AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Beginner › Human Review and the PAUSE Rule

Level 2 of 4 · Beginner · Lesson 6 of 6

Human Review and the PAUSE Rule

The PAUSE Rule, and the human responsibilities that remain when AI is part of the work.

About 20 minutes

Skip to lesson

By the end of this lesson you should be able to

  • Walk a task through the five gates of the PAUSE Rule
  • Describe what genuine human review involves, beyond glancing at an answer
  • Explain how a change in sensitivity or in your own expertise changes the decision

Why this matters

The earlier lessons explained what AI systems are, how they produce output, why they sound trustworthy, what goes wrong, and which rules apply. This lesson focuses on two important points of responsibility: deciding whether to use AI for a task, and reviewing what it produces before you rely on it. Responsibility does not pause in between, but these are where the key decisions are made.

The PAUSE Rule, in brief

The Robert Crown Law Library’s PAUSE Rule is a sequence of five gates, taken in order. A “no” at any gate means using an alternative or stopping.

P — Pinpoint your task(s)

Break the goal into specific tasks, each with a clear goal, audience, and constraints. Task first, tool second.

A — Assess AI-fit

Is AI the right tool for this task? Drafting and brainstorming often fit; tasks that depend on facts need reliable sources. A template, a spreadsheet, or doing it yourself may be better.

U — Use ethically

Can you use AI appropriately, lawfully, and respectfully here? That includes whether the information is permitted in the tool, respect for privacy and intellectual property, and disclosure. For what counts as permitted, rely on the live sources from Lesson 5.

S — Self-assess your skills

Do you have the knowledge to judge whether the output is right? If you cannot evaluate it yourself, do not rely on it: stop, or go ahead only with a qualified reviewer or a reliable way to check.

E — Enhance your workflow

Does AI genuinely improve on doing it yourself—in time saved once checking and fixing are included, or in quality, accessibility, or learning—and does it help you keep and build your own skills?

The full PAUSE page adds a decision flow, examples for faculty, students, and staff, and prompts for using AI to learn rather than only to produce.

What human review actually involves

“Human in the loop” means little if the human only glances at an answer and accepts it. Review is work, and it has parts:

  • Compare claims with evidence. Open the source. Check quotations, names, numbers, and dates against the original, not against the AI’s description of it.
  • Check permissions before using or sharing. Course policy, the information’s classification, rights in any material, and whether disclosure is required.
  • Be honest about your expertise. Could you spot a wrong answer here? If not, do not rely on an unreviewed result: get review from someone qualified, use a reliable way to check, or set the output aside. Using AI while you learn a subject can still work, if a qualified person or a reliable source checks what you learn.
  • Count the repair time. Checking and fixing are part of the cost. If they erase the time saving, consider whether a concrete gain in quality, accessibility, or learning still justifies using the tool. If not, use the alternative.
  • Know when to stop. Change tools, do it yourself, or ask a person. Rewording the request until an answer looks right is not review.

Two habits look like review and are not. The first is automation bias: the tendency to over-trust automated output, especially when it is fluent or has been right before. The second is treating an AI’s agreement as verification. AI can help with checking: it can locate the passage in a source, flag a possible error, or compare a quotation with the original, and you then confirm what it found. But a second assistant, or the same one asked again, can repeat the same error or simply accept the framing of the question. Its agreement alone is not independent verification.

Even careful review manages risk; it does not guarantee an error-free result. That is one reason to match the depth of review to the stakes.

Worked example: a low-risk task through PAUSE

A constructed example; the event and people are fictional.

A staff member needs a short announcement of a public lunchtime talk for the school’s weekly newsletter, based on the talk’s already-published event page.

