But we were only able to do this work because the training sets were available; for the most part, we don't know what the training sets for other models look like. Given this work using LAION as the example, hopefully others will be able to avoid mistakes of the past.