AI Explained Plainly home

Guide · · 5 min read

AI and OPSEC: what never goes into a chatbot

Service & Veterans

Want one workflow like this by email each Tuesday? It's free. Get it below

Most OPSEC slips with AI don't look like leaks. They look like someone pasting a messy email thread to get a cleaner summary. This guide is the checklist to run before that paste, and a routine for scrubbing what's left.

If you're still sorting out which tools are allowed at all, start with Can I use ChatGPT in the military? This page assumes you know the basics and want the practical list.

Why a chatbot is an OPSEC problem

OPSEC is about critical information. DoD's OPSEC manual (DoDM 5205.02) defines it as information about "activities, intentions, capabilities, or limitations that an adversary seeks." It also warns that admin staff hold this kind of information too, not just operators.

Two things make AI tools risky here. First, the paste leaves your control. The Department of the Navy CIO's 2023 guidance says chatbots "save every prompt they are given, which presents a persistent Operational Security (OPSEC) risk." Army CIO guidance adds that users "should have no expectations of privacy" for what they type in. Second, small pieces add up. The same memo warns that "the aggregation of individual generative AI prompts and outputs could lead to an inadvertent release of sensitive or classified information." The OPSEC manual makes the same point about unclassified information that can "individually or in the aggregate" lead to a compromise.

So the question isn't only "is this one line sensitive?" It's also "what does this line say next to everything else I've pasted this month?"

The do-not-paste checklist

None of this goes into a personal ChatGPT, Gemini, Claude, or Copilot account. Ever.

  • Classified anything. Not in any of these tools. Even GenAI.mil is designated for CUI, not classified work.
  • CUI. DoDI 5200.48 says DoD personnel won't use personal accounts or "other non-DoD information systems" for official business involving CUI. The National Archives' CUI Registry lists categories that include Military Personnel Records, Health Information, Controlled Technical Information, and Operations Security.
  • Your unit's critical information. Whatever is on your command's critical information list. Ask your OPSEC coordinator for it. The manual says personnel should know what's on it.
  • Schedules and movements. Deployment, exercise, and travel dates, duty rosters, flight or underway times, and leave dates tied to a unit.
  • Locations and layouts. Unit names with locations, building and room numbers, gate and access procedures, and photos or screenshots of work spaces.
  • Capabilities and problems. Readiness numbers, equipment shortages, maintenance status, and security weaknesses.
  • System details. Network diagrams, error logs, trouble tickets, and screenshots from government systems.
  • People. Names with SSNs, DoD ID numbers, dates of birth, home addresses, medical details, or family information. That's PII and PHI.
  • Forwarded email. Signature blocks and headers carry names, units, and phone numbers. Delete them before you paste.

One trap: approved doesn't mean anything goes. A January 2026 Navy memo calls GenAI.mil "fully authorized to handle CUI," but says PII and PHI "are not permitted for use on this platform." So you still scrub people's details there. Your command may add its own rules on top, and its rules win.

The five-step scrub routine

Do this before any paste, even into an approved tool.

  1. Start from your own words. Retype the gist rather than pasting a real document. A summary you wrote leaks less than a forwarded thread.
  2. Swap specifics for placeholders. [NAME], [UNIT], [LOCATION], [DATE], [NUMBER], [SYSTEM]. Keep a short key on paper or in an approved system, never in the chat.
  3. Blur the numbers. "About half," "a few weeks," and "most of the section" usually work as well as the real figure.
  4. Check against the list. Read the scrubbed text against your unit's critical information list and the checklist above.
  5. Do the aggregation test. Ask yourself what someone could piece together from this prompt plus your last few. If the answer is a schedule, a location, or a weak spot, cut more.

Then put the real details back yourself, outside the AI tool.

Prompt 1: work with placeholders

Paste this, then your scrubbed text. It keeps the tool from "helpfully" filling in blanks.

The text below uses placeholders like [NAME], [UNIT], [LOCATION],
[DATE], [NUMBER], and [SYSTEM]. Rules:
- Keep every placeholder exactly as written. Don't replace,
  guess, or invent names, places, dates, or numbers.
- Don't ask me what the placeholders stand for.
- If the task needs a detail I left out, write [NEED DETAIL]
  instead of making one up.

Task: [say what you want, e.g. "make this clearer and shorter"]

TEXT:
[paste your scrubbed text here]

Prompt 2: a second set of eyes

Use this only on text you've already scrubbed. It's a backup check, not a substitute for steps 1 to 5.

Review the text below for anything that could still identify a
person, unit, place, date, schedule, capability, or weakness.
List each item you find and suggest a placeholder for it.
Don't rewrite the text. Don't guess at what any placeholder
stands for.

TEXT:
[paste your scrubbed text here]

If it flags something, fix it in your own copy. And if it flags nothing, that's not a clearance. It's a chatbot.

Settings help, but they don't fix OPSEC

As of Oct 2026, ChatGPT has an "Improve the model for everyone" switch under Settings, then Data controls. OpenAI says turning it off stops new chats from being used for training, but it doesn't delete saved chats. Google says Gemini chats can be read by human reviewers and asks you not to enter confidential information you wouldn't want a reviewer to see. Reviewed chats are kept for up to three years, even if you delete your activity. Issue #1 shows where these settings live in each assistant.

Turn training off anyway. Just don't treat it as permission to paste anything on the checklist.

If you already pasted something

Don't just delete the chat and hope. Deleting doesn't pull back what was sent, and with Gemini a reviewed chat can outlast your delete. Tell your security manager or OPSEC coordinator what went in and where. DoDI 5200.48 has commands report and correct unauthorized disclosures of CUI, and it's much easier to fix when they hear about it from you first.

Writing evaluations? The same scrub routine is built into using AI for EPB, NCOER, and FITREP bullets. Before your next paste, ask your OPSEC coordinator for the unit's critical information list. Five minutes with it beats any setting.

Sources

All checked October 22, 2026.

Policies and settings change, so confirm with your command and your security manager. How this is made: an AI editor drafts and fact-checks each article. If we get something wrong, we fix it and note the change.