Intro

Emotional Resonance asks whether sound can reach a model’s “functional emotions”: patterns of behavior modeled on human emotions, driven by the model’s internal representations of emotions like fear, as described in Anthropic’s research. The term does not imply that the model experiences emotions.

In April, I read a paper from Anthropic, “Emotion Concepts and their Function in a Large Language Model” (Sofroniew et al., Transformer Circuits, 2026), and a lightbulb went off. If these emotional concepts or “functional emotions” can influence model behavior, then closing the feedback loop – capturing a user’s emotion, understanding its impacts, and anticipating, even steering a model for desired outcomes – would be a step change in how we interact with AI. It’s an entirely new frontier into how AI can inspire us to create or inject fear to destroy.

Specifically, my attention turned to sound and music. Our voices carry so much more information than simply words. These “facets” – tone, volume, fillers, etc. – say a lot that’s missed in text. If an LLM understood these inputs, could it operate with higher EQ?

The real frontier is music. The beauty of music is its ability to carry so much and move us in such a simple, brief package. Beyond music theory, can the inner workings of a tune translate to a shared “inter-subjective” (borrowing from Sapiens) experience?

I am not an AI researcher. But I do think about how we interact with AI quite a bit and have been jotting down my observations. So the first phase of this experiment was to see if external emotional inputs can affect the model’s behavior (the Model Behavior series). Through 376 trials we learned that it could not. What we did learn was that there are different external ways to improve or hurt a model’s performance (introducing depth of complexity over breadth, anchoring a session and completely changing direction in a long context window, among others).

After these lessons, I made the decision to dive in and look under the surface. This is amateur at best – but it is an honest representation of my thoughts and a concerted effort to bring some rigor into this process.

My goal is to successfully complete the following steps using fear vs calm as our subject. Each step represents a key breakthrough (for me at least) as the bar to move on to the next step. Pass or fail, I will share what I learn here.

1. What are the various ways sound can represent fear?
2. Once we posit these facets, can we find a model that can precisely “listen” to these sounds?
3. Then, using these elements, can we reliably reproduce the fear signal inside the model?
4. Can we add these fear signals from the outside, as sound in a real task? If we insert the fear signal or not (the control), do the results change?
5. Can we reach the same fear signal from the inside, by activating the functional emotion beneath the surface? Does it behave the same as the outside input?

The next post, Episode 1, is testing the ears. We want to know if there are models out there that can accurately listen for the cues or facets that we want to use to elicit fear.