Anthropic says new measures cut data harvesting from Claude by competitors by 90-99%
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Labs are building defences against having their models copied through their answers.
Anthropic described three measures against distillation, the training of other models on Claude's answers: controlling access to reasoning (including barring new API accounts from editing earlier context), a two-stage classifier for extraction requests with a false-positive rate below 0.1%, and stricter limits, account bans and charges for blocked requests in cases of clear abuse. According to the company's preliminary data, the volume of data collected by distillers fell by 90-99% and the number of reasoning tokens extracted by 10 to 100 times or more. Anthropic notes it has not yet seen "a highly sophisticated and persistent attempt" to circumvent them and expects a "cat-and-mouse game"; some developers had to update their integrations, and false positives are rare. The figures are the company's own.
Sources 1