft Copilot, etc.) with MHAS Data?
General Statement: Currently, large language models (LLMs) and other artificial intelligence (AI) tools may not be used to manage, process, or analyze data distributed by the Mexican Health and Aging Study (MHAS), including public-use files, imputed files, constructed variables, ancillary study data (e.g., Mex-Cog), Gateway Harmonized MHAS files, and any restricted or sensitive data products.
Under the MHAS data-use terms that all registered users accept, researchers are forbidden to transfer MHAS public-release data to a third party (except staff or students for whom the registered user is directly responsible). The same restriction applies, with even greater force, to restricted and sensitive data under the applicable Data Use Agreements. Providing MHAS microdata to an LLM or external AI service constitutes redistribution of the data and is therefore a violation of all existing MHAS data-use terms and agreements.
Important: This policy is consistent with NIH guidance on artificial intelligence in research. Unauthorized disclosure of research data (including when data are sent to an external AI provider) violates multiple NIH and institutional policies that protect participant privacy and confidentiality.
Classification of LLMs for Purposes of This Policy
Type 1: LLMs that retain user-provided data for any purpose, including training the model (e.g., publicly available versions of GPT, Llama, Gemini, Claude, etc.).
Type 2: LLMs that are licensed by an institution and whose terms of use prohibit the retention or use of user-provided data for training (e.g., enterprise or educational versions with data-protection guarantees).
Type 3: Type 2 LLMs that are isolated within a secure network or air-gapped environment with no access to the Internet.
We thank the Health and Retirement Study (HRS) at the University of Michigan, ICPSR, and Sebastian Karcher at Syracuse University for originating this taxonomy of LLMs.
Data Use by LLM Type
LLM Type
MHAS Data Use
Reason
Type 1
None
Type 1 LLMs ingest and may retain or reuse the data. This counts as redistributing the data to the company operating the LLM and is not permitted.
Type 2
None
Although Type 2 LLMs may not retain the data for training, they are not isolated from broader networks or the Internet. Their use does not comply with data-security expectations for restricted data and is currently not accepted for public-use data either.
Type 3
None (case-by-case review possible)
Type 3 systems are isolated, yet current evidence shows that complete isolation cannot yet be reliably guaranteed (risk of “sandbox escapes” or unintended access). At present MHAS does not routinely approve Type 3 use with microdata.
Study-Level Metadata and Documentation
It is permissible to use LLMs and other AI tools with MHAS public-facing documentation, codebooks, questionnaires, study-level metadata, and aggregated group or population estimates (tables or statistics that do not contain individual-level records). Person-level or household-level microdata must not be provided to or processed by any LLM or external AI service.
This restriction does not apply to other machine-learning methods (including generative machine-learning methods that run entirely under the researcher’s control and do not transmit microdata to external services), provided their use fully complies with applicable data-protection and data-use requirements.
AI and LLM Use Policy FAQs
Q: Is it acceptable to use an LLM to write analysis code (e.g., Stata, R, Python, or SAS scripts) based on MHAS data?
A: Yes, provided that no MHAS microdata (individual- or household-level records) are uploaded to or shared with the LLM. You may describe variables, ask for code templates, or paste non-identifying codebook excerpts.
Q: Is it acceptable to share derived analytic output (e.g., model coefficients, summary statistics, or regression tables) with an LLM?
A: Yes, as long as the output does not contain microdata and cannot be used to re-identify participants. Ordinary aggregate results are fine.
Q: Does the policy apply to all MHAS data products, including Gateway Harmonized files, imputed data, Mex-Cog, and restricted/sensitive data?
A: Yes. The policy applies to every data product distributed by or derived from MHAS, regardless of the distribution channel.
Q: Is it acceptable to use AI-assisted development tools that automatically index files in the working directory (e.g., AI features built into code editors or IDEs) when MHAS data files are present?
A: No. Users must disable any feature that indexes or uploads local files, or store MHAS microdata outside any directory to which such a tool has access.
Q: What if my institution has a fully isolated, air-gapped AI environment?
A: Please contact the MHAS team in advance. Limited Type 3 use may be considered on a case-by-case basis after review of the security environment and data-protection plan.
Questions or Requests for Clarification
If you have questions about this policy or wish to discuss a specific proposed use of AI tools, please contact the MHAS team: