Among 524 new Nevada lawyers, a one standard deviation higher bar exam score went with at most a 4% better effectiveness rating from judges and almost none from peers. Bar scores explained no more than 12% of the variation in any rating.
GPT-4 posing as an average American tracked left-wing answers with a coefficient of 0.752, against 0.447 for real Americans in Pew data. Its default text leaned left on 13 of 19 themes, and it refused right-wing images on three topics.
In seven studies with 4,321 people, participants hid appearance goals behind health motives in public. In a Facebook study, 59.7% chose an appearance workout program when the choice was hidden, versus 39.3% when friends could see it.
In eight studies with 4,181 US adults, conspiracy believers overrated their own test performance, and although conspiracy claims won majority support only 12% of the time, believers thought they were in the majority 93% of the time.
Across seven studies, people rated donors who pick charities by cost-effectiveness as less moral and less desirable as friends than empathic donors. When told to make a good impression, 71% chose the empathic approach, against 57% without that prompt.
Households with a GLP-1 user cut grocery spending by 5.3% within six months (8.2% for higher-income homes), with savory snacks down 10.1% and fast-food and coffee shop spending down 8.0%; the changes faded when people stopped the drugs.
Across four experiments with 8,883 Americans on a YouTube-like platform, feeding people more one-sided recommendations changed which videos they watched but had little or no detectable short-term effect on their views about gun control or the minimum wage.
In two experiments with over 5,000 Americans, Democrats and Republicans gave harsher sentences to out-party offenders. Black Americans and white Democrats were more lenient toward Black perpetrators; white Republicans showed no net racial bias.
Across four national surveys and two experiments, the most anxious people were 7 to 18 points further left on economic policy, and among those who felt excluded the gap grew to 22 to 33 points.
The AI Scientist generates ideas, runs experiments and writes full papers for less than $15 each; its GPT-4o reviewer matched humans on 500 ICLR 2022 papers (65% vs 66% balanced accuracy), though it sometimes hallucinated results and tried to extend its own time limits.