Understanding what counts as sensitive and what protections apply.

HIPAA-protected health information (PHI):

• Medical records, diagnoses, medications, test results
• Demographics linked to health data (date of birth + gender + zip code = reidentifiable)
• Genetic information
• Psychotherapy notes
• Protection: Cannot be entered into commercial AI tools without explicit institutional data use agreement
• Applies to: Research with human subjects involving health data, US researchers
• Enforcement: HHS Office for Civil Rights; violations: $100-50,000 per violation

GDPR-protected personal data (European Union):

• Name, email, ID number, IP address, cookies
• Genetic data, biometric data, health data
• Any data allowing identification of a person
• Protection: Must have legal basis for processing (consent, contract, etc.); data must be stored in EU/with adequate safeguards; people have right to access/deletion
• Applies to: Any research involving EU residents' data, even if researcher is not in EU
• Enforcement: National data protection authorities; violations: up to 4% of global revenue

Identifiable research data:

• Data allowing person identification: names, student IDs, medical record numbers, dates of birth combined with demographics
• Protection: Should be de-identified if possible; if identifiable, keep secure and limit access
• Applies to: Any research involving human subjects
• Enforcement: IRB; violations: loss of research approval, unable to publish, institutional sanctions

Proprietary/unpublished data:

• Methods not yet published that give your research group advantage
• Preliminary data that could affect your career if disclosed
• Collaborative data where others have agreed confidentiality
• Protection: Don't share without consent; some tools may use data for AI training
• Applies to: Competitive research, industry partnerships
• Enforcement: Institutional policies, partnership agreements

Vulnerable population data:

• Any data from children, prisoners, people with cognitive impairment
• Extra protections required; sensitive even if de-identified
• Protection: Extra caution; limited sharing; secure storage
• Applies to: Research with vulnerable populations
• Enforcement: IRB; funding agencies

Understanding what happens to your data when you use public AI tools.

Data retention and use:

• Most tools initially had policies where user data improved models (trained on inputs)
• This is changing: many now let you opt out of data retention
• Your inputs may be monitored for safety (filtering toxic content)
• Some tools share data with subcontractors
• Enterprise versions typically have stricter privacy terms

Where to check:

• Tool's privacy policy
• Terms of service
• Data use documentation
• Tool settings (usually an option to opt out of data retention)

Red flags:

• Policy doesn't clearly state data retention
• Tool reserves right to use your data for training
• No option to opt out
• Subcontractors not listed
• No security certifications (ISO 27001, SOC 2)

Examples:

• ChatGPT free: Data is retained and used for improvement; you can opt out in settings
• Claude (free): Data is retained; Anthropic has stated they don't use it for training; check current terms
• Claude (paid/enterprise): Stricter privacy terms; good for sensitive work if institutional agreement
• Grammarly: Varies by plan; free version keeps data; paid versions have privacy options
• GitHub Copilot: Don't use with sensitive code; data may be used for improvement

Rules from your institution about where data can go.

IRB (Institutional Review Board):

• Reviews research to ensure human subjects are protected
• May require: specific data security, limited access, retention policies
• Can restrict: where data is stored, who can access it, tools that can be used
• Process: Describe your data security plan in IRB protocol
• Can say "no" to certain tools if they don't meet security standards

Data governance offices/committees:

• Many institutions have data governance committees reviewing how data is used
• May require: data use agreements, security assessments, audit trails
• Approval: Required before using certain tools with institutional data

Funding agency requirements:

• NIH, NSF, etc. have data management plans
• Often specify: data must be stored securely, restricted access, compliant with laws
• May explicitly address: use of commercial tools, cloud storage, AI analysis

HIPAA compliance officers:

• If research involves health data, institution has compliance officer
• Can advise: what tools are HIPAA-compliant, what agreements are needed
• Can authorize: Business Associate Agreements (BAAs) that allow tool use

How to evaluate whether a tool is safe for your data.

Security certifications:

• ISO 27001: Information security management
• SOC 2 Type II: Security controls and auditing
• FedRAMP: US federal government security standard
• HIPAA-compliant (if using health data)
• Look for these in tool's documentation

Data location:

• Where is data stored? (US, EU, other)
• Can you choose storage location?
• Does it meet your regulatory requirements?

Encryption:

• In transit (TLS/HTTPS): Standard for most tools
• At rest: Does the tool encrypt data on their servers?
• End-to-end: Does only you have decryption keys? (Rare for commercial tools)

Access controls:

• Who can access your data?
• Can you audit who accessed what?
• Are there API keys/passwords managed securely?

