Signal Training data doubling every six months at the cost of quality control

Summary

As AI models grow larger, the problem of data quality becomes more serious. The data sets used to train them are doubling in size every six months. Gathering that much material can mean loosening quality checks, which raises the chance that models learn from fake or dubious data. EY therefore warns of a paradox in which the newest and most capable models could turn out to be less reliable and less trustworthy. This adds a technical dimension to the wider trust deficit in AI that EY describes alongside consumer and social concerns.