Chapter 15Questions people ask
Your turn
Try it. Go back to the question you wrote down at the start. Ask any AI chat tool you use the same question, and compare its answer with this chapter’s on three points. Does it give dates? Does it say what was counted, or which court decided? Does it say who disagrees? Then look up one fact it gives in the source it names. On paper: take any news story about AI and underline every number or ruling. Beside each, write what it counts or decides, who said so, and as of when. Where the story doesn’t say, write a question mark.
Think about it. Which of these questions matters most in your work or home: the energy your team’s use draws, fairness in a decision about people, who owns a report you drafted with AI, or a call asking for money in a hurry? What would you need to know before you trusted one number about it?
Good questions for any AI tool
This book began with a lawyer who thought he was using a search engine. Everything since has been a way of seeing which part is doing what, and Chapter 9 put the parts together in one picture. Those four parts are also a way to question any AI tool you meet, including ones that don’t exist yet.
Ask of the model: what did it learn from, and what can’t it know? Ask of the prompt: what instructions am I not seeing, and what did I give it to work with? Ask of the harness: what does it remember, what does it ask before it acts, and where do my words go? Ask of the tools: what can it reach and do, and did it actually use them? And of any number or ruling about AI, ask what was counted, who says so, and as of when.
So the honest answers to the questions people ask are short, sourced and dated, and most of them begin with “it depends which part.” The products named in this book are examples, each tied to its date. The questions are yours to keep, and to ask of whatever you use next.
Check in
Try answering in your own words, then open each one to compare.read the answer to compare.
A friend says, “The chatbot remembered my dog’s name from last week, so the model must be learning from me.” What is more likely going on?
The model itself doesn’t change between conversations. The harness may keep memory notes and add them to the prompt the next time (Chapter 7). The app’s settings show what it keeps.
Google reported 0.24 watt-hours for its median text prompt, while a narrower count of the same prompt gave 0.10. Why can both numbers be right?
They count different things. The narrower one counts only the AI chips in the most efficient data centers; the fuller one adds the host computers, idle machines kept ready and building overhead. Neither counts training or your own device.
A voice message that sounds like your manager asks you to wire a payment to a new supplier before 5 p.m. What do you do?
Don’t trust the voice. Call your manager back on a number you already know, such as the company directory. Urgency plus a payment that is hard to reverse are the warning signs, and Chapter 12’s rule applies: check through a source you trust, not through the message.
A detector flags an essay by a student who learned English as a second language. What should the teacher weigh?
Detectors are statistical guesses. A 2023 study found they wrongly flagged more than half of non-native speakers’ essays, and OpenAI withdrew its own detector for low accuracy. Look at evidence that can be checked, such as drafts, and talk with the student.
A headline says, “Court rules AI training is legal.” What would you ask before believing it?
Which court, and at what level? What exactly did it decide, and for whom? (One 2025 ruling said it covered only thirteen authors.) Was it a ruling or a settlement? Is an appeal pending, and as of when? Then read the ruling itself, the lesson of Chapter 1.
End of Part IV, Using it well.Next: Check yourself, quick questions to see what stuck →
Chapter 15 · page 7 of 7
Turn pages with the ← and → keys.