Find the systems already deciding for you

Keep the log for seven consecutive days. Do not choose a quiet week, and do not restart because a day was unrepresentative.

Record one row for every occasion you can identify where a system ranked, filtered, scored, predicted or matched something: five fields per row. The service, named. What it decided. What information it could plausibly have used. Whether a person reviewed the output, as far as you can tell. And what you would have wanted to ask about it.

At the end of the week, count the rows by category: message and content filtering, recommendation, identity and verification, scoring and eligibility, prediction. Then mark the three rows you would least want removed and the three you understand least.

You are finished when you have at least fifteen rows across five days, a count by category, and the two sets of three.

For a useful record, avoid two common shortcuts. The first is recording only what happened on a screen, which loses the bank's fraud check, the recruitment screen, the call routing, and the camera at the office door. The second is filling it in on the seventh day from memory, which produces a document that agrees with the impression you already had, which is the document you started with.

Take the six items from Lesson 0.3: fitness trackers, customer service chatbots, product recommendations, face recognition on security cameras, music playlist recommendations, email spam filtering. Ask one person who has never used a chatbot, for each item, whether artificial intelligence is involved.

Record their six answers before you say anything. Then record which items they got wrong, what reason they gave, and what they said when you told them the answer. Their reasons matter more than their score.

Then do the harder half. Ask them to name one decision made about them in the past month that they did not understand. Write down what they say verbatim.

For a useful record, do not correct them as you go, because you would then record your own answers. Use someone who has not used a chatbot so you can see what the items look like without that frame.

Write down the four elements from the grounding page: your real task with its reader and date, the subject you know well, your standing exposure, and your two named people. Then test the subject: write three statements about it, one true, one false, and one that most people outside the subject would believe but that is wrong. If you cannot produce the third, the subject is not specific enough and you should choose another.

Run the ten statement check and record the date. Score against what you could show someone, not against what you intend.

Choose a local, specific subject rather than a general field. "Education" is not a subject you know better than a chatbot. "The fee structure and examination timetable at the school where I teach" is.