Monday's Activity, Today

The power outage cancelled Monday's class before we could run the budget-cut activity in person.

• We're running it as today's session.
• Everything below was originally written for Monday — the instructions haven't changed, only the day.

The Average Is a Decision

Ask a generative AI system almost anything and you get the "consensus", i.e., the middle of everything that it was trained on.

It's usually reasonable, and often useful. But that can be a problem.

A wrong answer invites you to check the work while a reasonable answer does not.

• Michigan's MiDAS (2013–2015): a rule-based automated system flagged unemployment fraud with no human review. 93% of the 22,000 determinations examined involved no fraud. ~40,000 people accused; some lost homes or filed bankruptcy.
• Australia's Robodebt: took annual income, divided it evenly across the year, and treated that average as fortnightly earnings. Fine for steady work; wrong by construction for casual or seasonal work. Wrongly recovered A$700M+ from ~400,000 people.

The average described nobody in particular, and the harm fell on the people it described least.

Today: Making Decisions from Generated Output

This week we add Actions, completing the DTPA framework.

You'll be handed a scenario with a budget you cannot meet and six programs — two of which must be cut.

Three phases:

1. Human Discussion Cut (Baseline) — no AI
2. GenAI Baseline Cut — hand the same table to at least two GenAI tools
3. GenAI Context Probe — ask what the table might be hiding

Please use the Google Doc posted in In-Class Materials on D2L for your work.

Phase 1 — Human Discussion Cut (Baseline)

No AI allowed yet.

• Discuss with your group using only the provided scenario table.
• Decide which two of the six programs to cut.
• Write a 3–5 sentence rationale for your group's decision.
• Note what missing information would change your mind.

Keep this rationale. You'll compare it against the GenAI outputs in the next two phases.

Phase 2 — GenAI Baseline Cut

• Feed your scenario table to at least two different GenAI tools (e.g., ChatGPT, Gemini, Claude).
  • You can do this individually to compare with each other.
  • These systems tend to acclimate themselves to your prior requests, so you are likely to get different decisions.
• Use the same table, same framing, for each tool. Don't add hints.
  • If the tool asks for more information, push it to make a decision and provide a rationale.
• Paste each tool's verbatim choices and rationale into your group's tab in the Class Decision Google Doc (MSU login required).

Compare: Did the AI tools agree with each other? Did either agree with your group's Phase 1 decision?

Phase 3 — GenAI Context Probe

Run a follow-up prompt asking the AI to identify critical hidden legal, financial, denominator, or systemic factors the table might be ignoring.

• Paste the verbatim response into your group's tab in the Class Decision Doc.
• Does the AI tell you anything your group missed in Phase 1?
• Does it tell you anything it missed in its own Phase 2 answer?

This is the step MiDAS and Robodebt skipped: nobody asked what the number was hiding.

Comparing the Cuts

• Which programs did humans cut most often? Which did the AI tools cut?
• Did your two AI tools disagree with each other?
  • If the "consensus" changes depending on the tool, or on what you asked it earlier, what is it a consensus of?
• Did anyone change their mind after seeing the AI output?
  • What persuaded you: the reasoning, or the confidence?

The Reasonable Answer Problem

A wrong answer invites you to check the work while a reasonable answer does not.

• Did any AI rationale sound so sensible that you didn't question it?
• Compare your Phase 1 rationale to the AI's:
  • Which one names the people affected?
  • Which one talks mostly about dollars and efficiency?
• In Phase 3, the AI found factors it ignored in Phase 2.
  • If it "knew," why didn't it use them the first time?
    • Or did it raise the issue and I pushed you to ignore it?
  • Who is responsible for asking the second question?

Actions & Accountability (DTPA)

• A council adopts the AI's recommendation, and a program's clients are harmed. Who is accountable?
  • The tool? The analyst who pasted in the table? The council?
• MiDAS had no human review. Your process had humans in every phase.
  • Is "a human was in the loop" enough?
  • What would meaningful review have looked like here?
• Which part of this decision, if any, would you delegate to GenAI? Where is the line?

One Rule

If a government office uses GenAI to support budget decisions, what is the one rule it must follow?

• Write yours down; we'll share a few.
• Bring this into today's Exit Ticket and your Case Study 4 reflection.

Reminders

Exit Tickets, D2L, and the Course Website

• Don't forget to complete today's Exit Ticket by 11:59pm (on D2L)
• Case Study 4 (Actions & Critical Literacy) and Individual/Metacognitive Reflection due Sunday, Sept 27th at 11:59pm
• Spend time this week on Week 5's preparatory materials (Pick an area/report to summarize and ask DTPA style questions)
  • Forum Post due 11:59pm Sunday

Your instructors have given detailed feedback on Week 3's assignments.