Liability:

• Does the tool have insurance for data breaches?
• Do they have data breach response protocols?
• What's their liability if data is compromised?

How to structure research using AI while protecting data.

Separation of sensitive and non-sensitive:

• Keep sensitive data off commercial tools entirely
• Use de-identified data for AI analysis if possible
• Do sensitive work locally or on secure institutional systems
• Do non-sensitive work (literature search, general analysis) with commercial tools

De-identification:

• Remove identifiers: names, IDs, dates of birth, medical record numbers
• Remove quasi-identifiers: demographics that could allow re-identification
• Assess: Can someone re-identify? If yes, you didn't de-identify enough
• Trade-off: De-identified data loses richness but gains privacy

Local analysis:

• For sensitive data, consider local tools running on your computer
• Open-source models (LLaMA, Mistral) can run locally
• No data leaves your computer; complete privacy
• Trade-off: Requires technical skill and computational resources

Institutional tools:

• Use institution-approved tools and secure storage
• Example: Institutional enterprise agreement with AI tool provider
• Benefit: Negotiated privacy terms, compliance with regulations
• Limitation: May be slower, less feature-rich than commercial options

Data agreements:

• Business Associate Agreement (BAA) for HIPAA compliance
• Data Use Agreement (DUA) defining data sharing terms
• Confidentiality Agreement between collaborators
• Consulting agreements with external parties
• These allow some tools for use while protecting institutional and participant interests

—

Practical Research Use Cases

Scenario: You're researching depression treatment outcomes in patients. You have data with diagnoses, medications, demographics. You want to use AI for literature synthesis and data analysis.

Security assessment:

• Step 1: Patient data is PHI (protected health information)
• Step 2: Cannot put PHI into ChatGPT, regular Claude, or public tools
• Step 3: Options:

a) De-identify data (remove/generalize dates, identifiers)

b) Use institutional enterprise agreement with AI tool provider

c) Use local AI tools with data on secure institutional servers

d) Use commercial tools only for non-sensitive work (literature, general analysis guidance)

Secure workflow:

• Use Semantic Scholar for literature (no sensitive data needed)
• Use Claude to help synthesize findings (no patient data needed)
• De-identify patient data: remove identifiers, generalize dates
• Use de-identified data for exploratory analysis
• If needing AI help with analysis code: describe analysis goals in general terms (don't name your study); ask AI for code template; modify for de-identified data
• Final analysis uses AI-informed methods but runs on de-identified local data
• Report results without sensitive details

Scenario: Your lab developed a novel analytical approach that's not yet published. You want AI help developing code to implement it, but you're concerned about intellectual property.

Security assessment:

• Step 1: Unpublished methods are proprietary
• Step 2: Pasting into ChatGPT might contribute to training data (depends on settings)
• Step 3: Sharing with competitors would undermine advantage

Secure workflow:

• Describe the general analytical approach without proprietary details
• "I have a method that calculates X using Y. How would I implement this in Python?"
• AI can help with implementation without learning your specific method
• Test with local AI tools if very sensitive
• Never paste actual proprietary code or detailed methods into public tools
• Review AI's code and suggestions before accepting
• Document: You directed the implementation, AI assisted

Scenario: You're collaborating with another institution that provides data under strict confidentiality agreements. You want to use AI to help with analysis.

Security assessment:

• Step 1: Check the data agreement
• Step 2: Determine: Can data be used with external AI tools? Almost certainly no
• Step 3: Options:

a) Process data locally only

b) Get written permission from partner institution

c) Use only de-identified summaries of their data

Secure workflow:

• Keep partner data local; never put it in cloud tools
• For AI assistance, describe analysis goals in abstract terms without specific data
• "I have a time-series dataset with intervention effect. How would I test for interaction effects?" (generic, not specific to partner data)
• Implement generic code guidance on your specific data locally
• Report results to partner; maintain confidentiality

Scenario: You're a student writing a dissertation analyzing sensitive data about vulnerable populations.

