The results of two experiments support the hypothesis of "coder drift" which is defined as change that takes place while trained coders are using a system for a number of classroom observation sessions. The coding system used was a modification of the low-inference Flanders System of Interaction Analysis which calls for assigning observed teacher or student behaviors to ten relatively objective categories. The situations rated in these experiments consisted of videotapes of social studies classes which were conducted by experienced teachers. In each experiment, pairs of coders viewed the same series of ten videotapes, either working with their partners or working in separate rooms. In one condition, the pairs worked together for the first five sessions and apart for the other five. In the other condition, that order was reversed. Pairs working together for the first five sessions showed higher reliabilities than they did when they worked apart, and pairs working apart for the first five sessions showed even lower reliability. The same effect was found for five different methods of computing reliability. Randomly matched individuals showed no drift effects, nor did test-retest reliabilities between pairs working together and those working apart. (Author/CTM)