Thread · 3 tweets · 4 Feb 2022

In both cases you are designing a token to be miscategorised by an automated system. With the penny, the mechanism has properties (determinism, simplicity, transparency) such that we can accept it as a proxy for purchase from a human - which is then breached.

For the algorithm, well it depends on the algorithm (the cigarette box being one too). If it's explicitly string matching to filter only for Harvard people, then that's like direct résumé screening - so more like lying. If it's tricking an ML model and they don't know what...

...they're looking for, then it's a lot less clear. Can you lie to someone who doesn't know what question they're asking? (Disclaimer: I once got a job by passing a coding test with flying colours because I figured out how the answers were encoded in the page ;)