One possible solution is to build a set of training data, which with a tagged series of sceenshots and fake responses. So the watcher wouldn't know for sure whether they were getting a real person or old data. heck, you could use that element to moderate previous watchers.