the dog who barked too much
The dog barked at everything. That was the problem, and also, eventually, the point.
The idea was clean enough to be seductive. AI text runs lower entropy than human text, more predictable sequences, less information crammed into each token, and entropy is something you can measure. So you measure it. I am three sessions deep into it right now, which is the frame you need before I tell you the rest, since the rest is a man building a detector out of the exact material he once stood up in a lecture hall and refused to respect.
The first version was the zlib hack, and it was beautiful in the way that only wrong things get to be. Compress the text, divide compressed size by original size, and a lower ratio means more redundancy, machine. I ran it on the corpus and the numbers looked like a signal. Looked like one right up until I plotted them against document length, at which point the thing turned out to be a shadow. Compression ratio scales log-linear with word count at r = -0.967, which is the statistician's way of telling you your discovery has already explained itself and left no room for you to attach it to anything you cared about. The most AI-flagged post in the set was the longest thing I have ever written by hand. The least flagged was a short draft I built in an afternoon with heavy assistance, a machine practically holding the pen. LZMA gave me the same result with tidier decimals. And Huffman, may he rest in the particular discomfort reserved for professors whose classes turn out to matter later, gave me it again, the whole compression family barks at long documents and goes quiet at short ones and has, on the actual question of who wrote the page, nothing to say.
So I went to Shannon proper. Entropy per sentence, per paragraph, Miller-Madow correction for the small-sample bias, and the correction term at sentence level came in around +0.4 bits, larger than the signal it was fixing. Fine. Sliding windows to wash out the length dependence, and at a hundred words the entropy just converges on the English rate, roughly 6.3 to 6.5 bits a word. The window keeps finding the same value in everything, since it is measuring English, not me, not any model doing an impression of me. It sees the language and reads the tongue back into the page.
A compression ratio is a real number that correlates with a real property, and that property is redundancy, and none of it is authorship. The mistake was optimistic, technically legible, fatal, the qualities that make a mistake worth an afternoon and then three more of them. I had contempt for these algorithms once, loud, on record, in a room. Huffman did not need forgiving. He needed to be right just long enough to be annoying, and he has been, for three sessions, which is longer than I gave him credit for. Maybe that's the part I should have measured first.
The dog was never lying about the noise. It heard exactly what was there. What I am not sure of is whether the wrong thing was the tool, or the approach, or the ears I brought to it, or whether all three were the same mistake wearing different collars. The dog is still barking. I have not gone to the window.