Sensitive Data Handling in AI

Start here: sort the data, not the tool

UA classifies all data as Restricted, Internal Use, or Public. Under UA’s Generative AI Security Standard, Restricted and Internal Use data must not go into GenAI tools that have not been vetted for it. Public data is fine. When you are unsure, default to the higher one.

Class Typical staff examples AI rule of thumb
Restricted

Personnel and applicant files, SSNs, payroll and direct-deposit data, student records you handle at work, health and accommodation information, payment card and banking data, passwords

Do not enter it.

Not in a personal account, not in a UA-licensed tool, unless OIT has confirmed that tool is approved for this class.

Internal Use

Draft budgets, procurement and vendor records, contracts under negotiation, search committee materials, non-public meeting minutes, UA business email

Not in unapproved tools.

Use UA-provided tools and ask OIT if the material is sensitive or contract-covered.

Public

Published policies and forms, catalog and website copy, directory information, job descriptions, press statements, training materials

Go ahead.

This is the large majority of everyday AI use — drafting, summarizing, cleaning up language.

 

Five habits

      1. Sort before you type. One question before pasting: could this identify a student, an employee, an applicant, or a patient, or is it non-public UA business? If yes, describe the problem in general terms instead of pasting the record.

      2. Sign in with your UA account. UA provides Google Gemini, Microsoft Copilot, and Zoom AI Companion to all UA account holders, with vendor data-protection commitments a personal account does not carry. Same tool name, very different contract.

      3. Treat verification requests as the high-risk part of your job. Password resets, MFA re-enrollment, direct-deposit changes, and vendor banking updates are now requested by cloned voices and AI-written pretexts that get the details right. Confirm through a channel you initiated, and never let urgency or seniority shorten the process.

      4. Keep a human in the loop, and say so. The standard requires human review of AI output and attribution of the AI systems you used. Check numbers, names, dates, and policy citations before anything leaves your desk.

      5. Ask before you connect. New AI software goes through UA procurement, and unmanaged AI applications require a Generative AI Risk Awareness and Acknowledgement Form — including browser extensions and “assistants” that can read whatever is on your screen.

Two newer risks worth knowing

  • Prompt injection. AI tools that browse the web, read your email, or open attachments cannot reliably tell your instructions apart from text they encounter, so a malicious PDF or web page can instruct the tool on your behalf. Give them the narrowest access that works, and keep approval on anything that sends, deletes, or spends.

  • AI notetakers create records you did not intend to create. A meeting assistant turns a conversation into a searchable transcript that outlives the meeting and may be subject to retention schedules and public-records requests. Ask before one joins, keep them out of personnel, grievance, conduct, and health-related meetings, and use only the UA-provided option.

If you think something went out that should not have: report it to OIT Information Security immediately. Early reporting is what keeps a mistake small.