> It's probably old news for RISKS readers, but a very difficult concept for > lawmakers, that the semantic content of bit streams is in the eye of the > behold

> It's probably old news for RISKS readers, but a very difficult concept for > lawmakers, that the semantic content of bit streams is in the eye of the > beholder, and that the apparent correspondence between bits and semantics is > the result of engineering convention and not an inherent property of the > bits. Nice article with great points. I gave a talk about this in 2005 on why the more data intelligence agencies collect the worse their results if their analysis does not match their reach. It goes back to Quine and Shannon's ideas of salience as pre-agreed patterns of interpretation. The talk was actually about a spooky phenomenon called "listening in readiness". Mediums/charlatans and other cold-reading hucksters used EVP (electronic voice phenomena) in the 1930's and 40s, when radio, Theramins and such-like were more woo-woo and barely understood by ordinary people. If you play what is essentially noise/static to people and _tell_ (or suggest) to them that voices are saying something - they will hear that. The phenomenon is surprisingly reproducible. It works because the cochlea and auditory neural system (See a text like Nelken, King and Schupp's Auditory Neuroscience) can "listen in readiness". We have affective and sensory hairs and feedback loops in the cochlea that allow us to "tune" to what we _expect_ to hear. In simple words, people can hear what they expect/want to hear. When we apply AI and adaptive filters to data, a similar thing is happening. False positives, indeed very elaborate misinformation can be derived in intelligence work based on unsupervised (arbitrary mass surveillance) when the gatherer starts with an a-priori idea of what they are looking for and sifts through chatter. I am not sure exactly how yet, but I think this can be leveraged to some good use in privacy protection if, as in the Dissident scheme discussed in TFA, there is some "fuzzy" decryption and very many plausible but false decodings adjacent to cipher-text. This leads to the idea of a cupher that is apparently very easy to crack, but yields a false plaintext. When you "can't remember" the password, your adversary finds a low hanging "trap password" and smugly thinks they defeated your poor opsec.