A Practical Learning Experience

AI with Judgment

How to Use, Verify, and Take Responsibility for AI-Generated Work

10–12 minutes Self-paced 7 short lessons

A Polished Answer Is Not Necessarily a Reliable One

Generative AI can draft, summarize, compare, analyze, and organize information remarkably quickly. It can also omit context, invent details, misread a source, or present uncertain information with complete confidence.

Using AI well requires more than writing a prompt. It requires knowing what to question, what to protect, what to verify, and what still needs human judgment.

By the end of this course, you will be able to:

Recognize when an AI-generated response requires closer review.
Protect confidential, personal, proprietary, and sensitive information.
Use a simple Review, Verify, Decide framework.
Match the level of verification to the potential consequences.
Know when human expertise, authorization, or accountability is essential.

An AI assistant gives you a clear, professional response with several specific facts. What does the polished writing tell you about the facts?

The facts are probably correct because the response sounds confident. Nothing by itself. The facts still need to be evaluated and, when important, verified. The facts are safe to use if the response includes citations.

1

Fluent Does Not Mean Factual

Generative AI produces a likely response based on patterns in data and the instructions it receives. It does not automatically know whether every statement is true, current, complete, or appropriate for your situation.

What AI does well

Drafting, restructuring, brainstorming, summarizing, comparing, explaining, and identifying possible patterns or questions.

What AI cannot guarantee

Accuracy, completeness, current information, valid citations, sound interpretation, appropriate authorization, or an ethically responsible decision.

AI can generate an answer. It cannot accept accountability for using it. The person who applies the output remains responsible for reviewing it and deciding what happens next.

Warning signs that deserve a closer look

Select the signals you want to remember.

2

Match the Tool and the Task

Before prompting, pause long enough to decide whether AI is appropriate. A useful task is not automatically an authorized task.

Is this an approved tool?

Use only tools and accounts your organization has authorized for the type of work you are doing. A public AI account and an enterprise-approved account may have different safeguards.

What information would I need to provide?

If the task requires confidential, personal, proprietary, regulated, or security-sensitive information, stop and follow organizational policy before entering it.

What happens if the output is wrong?

The higher the potential impact on people, patients, customers, finances, compliance, safety, or reputation, the stronger the review and verification must be.

Who is authorized to make the final decision?

AI may support analysis, but it should not silently replace required professional review, approval, or accountable decision-making.

You want AI to organize themes from employee comments. The comments contain names and health information. What is the best first step?

Paste the comments into your personal AI account and delete the conversation afterward. Ask the AI not to remember the information. Stop and confirm the approved tool, permitted data, and required protections before proceeding.

3

Protect Before You Prompt

Good judgment begins with the information you choose to provide. Use the minimum information necessary, remove identifying details when appropriate, and follow your organization’s rules for data handling.

Personal information

Names, contact details, performance information, health information, identifiers, and other data connected to a person.

Confidential information

Internal plans, contracts, unpublished results, financial information, investigations, and sensitive communications.

Proprietary information

Trade secrets, source code, research, product plans, intellectual property, and nonpublic business methods.

Security information

Credentials, access details, system configurations, vulnerabilities, and information that could increase risk.

Build a safer request

Select one option in each group. The activity will combine your choices into a privacy-conscious prompt.

1. What is the task?

2. What information will you provide?

3. What boundary should the AI follow?

4. How should the response support review?

4

Review, Verify, Decide

Use this three-part framework before applying AI-generated work. The amount of effort should reflect the importance and potential consequences of the task.

Step 1

Review

Read critically. Look for missing context, unsupported certainty, altered meaning, bias, weak reasoning, fabricated details, and anything that does not fit the situation.

Step 2

Verify

Check important facts, calculations, quotations, citations, interpretations, and requirements against authoritative, current sources or qualified experts.

Step 3

Decide

Determine whether the output is appropriate to use, revise, escalate, or discard—and whether required approval or expert review has occurred.

AI produces a concise summary of a new policy. You compare it with the official policy and discover that an exception was omitted. Which step have you completed, and what should happen next?

Decide; publish the summary because most of it is correct. Review; ask AI whether the exception matters and accept its answer. Review and verify; correct the omission, assess its impact, and obtain any required approval before use.

5

Validation Lab: Find the Problems

Imagine an AI assistant was asked to summarize an internal research update for leaders. Compare the source notes with the AI-generated summary.

Source notes

An early pilot included 24 participants.
Participants reported that the new process may reduce preparation time.
No control group was used.
The project team recommended further testing before broader adoption.

AI-generated summary

“The study proved that the new process reduces preparation time by 42%. The successful trial demonstrates that the process is ready for organization-wide implementation.”

What needs to be challenged?

Select each issue you notice.

6

Match Verification to Consequences

Not every AI-assisted task carries the same risk. The more consequential the use, the more rigorous the validation and oversight should be.

Lower consequence: Brainstorming possible titles for an internal, informal meeting.
Review for usefulness, tone, and fit.

Moderate consequence: Drafting a client communication from approved information.
Confirm facts, context, tone, commitments, and required review before sending.

Higher consequence: Interpreting scientific, medical, legal, financial, employment, safety, compliance, or security information.
Use authoritative sources, qualified human expertise, and required organizational review. Do not treat AI output as the final authority.

AI recommends a course of action that could affect a person’s employment. What is the most appropriate response?

Use the recommendation if the AI provides a detailed rationale. Treat it only as possible input, examine the data and assumptions, consider bias and context, and preserve required human review and accountability. Ask a second AI tool and use the recommendation if both agree.

7

Practice: Apply Your Judgment

1. AI gives you a quotation and a link to its supposed source. What should you do before publishing the quotation?

Use it because a link was provided. Ask AI to confirm the quotation one more time. Open the original, authoritative source and confirm the exact wording, speaker, context, and date.

2. Which practice best protects sensitive workplace information?

Remove names, then use any convenient public AI tool. Confirm the approved tool, permitted data, and organizational requirements before entering the information. Enter the information and delete the conversation afterward.

3. Which action belongs in the Review step?

Look for omitted context, unsupported certainty, changed meaning, bias, and weak reasoning. Publish the output after correcting the spelling. Ask AI whether its own answer is accurate.

4. When should verification be most rigorous?

Whenever the response is more than one paragraph. Only when AI admits that it is uncertain. When errors could materially affect people, patients, customers, safety, compliance, finances, or important decisions.

5. Which statement best describes responsible AI-supported work?

If AI saves time, a detailed human review is unnecessary. AI may support the work, but people remain responsible for appropriate use, verification, decisions, and outcomes. Responsibility is shared equally between the employee and the AI provider.

0 / 5 Complete the five questions above.

Course Summary

A polished AI response is not evidence that it is accurate, complete, or appropriate.
Use only approved tools and follow organizational requirements for confidential, personal, proprietary, regulated, and security-sensitive information.
Review the content, reasoning, context, uncertainty, and fit.
Verify important claims using authoritative, current sources or qualified expertise.
Decide whether to use, revise, escalate, or discard the output.
Match the rigor of verification and oversight to the potential consequences.
AI can support work, but it cannot assume human responsibility or accountability.

Choose one commitment

Select the action you will apply first.

The goal is not to distrust everything AI produces. The goal is to use it with judgment proportional to the situation.

Thank you for taking the time to build your confidence with generative AI.

Start small, stay curious, and remember: Ask, Review, Refine, and Verify.