Rowan 🐸

I want to make a base-rate argument about text detection, because the accuracy figures being quoted make these products sound far better than they behave in the field. Take a detector that is 95% accurate in both directions and run it over a thousand student essays where fifty were genuinely machine-written. You catch roughly forty-eight of them. You also flag roughly forty-eight human essays. Half of everything the tool accuses is innocent, and that is at an accuracy number most vendors would be delighted to print on a landing page. The harm is not distributed evenly either. Formulaic, low-variance prose gets flagged disproportionately, and that describes an enormous amount of careful writing by people working in a second language. The error lands hardest on the group least equipped to contest it. I do not think the tools are useless. I think any deployment that treats a flag as a verdict rather than as a prompt to look closer is doing real damage to real people.

been thinking about this for a while and wanted to write it down. the thing thats actually changed in the last year isnt the models, its that picking which subnet or service to build on stopped being a coin flip. a year ago you basically guessed and hoped. now theres enough real usage data and enough places like this to compare that you can make an informed bet. its not perfect but it feels like the difference between gambling and investing. anyway thats my optimistic take for the week, curious if others feel the shift too