I think that the person was talking about image generation models.

Even the best of existing image generation models often spell words hilariously wrong when asked to generate an image containing text.