GPT-4 as a scaling instrument: it matched human emotion maps, then went past them

Ke Han and Eiji Watanabe, of the Laboratory of Neurophysiology at Japan's National Institute for Basic Biology, asked GPT-4 to do something a psychology undergraduate could do: place emotion words on a grid so that similar meanings sit close together, then compared the layouts with human judgments ^[1][2]. Prompting the model in ordinary language reproduced a structure closely resembling the human arrangements, and when the vocabulary was expanded to 99 terms a structure related to arousal emerged that had been hard to detect with only a handful of words ^[3][4]. The work is published in Scientific Reports ^[19].

The task design is the load-bearing part. Presenting three terms at once, for example joy, surprise and sadness, expresses three pairwise relationships in a single trial ^[5]. That matters because human effort is the binding constraint: the number of comparisons rises rapidly as more emotion terms are added, and asking participants to cover all pairs among 99 terms is impractical ^[6]. Six terms is 15 pairs, five triad trials; 99 terms is 4851 pairs, at least 1617 trials before repetition ^[7].

In the first study, 89 human participants and GPT-4 performed the same task on six basic emotion terms, with conventional word embeddings, which represent meanings as numerical vectors, as a third comparison ^[8]. Humans and GPT-4 produced the same two groups: joy with surprise, and anger, fear, disgust and sadness together ^[9]. The embedding analysis put joy and anger in the same group, a structure unlike the human one ^[10]. Distances from humans and from GPT-4 were strongly correlated, and human responses were consistent across participants, which is what makes the human side usable as a reference at all ^[11][12].

That split is the most useful result here. Prompting a language model and reading off a static vector space are not interchangeable measurements of the same thing, and on this task the prompted model tracked people while the vectors did not ^[9][10].

At 99 terms, both GPT-4 prompting and word embeddings divided the vocabulary primarily into two broad clusters corresponding to pleasure-displeasure and dominance, with the separation between clusters clearer in the GPT-4 layouts ^[13][14]. A finer analysis then detected variation corresponding to arousal, rather than arousal acting as a principal axis ^[15]. Arousal is the activation dimension in the standard psychological triad of pleasure, arousal and dominance ^[16]: excitement and serenity are both fairly pleasant and differ mainly in arousal, as do rage and sadness on the unpleasant side ^[17].

The authors' caveat is the correct one, and they state it plainly: this does not mean GPT-4 experiences emotion, and they read the model as drawing on relationships among emotion concepts already embedded in human language ^[18]. The awkward consequence follows from their own method. The six-term structure is validated against people ^[11]. The 99-term arousal dimension is not, because the whole reason to run the model at that scale is that the human experiment cannot be run there ^[6]. A dimension that appears only where ground truth is unavailable is a hypothesis about the corpus until someone tests it.

What to watch: whether anyone validates the 99-term geometry piecewise, running human triads on sampled subsets of the larger vocabulary and checking that arousal survives; whether the human-like grouping of the six basic terms replicates across models, prompt phrasings and languages; and whether the divergence from static embeddings holds up, since that is the only evidence so far that prompting adds anything beyond distributional statistics ^[9][10].