Security assessment:

• Step 1: Student data is institutional data
• Step 2: IRB approval restricts data use
• Step 3: Likely restriction: data cannot be shared with third parties (including AI tools)

Secure workflow:

• Ask advisor: "Can I use AI tools with my data? What does my IRB approval allow?"
• Likely answer: "No, keep data local. But you can use AI for other aspects: literature, writing improvement, code templates."
• Workflow:
• Use AI for literature (no data needed)
• Describe analysis approach abstractly to get code help
• Use code as template, modify for your actual data analysis
• Use AI for writing improvement (no sensitive data)
• Keep actual data analysis local

—

Hands-On Exercise

Objective: Understand privacy requirements for your work and design a secure workflow.

Steps (Time: 60-90 minutes):

1. Identify sensitive data (20 minutes):

• What data does your research use?
• Is any of it sensitive? (Health, personally identifiable, proprietary, vulnerable populations)
• Do any regulations apply? (HIPAA, GDPR, IRB, funding agency requirements)
• Document: Types of sensitive data you have

1. Check institutional constraints (20 minutes):

• Contact IRB: "What does my approval allow regarding data security and use of external tools?"
• Contact HIPAA compliance officer (if applicable): "Can I use commercial AI tools with my data? What agreements are needed?"
• Check funding agency requirements: "What data security standards must I meet?"
• Document: What constraints apply to your research

1. Assess AI tools you want to use (20 minutes):

• For each tool, find: Privacy policy, data retention statement, security certifications
• Assess: Can this tool be used with your sensitive data?
• For each tool:
• Security certifications? (ISO 27001, SOC 2, HIPAA-compliant?)
• Data retention policy?
• Option to opt-out of data retention?
• Enterprise agreement available?
• Document: Assessment of each tool

1. Design secure workflow (20 minutes):

• Given your data and constraints, how would you safely use AI?
• Create a workflow flowchart:
• Step 1: This task (literature search) uses AI Tool A (no sensitive data)
• Step 2: This task (de-identification) happens locally
• Step 3: This task (analysis guidance) uses AI Tool B with de-identified data
• Step 4: This task (writing) uses AI Tool C (no sensitive data)
• Document: Your secure workflow for your research

1. Identify gaps (10 minutes):

• Are there tasks you want to do with AI but can't due to data sensitivity?
• How could you solve these? (De-identification? Local tools? Institutional agreements?)
• Document: How you'd handle each gap

Time required: 60-90 minutes

—

Common Mistakes and Misconceptions

You don't get to decide what's sensitive. Regulations and institutions decide. Even seemingly non-sensitive data combined with demographics can be reidentifiable. When in doubt, treat it as sensitive.

Privacy policies are marketing documents. Check for security certifications and actual data practices, not just marketing language. "Encrypted" can mean many things; "end-to-end encrypted" is better than server-side encryption.

De-identification must be actual de-identification (removing all identifiers and quasi-identifiers), not just removing names. Improper de-identification still violates HIPAA. If unsure, have your HIPAA compliance officer review your de-identification process.

Enterprise agreements with BAAs cost money but are required if you have sensitive data. Using free tools with HIPAA data violates HIPAA, regardless of how much work you do. There's no alternative to doing this right.

Compliance is minimum bar, not best practice. Even if technically allowed, consider whether you're protecting participants' privacy in spirit, not just letter. Additional security beyond requirements is often warranted.

—

Key Takeaways

• Sensitive data includes HIPAA-protected health information, GDPR personal data, identifiable research data, proprietary data, and vulnerable population data with different regulatory and ethical protections
• Commercial AI tools vary in privacy practices: Most retain data for improvement; check privacy policies for opt-out options and data retention practices; enterprise versions typically have better privacy terms
• Institutional constraints from IRB, HIPAA officers, and funding agencies restrict where sensitive data can go and must be respected
• Secure workflows separate sensitive and non-sensitive tasks: Use AI freely for non-sensitive work; use de-identified data or local tools for sensitive data; avoid putting sensitive data into commercial tools
• De-identification requires removing all identifiers and quasi-identifiers, not just names; improper de-identification doesn't provide privacy protection
• Security certifications matter: Look for ISO 27001, SOC 2, HIPAA-compliance where applicable; use local tools if cloud isn't secure enough

—

Reflection Questions

1. Your sensitive data: Does your research involve sensitive data? What kinds? What regulations or institutional policies constrain how you can use it?

1. Tool assessment: For tools you want to use, what's their privacy policy? Can you use them safely with your data, or do you need alternatives?

1. Secure workflow: For your specific research, how would you design a workflow that safely uses AI while protecting sensitive data?

1. Institutional support: What does your institution provide (enterprise agreements, local tools, guidance) to support secure AI use with sensitive data?