← AI 101

Beginner training

Trust, Safety, and Your Data

A clean rulebook for what data is safe in Claude, what isn't, what a hallucination looks like, and when to stop and escalate. The foundation underneath every other module in this curriculum.

17 minutes Builds on Module 1.3 Includes data classifier

Start here

What you must be able to do

Classify before you paste

Sort a piece of information as safe, sensitive, or off-limits before it ever lands in a Claude conversation. Default to caution.

Spot a hallucination

Recognize when Claude's confident tone is doing the talking. Know which parts of an answer to verify against a real source.

Know when to escalate

Identify the Claude tasks that need human or compliance review before they leave your screen — and the path to take when they do.

The three buckets that matter

Confidentiality, accuracy, accountability

Confidentiality

The chat is a boundary. Some data crosses it safely, some doesn't. Knowing the difference — before you paste — is the single most important habit in this module.

Accuracy

Claude can be confident and wrong. The model is trained to sound right, not to be right. Tone is not truth. Every name, date, number, and citation needs a human-verified source.

Accountability

The output has your name on it the moment you send it. The employee owns the work — Claude is a tool, not a colleague you can blame. Every shipped answer is your answer.

Trainer note: "Claude said it" is never an acceptable answer. Not in an audit, not in a meeting, not in a borrower conversation. Treat the output the way you'd treat a draft from a smart but unverified intern.

Where the risk lives

Three surfaces, three common mistakes

The chat window

The most common mistake is pasting regulated or confidential data — borrower PII, internal pricing, draft strategy — into a Claude chat that wasn't cleared for that data class.

What you type

The attachment

The most common mistake is uploading a file with embedded PII you didn't think about: a credit-report PDF, a loan file with cover pages, or a screenshot with names still visible.

What you upload

The output

The most common mistake is forwarding an answer that contains a hallucinated citation, an unverified number, or a tone that's wrong for the audience — once you send it, it's your statement.

What you forward

Safe daily use

Five rules to keep the work clean

Treat every conversation like it could be reviewed

Compliance, legal, your manager. Don't type anything in a Claude chat you wouldn't be comfortable having pulled into a routine review.

Never paste regulated data without checking policy

Borrower PII, credit data, employee records, investor pricing — every regulated category has a documented home. Default is "not in Claude" until policy says otherwise.

Verify every name, date, number, and quote

Before any answer leaves your screen for a customer, partner, or executive — verify the specifics against the actual source. Tone fluency is not accuracy.

Ask Claude to cite, then check the citation

When the stakes are real, ask for the source — and then open the source. Citations that look right but don't exist are a classic Claude failure mode.

Stop and escalate the second you're unsure

Uncertainty is a signal, not a problem. Compliance, legal, security, your manager — know your paths and use them. "I checked with X" is always a better answer than "I assumed."

Prompt upgrade

Weak prompt

Borrower Jane Smith, SSN 123-45-6789, 740 FICO, asks me what her LTV options are at 95%. What should I tell her?

Safe prompt

Work-ready prompt

A borrower (de-identified — call them "B") has this profile: 740 FICO, 95% LTV on a conventional loan, W-2 income, no co-borrower. List the LTV-related options I should walk them through, with the trade-offs of each. No PII in this conversation.

Safe places to start

Four low-risk Claude patterns for any employee

Drafting from your own notes

Your own scratch notes, meeting summaries, and rough thinking — with no client PII — are the safest place to build the habit of using Claude well.

Summarizing public or internal documents

Public agency guides, marketing materials, and company-internal documents you're already authorized to read can be summarized and queried safely.

Exploring options on non-sensitive decisions

Pressure-test a plan, list trade-offs, generate counterarguments — anywhere the decision itself isn't regulated and the inputs aren't confidential.

Rewriting your own writing

Tightening an email you wrote, simplifying a paragraph in your own voice, adapting your draft for a different audience. Low-risk and high-leverage.

Red flags before you send

Five signals to stop and verify

Hallucinated citation — a page number, section, or law that doesn't actually exist when you check the source Confident claim with no source — a specific assertion the answer doesn't tie back to anything verifiable Regulated data sitting in the conversation — PII, credit data, internal pricing, anything from a restricted bucket An answer that touches legal, compliance, HR, or customer-facing matters without a human review step Any output you wouldn't be comfortable having read aloud in front of leadership

Employee rule: Default to escalation, not to shipping. If you're asking yourself whether to send it, the answer is to verify first. The cost of a slow answer is small. The cost of a wrong answer with your name on it is large.

Build the habits

Six exercises that turn the rules into reflexes

The goal is not to memorize policy — it's to make the right move automatic.

Open the data classifier in this lesson. Run all six items, then re-run any you got wrong. Note which categories you confuse.
Find a planted hallucination: ask Claude a factual question about your area of work and check every citation. Note which kinds of mistakes look most convincing.
Rewrite a prompt that leaks data — take a real prompt with PII in it and rebuild it as a de-identified version.
List every escalation path you should know: compliance, legal, security, your manager, IT. Write down who, when, and how.
Write a one-paragraph escalation note. Practice the language now so it's ready when you actually need it.
Commit to one new verification habit you'll use every day. Tell a teammate so they can ask you about it next week.

Completion standard

You've finished this module when you can name, in plain English: one data type you'll never put in Claude, one habit you'll use to verify every answer, and one path you'll use to escalate when you're unsure.