Skip to content
Idaho AI Classroom
All lessons
AI Technical AptitudeAI Methods9-12.AITA.AIM.7

Human in the Loop, On Purpose

9–12 CS / engineering50 minutes90 minutes (lab extension)

Standard quoted exactly

Plan safeguards for AI systems that protect human well-being and privacy while ensuring meaningful human involvement in decision-making.

Example from the standards. Web search can return anything, or you can filter it for child-appropriate content.

Student-friendly learning targets

  • I can plan safeguards that protect well-being and privacy, not only accuracy.
  • I can place a meaningful human-in-the-loop step where a person can actually change the outcome.
  • I can use filtered vs. unfiltered web search as a concrete safeguard example.

Essential questions

  1. What is the difference between a human who rubber-stamps and a human who can say no?
  2. Which harms are about well-being, which are about privacy, and which are both?
  3. If we filter web search for a classroom agent, what do we lose and what do we protect?

Objectives

  1. Inventory harms for one Idaho system (wildfire paging, school help agent, load-shed suggestion).
  2. Plan at least five safeguards: filter, approval, confidence escalate, PII strip, logging/kill switch.
  3. Mark humans as meaningful (they have time, information, and authority) not decorative.
  4. Apply the standards example: unfiltered vs. child-appropriate filtered search.

Key vocabulary

Safeguard
A designed control that reduces harm: filters, approvals, rate limits, PII stripping, kill switches, confidence gates.
Human-in-the-loop
A person must verify, approve, or correct before the system acts. Meaningful means they can change the action, not just watch it.
Well-being
Physical and psychological safety: crews not paged into nothing, students not shown harmful content, patients not dosed by a model.
PII strip
Removing names, IDs, addresses, and other identifiers before data hits a model or a log.
Content filter
A rule layer that blocks categories of web or model output (e.g., child-inappropriate search hits) before a user sees them.

Teacher background

AIM.2 required an approval diamond; AIM.7 is a full safeguard plan for well-being, privacy, and meaningful humans. Teach the web-search example first: an agent with unfiltered search can pull anything; a classroom or library agent should use a child-appropriate filter — a rule layer around a data-driven tool (AIM.4 combo). Then plan a wildfire 'page crews' system: confidence threshold, human dispatcher, no auto-tone-out, logging, privacy of camera homes. Rubber-stamp humans (0.3 seconds to click OK) fail the standard. Spreadsheet of harm × safeguard is the artifact. Offline card sort of safeguards onto a system diagram. No live unfiltered search demos of harmful content.

Materials and prep

Materials

  • System-at-risk one-pagers: (A) classroom web-search agent, (B) wildfire page-crews recommender, (C) school lab-assistant from AIM.2.
  • Safeguard cards: filtered search, unfiltered search, human approve, confidence escalate, PII strip, rate limit, log+audit, kill switch, age gate, no-retention.
  • Meaningful-human checklist: time, information, authority, backup person, no punishment for saying no.
  • Plan template: harm, who is hurt, safeguard, where in the flow, how we know it works.

Before class

  • Do not open unfiltered search on a projector. Describe the risk; use the filter as the live path if you demo at all.
  • Coordinate with AIM.9: paging crews will return as false positives. Today is the human gate; next is the metric.

Instructional sequence

Search can return anything

5 min
  1. Read the standards example. Students list three categories a 9th-grade lab agent should not display.
  2. Ask: is a filter a model or a rule? (Often a rule or a separate classifier with a human policy.)

Well-being, privacy, meaningful humans

10 min
  1. Three goals from the standard. Map safeguards to each (filter → well-being; PII strip → privacy; approval with authority → meaningful involvement).
  2. Show a fake loop: a dispatcher has 80 alerts/minute. That human is not meaningful. Redesign: fewer alerts, better ranking, legal power to ignore.
  3. Kill switch: a labeled way to turn the recommender off during a real incident.

