CLIP is trained to recognise images using a mountain of raw data drawn from the internet, where people routinely upload images and identify them with captions.