(Author here.) This is generally checked via calibration charts, e.g. bucketing markets at various points in time into 0-5%, 5-10%, etc; then counting how often the underlying events actually happen. The more they match, the more it's reasonable to interpret the market prices as probabilities.
Google published [1] one such calibration chart on its current prediction market in late 2021. Also, the 2009 paper in the article [2] on Google's first prediction market published one too.
[1] https://cloud.google.com/blog/topics/solutions-how-tos/desig... [2] https://static.googleusercontent.com/media/services.google.c...