Filter the search agent

12 min
  1. On system A, place unfiltered vs. filtered search cards. Write who sets the policy (librarian, admin, vendor — debate).
  2. Add PII strip (no student names in queries stored) and no-retention.
  3. Run the meaningful-human checklist on 'teacher can override the filter.' Who audits the override?

Plan for paging or the lab agent

10 min
  1. Pairs take system B or C and fill five rows of the plan template, including at least one well-being, one privacy, and one human-in-the-loop safeguard.
  2. They mark any human step that is currently a rubber stamp and fix it.
  3. Offline: cards on the one-pager diagram.

Real-world examples

  • Classroom search: filtered vs. raw web, exactly the standards example.
  • Wildfire: a model that auto-pages crews without a dispatcher harms well-being (fatigue, missed real fires) and can leak camera views of homes (privacy).
  • Idaho Power load-shed suggestions must not auto-cut a hospital feeder; a human operator with authority is the safeguard.
  • A counseling chatbot that cannot escalate to a human fails well-being even if it 'respects privacy.'

Hands-on activity

Red-team the plan

8 min
  1. Neighbor tries to route around the human (script the approval, flood the queue, put PII in a screenshot).
  2. Pairs add one more safeguard and initial the checklist.
  3. Share one failed rubber-stamp they caught.

Discussion questions

  1. Who is harmed if the child-appropriate filter is too aggressive? Too weak?
  2. Can logging be both a safeguard and a privacy hole?
  3. Should a kill switch live in software, in a physical button, or both?
  4. When is 'the human is in the loop' a slogan that hides an unsafe system?

Differentiation

Support

  • System A only (search filter) with three safeguards pre-named to explain.
  • Checklist with yes/no boxes.

Challenge

  • Write a policy paragraph on filter overrides: who, logging, time limit.
  • Design the wildfire queue so a dispatcher can be meaningful at 2 a.m. (rate, grouping, map).

Multilingual learners

  • Safeguard names with glosses; plans may be bilingual.
  • Role-play the approval conversation.

IEP / 504

  • Card placement is sufficient; writing reduced to three rows.
  • Avoid graphic wildfire imagery; stick to process.

Assessment

Formative

  • Three categories listed in the warm-up.
  • Meaningful-human checklist used in guided practice.

Summative

  • Plan template with five safeguards covering well-being, privacy, and a non-rubber-stamp human step.
  • Must include filtered search or an analogous content/action filter.

Success criteria

  • Student's human can say no with time and authority.
  • Student names a privacy safeguard (PII, retention, cameras).
  • Student treats filters as a designed control, not as censorship trivia.

Responsible use, ethics, and privacy

Responsible use

Do not demo harmful unfiltered results. Plan on paper. District filter policy wins over a vendor default.

Ethics

Safeguards can be paternalistic or captured by vendors. Students should name who sets the filter and how it is reviewed.

Privacy

FERPA: student queries are education records if stored. Default to no-retention. Camera systems must not enroll student faces.

Reflection

  1. Where was your human not actually meaningful on the first draft?
  2. Which safeguard would you fund first with a small budget?
  3. How will you explain filtered search without sneering at it?

Homework

Revise the AIM.2 agent flowchart with three labeled safeguards (filter, approval, PII). Write four sentences on what makes the human meaningful. Paper; no live agents.

Closing

Web search can return anything — or you can filter it. Crews can be auto-paged — or a person with authority can say no. Safeguards are designed, not hoped. Next we look at the whole training pipeline from the methods side, including leakage.

Extensions and cross-curricular links

Go further

  • 90-minute block: tabletop incident — unfiltered query, then a wildfire false alarm flood; revise the plan.
  • Python extension: a stub filter list and an approve() gate in code comments as a design, not a product.
  • Connect to AIM.2 flowcharts and upgrade their diamonds using today's checklist.
  • Printed safeguard card sort for subs.
Health / counseling
Escalation to a human is a well-being safeguard, not a model failure.
Government / library science
Collection filters and challenge policies already exist; AI search inherits that debate.
Emergency services CTE
Dispatch already uses human-in-the-loop; models must not erase it.