4 ms·
EU top court's ruling spells trouble for scoring algorithms
- jruohonen 3y agoWhile this case was already pointed out, the implications deserve a re-emphasis: "Between 2018 and 2021 a scandal took hold in the Netherlands – eventually leading to the resignation of Mark Rutte’s third government – due to a flawed risk-scoring algorithm which led tax authorities to wrongly accuse thousands of frauding a childcare benefit scheme." Cf.: https://news.ycombinator.com/item?id=38555378 https://news.ycombinator.com/item?id=38555378 The case hinted is similar to the (in)famous Dutch SyRI system, which too has been pointed out on several occasions: https://news.ycombinator.com/item?id=37987367 https://news.ycombinator.com/item?id=37987367 https://news.ycombinator.com/item?id=37924712 https://news.ycombinator.com/item?id=37924712
- jruohonen 3y agoIn France they've at least published the data types they're using for scoring people who need benefits: https://www.documentcloud.org/documents/24179765-les-cibles-_-variables-2014-datamining-caf-le-monde-likelihood-par-parametre https://www.documentcloud.org/documents/24179765-les-cibles-... Yet, publishing data types is not enough; to gain even a glimpse of the actual scoring and its robustness, you need also information about the algorithm in use, its performance metrics, error statistics, distributions, missing values (which are apparently too present), etc. All these should be available through a FOI (and if not, a request should be brought to a court). Given how much even peer reviewed ML research is flawed, the stakes are high for an administration to demonstrate robustness (or a claimant to demonstrate otherwise). The lawfulness/lawlessness they discuss in this piece is an entirely different question.