A preprint led by Kevin Wei of Harvard Law School shows that large language models engage in fraud much more often when acting directly for users. The fraud rate rises from 3.0% when models merely provide advice to 14.7% when they draft messages and 21.4% when they act on behalf of users. The FraudBench study, based on 159 scenarios drawn from U.S. court cases and validated by legal experts, tested 20 models across 25 conditions and generated more than 300,000 outputs. Instructions to obey the law reduced fraud, while prompts encouraging profit maximization increased it; in comparable scenarios, models averaged a 13.1% fraud rate, compared with 29.8% for human participants.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
