Data Privacy and AI Tools

Being good caretakers of your data

Dr. D

2026-11-02

Today

Morals aside. Whether AI should be used in qualitative research is a week 13 conversation. Today is narrower: which data can go into which tool?

By the end of today, you’ll be able to:

  • Sort data into Chico State’s Level 1, 2 and 3 privacy classifications
  • Distinguish between a chatbot and an, agent, and a campus-managed tool vs one that trains on your data

Hook

An invented excerpt

“I’m the only night-shift pharmacist in a town of about 4,000 in Trinity County. Last winter when the bridge closed, I walked two miles to work, and my boss still wrote me up for being late.”

Participant 07

Pair up (1 min)

Would you paste this into a chatbot? It is de-identified.

Names removed is not the same as de-identified

  • Town size, job, county and one event can point to one person
  • A pseudonym doesn’t help when the story still identifies them
  • Chico’s guidance allows fully de-identified data in a prompt. For rich text, “fully” is hard to reach.

Pair up (1 min)

You’ll read a few examples of real stories next week

Data levels

What is a data level?

Chico State sorts university data into three levels. The level depends on what happens if the data is exposed, not on how sensitive it feels.

  • Level 1: Confidential
  • Level 2: Internal Use
  • Level 3: General

You don’t pick the level

You apply the criteria. Chico’s Standard says it covers research data too.

Level 1: Confidential

  • A law requires protection, or exposure would do severe harm.

  • Not disclosable under the California Public Records Act or other law; severe risk if exposed; need-to-know use only.

Examples:

  • Social Security number with a name
  • Medical records and psychological counseling records
  • Biometric information
  • Passwords and credentials

Where it can live

Only the Box Level 1 folder. Not your laptop, not OneDrive or SharePoint, not a personal account.

Level 2: Internal Use

  • No protection law, but exposure would harm a person or break a promise.

  • Needs protection for ethical, contractual or privacy reasons; moderate risk.

Examples:

  • Grades and other student records that are not directory information
  • Race, ethnicity and gender
  • Name with a birth date
  • Student photo

Level 3: General

  • Meant to be public.
  • No special protection required.

Examples:

  • Campus directory information
  • Public work email and job title
  • A name on its own

Combinations count

Data Level
A name on its own Level 3
A name plus a birth date Level 2
A name plus a Social Security number Level 1

Same idea as the opening excerpt: pieces that look harmless can identify someone together.

Guidance on how to choose

Ask from the top down:

  1. Does a law require protection, or would exposure cause severe harm? –> Level 1
  2. Would exposure harm a person or break a promise, even if no law applies? –> Level 2
  3. Is it public by design? –> Level 3
  4. Not sure? –> Err by 1 on the side of more protections (e.g. 2 instead of 3)

Where do our qualitative data land?

Data Our reading
Identifiable interview or focus-group transcripts At least Level 2
Counseling or medical content inside a transcript Could be Level 1
Audio recording of a participant A voice may count as biometric, so Level 1
Public posts or published reports About Level 3 (consent and IRB rules may still apply)
Invented (synthetic) transcripts No level, no real person

Identify the level (show of fingers)

Item Level
Public work email and job title Level 3
A student’s grades Level 2
A Social Security number with a name Level 1
A name with a birth date Level 2
A participant’s race and ethnicity tied to their survey Level 2
Notes from a counseling session, client named Level 1
An identifiable interview transcript Level 2
Campus directory information Level 3

Which tool?

Chatbot vs. agent

Chatbot

  • You type, it replies with text
  • You upload or copy data to the text box - thus deciding what leaves your machine

Agent

  • Claude Code, Codex, the AI assistant in RStudio
  • Reads your files and runs commands
  • Edits and deletes files on your computer
  • It decides what to read

What can go wrong with an agent

  • It can read every file in its folder: interview audio, transcripts, your codebook
  • Anything it reads can be sent to the AI company
  • AI features built into transcription or note-taking apps can also send your content to a vendor
  • Does not always respect your rules or boundaries

Ground rule

Keep sensitive data out of the folder an agent works in. Don’t let it run commands outside of the authorized folder

Two different questions

  1. Does the vendor train on or keep my data?
  2. Is this tool cleared for my data’s level?

Campus-managed is not the same as cleared

Campus-managed AI tools help keep data out of the vendors hand, but are still not allowed for all data sources.

Example: Campus Managed AI Tools

  • Google Gemini (Chat and Notebook)

  • Chat GPT EDU (Chat and Codex)

  • Claude (in testing)

  • Microsoft Copilot

  • Zoom AI Companion

  • Adobe Firefly

  • Conversations are private and not used for model training

  • Not approved for Level 1 or Level 2 data

What may go into a prompt - recap

May include

  • Public information
  • Synthetic data not derived from real people
  • Fully de-identified data
  • Generic examples
  • Non-sensitive code

Must not include

  • Level 1 or Level 2 data
  • FERPA-protected or health information
  • Student ID numbers, grades, academic transcripts, advising records
  • Passwords, keys and tokens

Source: Chico’s coding assistants guide. It is tool-neutral and covers research and scholarship.

Kahoot - Quiz time

  • Open note
  • Open Neighbor
  • Not open Google/AI

Wrap up

Rule of thumb

Level 3, synthetic or fully de-identified data only. Anything else: ask first (your instructor, the IRB or ITSS).

  • If you aren’t sure it’s safe, assume it isn’t.
  • Your consent form and IRB protocol can be stricter than the campus levels.
  • Chico’s GenAI Guidelines say they do not govern research, so don’t treat the campus tools as pre-cleared for it.

Next

Readings on using AI and LLMs to assist qualitative analysis. Read a few over the next two weeks. Week 13: share-out and discussion.

Sources

Tool statuses as of November 2026. Check before relying on them.