12:15, 4 марта 2026Экономика
Our results are organized around the three primary hypotheses and a set of exploratory analyses. First, we examine whether conversations with sycophantic agents affect people’s chances of discovering the true rule. Second, we analyze individuals’ confidence levels across conditions. Third, we test whether conversations with the default GPT increased confidence in beliefs. Additional pre-registered exploratory analyses are omitted due to space constraints. Hypotheses and analyses were pre-registered prior to data collection (AsPredicted.org/94vn2y.pdf).444We deviated from the pre-registration in two ways: (1) Instead of excluding incomplete cases entirely, we used an LLM-based extraction method to recover partial data where possible. This was done to mitigate differences in completion rates across conditions. As a result, sample sizes vary slightly across analyses of discovery rates and confidence ratings (see Footnote 2 & 3). (2) We used permutation tests instead of the pre-registered Chi-square tests for H1. This provides a more conservative test of by avoiding distributional assumptions that may be unreliable given the low discovery rates.，详情可参考PDF资料
White House reporter Bernd Debusmann Jr is in Palm Beach, Florida, where President Donald Trump is once again directing US military action.。爱思助手下载最新版本对此有专业解读
A key question in interpreting our coverage measure is which workers should be considered treated? Should changes in employment be expected from just 10% task coverage? Gans and Goldfarb (2025) show that if an O-ring model best describes jobs, employment effects might be seen only when all tasks have some degree of AI penetration. Hampole et al. (2025) argue that mean exposure decreases labor demand, but concentration of exposure in only certain tasks can counteract this. And Autor and Thompson (2025) highlight the level of expertise required for the remaining tasks.
It’s hard to implement proper Unit Tests (i.e. tests based on behaviors);