https://josephausterweil.github.io/probintro/ethics/index.htmlThe Mascot’s Code. At year’s end, Chibany audits what the campus systems actually learned — and writes a code of conduct for machines that learn from people. Can the kiosk be fooled with a sticker? Is the cafeteria recommender fair to every dorm — and what would “fair” even mean, stated as a conditional probability? Whose priors did the models inherit from their data? And when we trained the robot trainee with rewards, what did we actually teach it?Hugoen-usThu, 16 Jul 2026 00:00:00 +0000https://josephausterweil.github.io/probintro/ethics/adversarial-examples/index.htmlThu, 16 Jul 2026 00:00:00 +0000https://josephausterweil.github.io/probintro/ethics/adversarial-examples/index.htmlThe Sticker on the Bento The kiosk in Cafeteria A has been faultless for months, so at first nobody believes Chibany. A student holds up the lunch bento — plainly tonkatsu, golden cutlet, the label even says so — and the kiosk chirps hamburger. Not “maybe hamburger.” The screen prints a confidence: 97% hamburger. Chibany does what Chibany always does: investigates. The bento goes on the scale — 500 grams, tonkatsu’s weight. A chopstick tap — crunch 8, tonkatsu’s crunch. Chibany turns the box in the light, sniffs it, compares it against yesterday’s photo in the Bento Journal. Nothing a human would notice has changed. Then the trainee robot, peering very closely, points a manipulator at a corner of the lid: a small, cheap sticker, the kind that comes on a sheet of a hundred. A prankster from the robotics club, it turns out, printed it on purpose.https://josephausterweil.github.io/probintro/ethics/fairness-formalisms/index.htmlThu, 16 Jul 2026 00:00:00 +0000https://josephausterweil.github.io/probintro/ethics/fairness-formalisms/index.htmlWhich Fairness Does Chibany Mean? The campus bento-shop recommender has been running all semester. Every student who opens it gets a short list of shops to try tonight, ranked. This week the complaints arrived: Sakura dorm says the recommender keeps steering them to the same three tired shops, while Kaede dorm gets the good new ones. Kaede dorm says the opposite. Both feel cheated. Chibany, who has been keeping score in the Bento Journal all year, pads over to the grad-student office with the single question a mascot always asks: is it fair?https://josephausterweil.github.io/probintro/ethics/bias-in-data/index.htmlThu, 16 Jul 2026 00:00:00 +0000https://josephausterweil.github.io/probintro/ethics/bias-in-data/index.htmlThe Autocomplete That Knew Too Much It is the last week of the school year, and Chibany is doing what the mascot does at year’s end: auditing. The campus systems have had a full year to learn — the bento kiosk that sees, the chatbot that talks — and Chibany, notebook open, is checking what they actually learned. The chatbot is first. Chibany types half a sentence and lets the model finish it.https://josephausterweil.github.io/probintro/ethics/alignment-safety/index.htmlThu, 16 Jul 2026 00:00:00 +0000https://josephausterweil.github.io/probintro/ethics/alignment-safety/index.htmlThe Audit’s Last Line Three questions down. Chibany has poked the bento kiosk with a sticker and watched it call a hamburger tonkatsu; has asked whether the cafeteria recommender is fair to every dorm and learned it cannot be fair in every sense at once; has traced whose associations the campus systems soaked up from their training data. The Bento Journal’s audit page has one line left, and it is about the robot mascot trainee — the one the SDS students spent all spring teaching.