that's a gross oversimplification and an argument that was made in bad faith that has now spread like a meme.
The gist was if you overtrain a model, and try to recreate an exact image by prompting very similar things, and run it several thousand times, then you can recreate an image. If you seriously crank up the overtraining it's even easier. But normal use of the models do not just pump out recreations of training data.