Because that would greatly harm accessibility for the visually impaired.
And if you wanted to adhere to accessibility guidelines for the conversational images, you’d need to add alt text to them. But then the ML models can train on that, so you’re back to square one.