The AI Learning Hub Journal

Supervising Juniors Who Use These Tools

Supervising people who use these toolsthe aim is not to police use — it is to make verification the normal, sayable thingSET EXPECTATIONS FIRSTSay which tasks may use atool and what verificationis expected, before thework starts — not after.ASK TO SEE SOURCESAsk for the underlyingmaterial, not a confidentsummary of it. A sourceyou can open, or nothing.REVIEW THE REASONINGAsk how the conclusion wasreached. Output can beright by accident; thereasoning cannot.MAKE ADMITTING IT SAFEA junior has to be able tosay the tool produced thisand I could not verify it,without it costing them.THE FAILURE MODE THIS PREVENTSTIME PRESSUREa deadline that leavesno room to checkSILENCEadmitting tool usefeels like a riskNO SUPERVISIONnobody knows there isanything to checkUNVERIFIED WORKgoes out under asupervisor's nameFear of the second box is what produces the fourth — removing that fear is supervisionWHAT THE SUPERVISOR IS ACTUALLY ACCOUNTABLE FORThe output carries your name as well as theirs — reviewing the reasoning is how you supervise itCommon professional principles, described generally; how supervision duties apply to you varies by jurisdictionVerification you can see beats verification you were promised
A junior who cannot safely say the tool produced this and I could not verify it will hand you something unverified instead

Assume Use Unless You Have Made Non-Use Realistic

Supervisors who assume juniors are not using AI are usually wrong, and the assumption makes things worse: undisclosed use is unsupervised use. Where a firm has a restrictive policy but demanding time targets, the predictable outcome is quiet use and no disclosure, which removes the supervisor's ability to check anything. The more effective posture is to make disclosure of AI use routine and consequence-free, so that supervision can attach to it. That means asking as a normal part of reviewing work — which parts were AI-assisted, which tool, what was verified — and reacting to the answer as information rather than as an admission. A junior who reports a fabricated citation they caught themselves should be treated as having done the job correctly.

  • Undisclosed use is unsupervised use — a restrictive policy plus tight targets produces exactly that
  • Make disclosure routine and consequence-free so supervision can attach to it
  • Ask which parts were AI-assisted, with which tool, and what was verified — as a normal review question
  • Treat a self-reported near-miss as the process working, not as a performance problem

Supervision Duty Does Not Delegate

Across jurisdictions, supervisory responsibility for the work of those you oversee is a recurring principle, and it applies to work produced with AI assistance just as it does to work produced without it. In practice this means a supervisor cannot discharge the duty by relying on an assurance that the junior checked. Review of AI-assisted work should include independently verifying a sample of citations rather than accepting the verification record at face value, because the record can be completed without the checks being done. The specific rules and their scope differ by jurisdiction, and firms should confirm their own position, but the underlying expectation — that supervision is active — is broadly shared.

  • Supervisory responsibility is a recurring principle across jurisdictions and extends to AI-assisted work
  • An assurance that checking happened is not the same as checking having happened
  • Independently verify a sample of citations rather than relying solely on the verification record
  • Specific rules and scope differ by jurisdiction — confirm your own before setting firm policy
  • In the US, ABA Model Rules 5.1 and 5.3 frame supervisory responsibility; states adopt their own versions, so check yours

Protecting the Development Pipeline

There is a longer-term supervision problem that sits beneath the immediate risk. The ability to review AI output well is built by having done the underlying work manually — you spot a mischaracterised holding because you have read enough judgments to feel the mismatch. If juniors move directly to reviewing machine output, that intuition never develops, and the firm ends up with reviewers who cannot review. Deliberate countermeasures help: research exercises done without AI, requiring juniors to read full judgments rather than retrieved extracts, and periodically having them find the error in a deliberately flawed AI draft. This is a training investment with no short-term return and a substantial long-term one.

  • Reviewing well depends on having done the underlying work manually often enough to feel a mismatch
  • Straight-to-review juniors become reviewers who cannot detect what they have never produced
  • Keep unassisted research exercises and full-judgment reading in the development programme
  • Error-spotting drills on deliberately flawed AI drafts build the specific skill that matters

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.