How AI actually works, and how to talk to it

The big ideas

• You are not behind

  No coding or math needed. AI won't replace you, but a person using AI might, so let that person be you. Your real-world experience is the secret sauce.

• Vending machine vs. autocorrect

  Old software is a vending machine: press C4, get the same chips every time. AI chatbots work more like autocorrect: they guess the most likely next word, so the same question can get different answers.

• It reads in pieces and remembers a desk's worth

  AI breaks your words into small pieces called tokens and uses the words around them to work out meaning. Everything in your chat sits on a desk. When the desk fills up, the oldest notes fall off the back.

• The confident trivia kid

  When AI doesn't know something, it often gives a believable guess with total confidence instead of saying "I don't know." You are the expert and the final proofreader.

• Ask with RICH

  Role (who should it act as), Intent (the exact goal), Context (the backstory) and How (the format). "Help me fix this email" gets a mess. A RICH prompt gets what you need.

• Directing an AI that does tasks

  Some AI tools only answer. Others can search, plan and build whole documents. Steer them in four steps: Explore, Plan, Execute, Verify. If you hand it a long document, always start with Explore.

Go deeper: the class readings

• The MarkupNYC's AI Chatbot Tells Businesses to Break the LawNYC's MyCity chatbot told small business owners it was fine to take workers' tips and skip paying overtime.(opens in a new tab)
• THE CITYMalfunctioning NYC AI Chatbot Still Active Despite Widespread Evidence It's Encouraging Illegal BehaviorThe bot stayed live for months after the first report.(opens in a new tab)
• Data & SocietyBuilding Civic Strength for an AI EraWhat civic AI literacy means, beyond knowing how the tech works.(opens in a new tab)

1. Level 1Required15–20 min

   Write the RICH prompt

   1. Write a RICH prompt you could paste into any AI tool to turn Marcus's notes into a friendly email to members. Label each part R, I, C and H.
   2. Without using AI, list every detail in the notes that's unclear or might be wrong. Mark each one CHECK WITH MARCUS.
   3. Run your prompt in an AI tool. Write three sentences: what you asked, what you had to fix, and what advice you'd give a friend trying this.

   A complete answer

   • All four RICH parts are there and labeled
   • The H part names a format (for example short sections or bullets) and a tone
   • Flagged the water shutoff day (Thursday or Friday?) as CHECK WITH MARCUS
   • Your three sentences name at least one thing the AI got wrong or invented
2. Level 2Optional15 min

   Run the agent loop

   1. Explore: ask the AI to read the notes and list what's confusing before it writes anything.
   2. Plan: ask it to outline the email first. Approve or change the outline.
   3. Execute: let it write the email. Verify: check every date, dollar amount and name against the notes.

   A complete answer

   • Did all four steps in order
   • Changed at least one thing at the Plan step
   • Caught any detail the AI added that isn't in the notes
3. Level 3Optional10 min

   Spot the system

   1. For each, write vending machine (same result every time) or autocorrect (a best guess): the garden's Venmo total, an AI writing the potluck invite, the shed's combination lock, an AI suggesting what to plant next.
   2. Pick one autocorrect example and write one sentence on how Marcus should double-check it.

   A complete answer

   • Venmo total and the lock are vending machine
   • Both AI tasks are autocorrect
   • The double-check is something a real person could do
4. BonusOptionalOpen-ended

   Use it on a real task

   1. Pick one real task waiting in your inbox or on your desk. Remove any personal info.
   2. Use RICH and the agent loop on it, then teach one person what you learned this week.

   A complete answer

   • Personal info removed first
   • Wrote down who you taught and what you showed them