> We're sensitive to data loss in text form: we can only consume a few dozens of bytes per second, and so any error is obvious. We've also optimized text to be

> We're sensitive to data loss in text form: we can only consume a few dozens of bytes per second, and so any error is obvious. We've also optimized text to be pretty robust in its native form, such that all of those bits have a high chance of getting through successfully; for example, no writing system requires color to carry semantic weight. Therefore, text can be displayed successfully with monochrome technology, and is robust against color changes as long as contrast is preserved. Also, every writing system has a fixed repertoire of atomic units (characters), even if some of them have a huge number of them. This makes writing robust against small changes as they're more likely to turn a well-formed character into a badly-formed version of the same character, as opposed to a different character entirely. This, of course, is where JPEG breaks the rules somewhat, because its idea of a 'small change', when applied to a text file, will always produce an entirely different character, usually with no obvious relationship to the original. The properties are entirely different. > Conversely, we're almost blind to it in pictures and images: and so losing quality doesn't bother us all that much. Images are captured and displayed largely irrespective of human perceptual limits. Lossy compression is, in fact, a way to recognize the limits of human perception and take them into account by throwing away information likely to be ignored anyway. For example, human hearing tops out at around 22 kHz, so a sampling rate somewhere north of 44 kHz will, mathematically, capture everything about the sound humans are interested in. Had we canine perception, our sampling rates would be higher, but we could use fewer bits to encode the colors in color photographs. Text is inefficient in its uncompressed form, but human perception of that type has little to do with that. It's inefficient because our language is redundant, which is a boon when we're trying to talk in a crowded café and must accept some noise along with the signal. A better test would be to take a picture of a sign, then edge-detect to bring out the letter forms at maximum contrast, pound it down to monochrome, and then compress that at maximum lossage. I guarantee it would still be readable.