It's not delaying anything. It won't work outside of the paper.

If you draw fantasy cats, and you bias towards pointillism dogs, and someone else draws cubist cats, so they bias towards anime dogs, you dilute the effect of the biasing data as multiple axes are flipped.

And this assumes that all artists drawing cats agree on biasing towards dogs and not that some cat artists bias towards horses and others towards cows, which again dilutes any signal to just be noise.

It had a measurable effect in what were effectively artificial lab conditions to get the authors attention with a clickbaity pitch for the paper, but in the real world this is completely worthless right out of the gate.

No idea, but it's worth a shot, worst it can do is nothing.

If this takes off, it will be bypassed within a month. Adversarial training is something Stable Diffusion users already invented, and we use it to make our artwork better by poisoning the dataset to teach the network what a wrong result looks like. They reinvented our wheel.

It's fine, as outside of laboratory conditions it won't work anyways (diverse image "reverse labels" would erase their signal to noise ratio of biased pixels across aggregate real world training data), so no need to stress about any kind of reaction to it.

Wouldn't surprise me

Chuckles, "I like that, but I suspect they're using calling it poison because of the phrase 'poisoning the well'