Newer artificial intelligence models missed more fraud in Coinbase testing

Coinbase reported on October 7 that newer versions of three major artificial intelligence model families caught fewer fraudulent payments in a screening test for its Onramp service, according to CryptoSlate. The exchange replayed 16,140 transactions across historical data while keeping its decision policy unchanged. The findings challenged the assumption that upgrading an artificial intelligence model automatically improves an existing payment screener. [1]

CryptoSlate reported that the fixed historical replay showed weaker fraud coverage across three model upgrades, including newer Opus, Sonnet, and GPT models. The updated models missed a larger share of total fraud value compared to their predecessors. However, the precision of the GPT model improved during the benchmark evaluation. [1]

Coinbase is a publicly traded cryptocurrency exchange that went public on the NASDAQ in April 2021, Grokipedia notes. The artificial intelligence models tested included Sonnet and Opus, which are part of the Claude family of large language models developed by Anthropic, according to Wikipedia. [3][5]

Coinbase concluded that artificial intelligence providers need task specific upgrade tests before deploying new models into financial production environments, CryptoSlate reported. The company stated that standard model updates can inadvertently reduce the effectiveness of fraud detection systems. [1]