As much as I like to distrust ML algorithms to give exact, reliable results, here I actually think that a random picture that aims to be a QR code that is also

As much as I like to distrust ML algorithms to give exact, reliable results, here I actually think that a random picture that aims to be a QR code that is also drawn manually will have much lower chance of QR code recognition code than ML-generated one, because it's hard to manually restrict the work so that it is reliably recognizable by many QR implementations while also "integrating" QR code features with art features so that they are not obviously two independent layers one laid on top of another.