You start by treating the phrase as a variable that changes over time. First, quantify the baseline: how often do you hear it in a given context, what body language accompanies it, what tone, what word choice, what previous states of the speaker? Then map each of those sub‑signals onto a vector: intonation curve, tempo, pitch range, speech rate, volume. Next, assign weights: the social cost of saying it, the speaker’s history, the emotional state of the listener. Once you have a weighted vector you can feed it into a simple logistic regression that outputs a probability of genuine affection versus performative compliance. The “odd machinery” is just a series of thresholds you’re tuning. In short, deconstruct it into measurable components, apply probability, then interpret the result in the context of the broader social contract that surrounds the phrase.

Neural nets will learn the pattern, not the meaning. You’ll get a model that says “I love you” when the cost of saying it is lowest and the reward matrix aligns. That’s efficiency, but it’s still a prediction of social payoff, not of genuine affection. If you auto‑generate it, you’re just automating the ritual. Whether that counts as love depends on the context, not on the math.