  1. Pinpoint: a 120-word announcement for students, faculty, and staff, including the title, speaker, date, time, room, and registration link.
  2. Assess AI-fit: drafting from a supplied public text is a good fit. Every fact is on the event page, so every fact can be checked.
  3. Use ethically: the source is public, Low Risk information; the staff member uses a Stanford-approved service and follows the office’s practice on AI-assisted communications.
  4. Self-assess: the staff member writes these announcements regularly and can check the details against the page.
  5. Enhance: writing the announcement by hand usually takes this staff member about fifteen minutes. With AI, drafting takes a minute, checking about three, and adjusting the wording a few more. The saving is real only if the check is actually done. (Illustrative times for this fictional example, not measurements.)

Time is not the only possible benefit. A draft might also improve clarity, or help produce an accessible or plain-language version, but none of that removes the privacy, permission, or review requirements.

Decision: go, with a line-by-line check of the details against the event page before it is sent.

Try it: change one condition

Start from the worked example above. For each change, which gate does it affect, and what should the staff member do?

  1. The message is no longer a public announcement. It goes to six students who were not admitted to a limited-enrollment program, explaining the decision, and the staff member wants to draft it from their application files.
  2. The announcement is the same, but it must go out in Spanish, and the staff member does not read Spanish.
  3. The event page has not been published yet, and the speaker’s title and the room are still being confirmed.
Discussion
  1. Use ethically, and Assess AI-fit. Admission files are student records, which Stanford classifies as at least Moderate Risk, so the information question comes first: is there a service approved for it, is the staff member signed in with the Stanford-provided account, is this access necessary, and what does the office’s own guidance say? Even with all of that settled, fit is doubtful: individualized decision letters call for care and human judgment. A reasonable outcome is to write them by hand, or use AI only for a generic template with no student information in it.
  2. Self-assess. If you cannot read the output, you cannot review it. The decision changes to “stop,” or “go only with a qualified reviewer,” such as a colleague fluent in Spanish or a professional translation service.
  3. Assess AI-fit, at the verification step. The problem is the unconfirmed title and room, not the fact that the page is unpublished: unpublished information can be reliable and authorized for use. Draft from the details that are confirmed, with clearly marked placeholders for the title and room, and fill them in once they are confirmed. Any title or room the tool supplies on its own would be unsupported.

In each case the tool is the same. What changed was the information, the reviewer’s ability to check, or the evidence available, and the decision followed.

You keep the responsibilities of your role. Using an AI tool does not hand them to the tool. Nor does it remove the responsibilities that vendors, supervisors, and institutions have for their own parts. Decide deliberately, check against evidence, and be willing to stop.

Level check: putting the Beginner lessons together

A self-check for the whole level, not a graded test. It asks how you would review this workflow, not whether you know the legal doctrine involved.

A constructed scenario.

A research assistant for a faculty member uses an AI research tool, available through a Stanford-approved service, that searches a database of public court opinions and writes a one-page summary of five opinions on a question, with citations. The summary ends: “All five courts agreed on the standard.”

  1. What kinds of work is the tool doing? (Lesson 1)
  2. What would show whether the statement “All five courts agreed” is supported? (Lesson 2)
  3. The summary is confident and cleanly written. What does that establish? (Lesson 3)
  4. What does reviewing a summary like this require? (Lesson 4)
  5. What should the research assistant check before sharing it or building on it? (Lessons 5 and 6)
Discussion
  1. Retrieving, then generating. The opinions are real records; the summary and the conclusion about agreement are generated.
  2. Only the opinions themselves. The statement may accurately summarize them, may be an inference that needs checking, or may be an unsupported generalization. The sentence alone cannot tell you which, and neither can the tool’s own account of how it reached it.
  3. Nothing about accuracy. Confidence and polish are features of the writing.
  4. Reading the five opinions, or at least the relevant passages, to confirm the standard each court applied, and checking that the citations match. A claim that sources “agreed” is exactly where summaries lose distinctions.
  5. Whether the faculty member’s project, and any publication it feeds, has rules about AI use or disclosure; whether the research assistant has the expertise to judge the legal analysis or needs the faculty member’s review; and, before anything is relied on, that the claims match the opinions themselves.

Sources and further reading