AI Toys: small pages that break a misconception

Each one is a single self-contained page built to kill one specific wrong idea about AI and machine learning that slides reliably fail to kill. No install, no account, no network, no real model. Save the page and it still works. Companion to Security Toys.

Threshold Dial

Kills: “overall accuracy means it works”. The prettiest headline on the dial ships 29 of 50 faults, and a detector that detects nothing scores 95%; the threshold is a business decision about which error hurts more.

Ready

Data or Command?

Kills: “a document is just data”. Drag one hidden sentence into a document and watch the summary flip. Nothing on screen looked different.

Ready

K-Anonymity

Kills: “no names, so it’s anonymous”. The nameless table opens at k=1, and one row is uniquely the CFO. Coarsen the columns and watch them vanish into a crowd.

Ready

Next Token

Kills: “the model knows things”. Same prompt, two runs, two answers; Sydney was always sitting there at 24%. It weighs, it doesn’t know.

Ready

Overfit

Kills: “higher training accuracy is better”. Drag complexity up: train climbs to 100% while test falls to 73%. The model memorised the noise.

Ready

Drift Compounds

Kills: “95% per step is basically reliable”. Across a 12-step agent workflow it’s a coin flip, and the failures arrive polished. Checkpoints are placement, not diligence.

Ready

Jevons Dial

Kills: “more efficient means less used”. Make it 10× cheaper per task and the bill quadruples; the whole paradox hangs on one toggle.

Ready

K-Means Stepper

Kills: “the clusters are in the data”. Step the algorithm by hand; two starts, two confident truths, ten documents that change sides. A cluster is a choice until a human names it.

Ready

House rules