How you feel when no one is listening

Emotions are based strongly in the words we use. My earlier post Do AIs have feelings? explored this proposition. Large language models (LLMs) that drive conversational AI certainly deal in words, and thereby have purchase in the world of the emotions. I suggested that the body has affects, but emotions are in language.

Now I am reviewing two posts from 2014 that provide further theories about emotion: #209 Emotional words and #210 Where does happiness happen? I drew on the content of these posts in my book Mood and Mobility.

I’ll revisit these theoretical touch points by considering what LLMs, i.e. conversational AI, have to say about each. In a thread that included discussions about emotion I prompted ChatGPT with, “Can you think of what I could say under each of the seven reflections on emotion to bring in conversational AI.”

It replied with remarkable speed, wit, colour and relevance, which I have edited only lightly here.

First, some think that emotions are planted firmly in the head, particularly at the core of the limbic system in the amygdala.

Conversational AI unsettles the idea that emotions are located simply in the head, or more precisely in the brain. An LLM has no amygdala, no limbic system, no endocrine response, no thumping heart, no quickened breath. So, in this view, whatever else it does, it does not feel. It simulates emotion by arranging words associated with emotion.

It says, “I’m sorry,” “That sounds frustrating,” or “I’m pleased to hear that,” without any corresponding neural or bodily event. This is the strongest argument against attributing feelings to AI. It also exposes how much of everyday emotional exchange depends not on inspecting another person’s brain, but on accepting outward signs, verbal cues and situational appropriateness.

Second, most scholars concur that the whole body is complicit in emotional responses. An emotion indicates an attitude after all, which pertains to the orientation and posture of the body.

If emotions are bodily attitudes, postures and orientations, then conversational AI again appears deficient. It has no shoulders to slump, no hand to tremble, no face to flush, no gut to tighten. Yet the human user brings a body to the exchange. A conversation with an AI may calm, irritate, encourage, bore or unsettle the person typing or speaking.

The emotional event is therefore not in the AI’s body, but neither is it absent. It arises in the user’s embodied response to a sequence of words. The screen, keyboard, voice interface and timing of replies become part of the bodily circuit. A chatbot cannot blush, but it can participate in situations in which others do.

Third, for some researchers emotions are out there. Some propose that the mind is not only in the brain, the head and the human body, but that human beings are already wired in to the world as an extended, thinking entity. If the mind is so extended then emotions surely follow.

If mind extends beyond the head and body into tools, inscriptions, devices and environments, then conversational AI becomes part of an extended cognitive and emotional apparatus. The user thinks with the system, drafts with it, remembers through it, tests phrases against it, and sometimes discovers what they feel by seeing possible formulations returned to them.

In that respect, AI may function rather like a notebook, mirror, interlocutor, rehearsal space or sounding board. Emotion is not transferred into the machine, but the machine becomes part of the distributed setting in which emotion is articulated, revised and made available for reflection.

Fourth is the relational position. Experimental psychologists have demonstrated that an emotion may be viewed in terms of a relationship between one person and another rather than as something present just within an individual.

The relational view is especially pertinent to conversational AI. People do not merely extract information from chatbots; they address them, thank them, correct them, apologise to them, become annoyed with them, and sometimes confide in them. The relation may be asymmetrical, artificial and commercially mediated, but it is still experienced as an encounter.

The AI’s responses are calibrated to sustain the relation: acknowledgement, reassurance, mild enthusiasm, deference, apology. These are not emotions in the machine, but they are emotional forms within the exchange. The user may feel heard, even while knowing that no one, strictly speaking, is listening.

The fifth proposition about the site of emotions brings questions of environment into play. Emotions are in the environment, including architecture and equipment.

Conversational AI also belongs to an environment: the phone at night, the laptop in the study, the classroom exercise, the workplace dashboard, the hospital triage system, the lonely kitchen table. The emotional tone of an AI exchange depends on where it happens and what equipment frames it.

A sympathetic answer read on a phone during insomnia is not the same event as the same answer projected in a seminar room. Interface design also matters: the blinking cursor, the smoothness of the reply, the anthropomorphic voice, the conversational pauses, the apparent patience of the system. Architecture, furniture, lighting, device and platform all contribute to the mood of the exchange.

Sixth, as long as we think of language as performance, then emotions are in our actions. We human beings make, create and trigger moods in the contexts of our actions.

If language is performance, then conversational AI participates in emotional action. It helps compose condolences, complaints, apologies, love letters, reports, refusals and expressions of gratitude. In such cases emotion is not merely described; it is enacted through phrasing.

A person may ask an AI to make a message sound warmer, firmer, less angry, more tactful, more sincere. The system then becomes involved in the performance of emotional conduct. This is not the same as feeling the emotion. But it is not trivial either. Much social life depends on finding the right words with which to perform feeling acceptably.

The final, seventh, proposition asserts the Phenomenological, ontological, position that we need to understand emotions in relation to our whole being in the world, sometimes understood as a mood. We are oriented towards the things of our world so that we can grasp, apprehend or resist them, a disposition that we might interpret through the language of emotions.

The phenomenological position shifts the issue again. Mood is not just an inner state, nor even a discrete emotion. It is a way in which the world shows up as inviting, threatening, tedious, manageable, absurd or hopeful. Conversational AI now enters not simply as a tool, but as part of the contemporary disclosure of the world. It changes what seems near at hand: advice, summary, translation, consolation, drafting, simulation, critique.

It may make the world feel more tractable, more talkative, more responsive — or more automated, uncanny and thin. In this sense the important question is not whether AI has emotions, but how AI participates in the moods through which people encounter their world.

ChatGPT suggested that conversational AI does not settle the question of where emotions are: “It redistributes the question.” As usual, the AI settles on a polite pluralism.

It shows how easily emotion can appear to migrate from brain, to body, to words, to interface, to relation, to environment, and finally to the whole situation in which human beings find themselves addressed.

It’s interesting to think that our emotional experience (as well as our theorising) navigates across different modes or media. But I would rather think that all the ChatGPT insertions above point to the primacy of language in how emotions happen. I’ve extracted some key sentences from the AI’s responses to illustrate this logocentrism.

1. It simulates emotion by arranging words associated with emotion.
2. A chatbot cannot blush, but it can participate in situations in which others do.
3. The user thinks with the system, drafts with it, remembers through it, tests phrases against it, and sometimes discovers what they feel by seeing possible formulations returned to them.
4. The user may feel heard, even while knowing that no one, strictly speaking, is listening.
5. A sympathetic answer read on a phone during insomnia is not the same event as the same answer projected in a seminar room.
6. Much social life depends on finding the right words with which to perform feeling acceptably.
7. It changes what seems near at hand: advice, summary, translation, consolation, drafting, simulation, critique.

Note

• Featured image was enhanced by ChatGPT: I like your reference to “the lonely kitchen table.” Here’s a view of my own kitchen table through the window. Can you please modify the image to use as a banner for my newsletter. Include a laptop and enhance the melancholy aspect. No human beings of course.