skeptical reaction with one expression of support: https://statmodeling.stat.columbia.edu/2023/05/31/jurassic-ai-extinction/
and generally “beware the one of just one study”
Given that in more than a third of the cases where GPT and the answer set disagreed you thought GPT was right and the answer set was wrong, did you check for cases where GPT and the answer set agreed on an answer you thought was wrong?
“Finally, we can test our entire pipeline by deliberately training misaligned models, and confirming that our techniques detect the worst kinds of misalignments (adversarial testing).”
In 26 models taken from volumes 21 to 25 of the journal Law and Human Behavior, the highest R-squared -proportion of VARIANCE, not variation, explained was 40% and the second highest 24%
Is martingale different from conservation of expected evidence?
This seems to have stopped in July 2022.
In a “heatplot” or plots cf https://www.elsblog.org/the_empirical_legal_studi/2023/05/heatplots-for-correlation-coefficients-graphs.html
You could also study the distribution of correlation strengths found over the range of correlations tested, possible, seeing how it compares to what would be expected by chance.