Independent resource · not affiliated with or endorsed by Clemson University

AI Workshops by ProvenLabs

Resource · Guide

Security basics by data type.

Every credible institution lands on the same method: classify the data before you pick the tool. Get that order right and most of what could go wrong never comes up.

This page teaches the method. The classifications below follow Clemson’s published policy as verified on August 5, 2026; where this page and the official sources disagree, the official sources win.

The data classes

Clemson, like most universities and the State of South Carolina, sorts data into four tiers. Before anything goes into an AI tool, know which tier it belongs to. That single habit prevents most of what goes wrong.

Public

Developed and intended for public disclosure.

Public websites, published research data, campus maps, non-private contact info.

Internal Use

Not public, minimal adverse impact if breached, used in daily operations.

Non-public policies, training materials, internal emails, memos, budgets, reports.

Confidential

Sensitive, with adverse impact if breached.

Student records and FERPA-protected data, HR records, PII, donor info, survey data with identifiers, birth dates and addresses.

Restricted

Highly sensitive, protected by law, with statutory notification obligations if breached.

SSNs, health information under HIPAA, financial aid data, card and bank numbers, passport numbers, export-controlled data.

The rules that follow

  1. 1. Classify before you paste.

    Ask what tier the data is first, then which tools are approved for that tier. Michigan, Stanford, and Harvard all structure their guidance exactly this way; none of them start with the tool.

  2. 2. Consumer tools get Public data only, whatever account you’re signed in with.

    Anything above the lowest tier stays out of free ChatGPT, Claude, Gemini, and the rest. Protection comes from an institutional contract with a specific tool, and at Clemson the contracted environment is ChatGPT Edu.

  3. 3. Know what the institutional contract changes.

    An Edu agreement typically buys no-training-on-your-data, encryption, and institutional ownership. It does not make every data type eligible: classification ceilings still apply, and some data (student records among them) carries extra conditions like approval from the Data Trustee and a University-managed device. Check the official guidance for the current rules.

  4. 4. Treat connectors as a data-sharing decision.

    Once an AI can read your mail, drive, or calendar, everything in there is exposed to wherever the AI sends data, and security researchers demonstrated in 2025 that a single poisoned shared document can silently exfiltrate connected files with zero clicks. This is why connectors require a submitted approval request before use.

  5. 5. De-identify properly, and know its limits.

    Stripping identifiers is the legitimate route for working with sensitive material, and the guidelines endorse it. But de-identification is harder than it looks at small numbers, and long AI context windows make re-linking easier. When the population is small, treat the data as identified.

  6. 6. If it goes wrong: stop, preserve, report. Fast.

    Sensitive data in the wrong tool is a reportable security incident. Don’t delete the chat, that destroys the evidence the security team needs and recalls nothing. Call the CCIT Support Center at 864-656-3494 the same day. Early reporting is what makes it a contained problem, and every credible institution treats self-reporting as protection for the person who reports.

Common misconceptions, retired

“Temporary chat means it's safe.”

Temporary chat changes what gets saved to your history. The data itself still transmits to the provider's servers and is retained for a period for abuse monitoring, and in 2025 a federal court order required one provider to preserve even deleted and temporary chats for litigation. Data barred from a tool is still barred in a temporary chat.

“I deleted the conversation, so it's gone.”

Deletion removes it from your view. Server-side purging takes time, legal holds can suspend it entirely, and a separate memory feature may have already distilled facts that survive the chat's deletion. Deleting a chat, deleting a memory, and opting out of training are three different switches.

“I logged in with my university email, so it's covered.”

Coverage comes from the institution's contract with that specific tool. A .edu address on a free consumer account is still a consumer account. Peer universities are explicit about this: an uncontracted tool gets low-sensitivity data only, whatever account you used.

“I removed the names, so it's de-identified.”

The federal standard asks whether identity could be recovered, including by linking with other available information. In a class of twelve, “the transfer student who was out sick in October” identifies someone. Small populations are effectively never de-identified by name removal alone.

“It's the enterprise version, so anything goes.”

No-training and encryption are not the same as authorization. Institutions set classification ceilings on their own contracted tools, and those ceilings decide what you may enter, whatever the vendor's marketing says.

“Sharing the chat just sends a link.”

A share link is a publish button. In 2025, thousands of shared AI conversations, including resumes and sensitive disclosures, ended up indexed by Google through a sharing feature users didn't fully understand.

Sources