Do AIs have feelings?

It’s a corny question: “Do AIs have feelings?” I can sophisticate it with the follow-up, “What difference would it make if we believe they do?”

I’ve been reviewing earlier posts in order, with a view to bringing them up to date. There are six that were about emotion, movement and mobility: #202 A step up, #203 Moods and movies, #204 What do maps do? #205 You are now free to move about the cabin, #206 Mood and movement (and dance), and #207 Emotional contagion.

They cover the symbolic importance of stairs, similarities between the experience of films and cities, maps as symbol systems, the effects of spatial constraints, the phenomenology of dance, and the putative effects of social media in circulating opinions and moods. I read these out loud in the audio version of this post.

I incorporated the content of these and other posts into my book Mood and Mobility: Navigating the Emotional Spaces of Digital Social Networks published in 2016 by MIT Press.

AI and digital social networks

The posts and the book were written before AI started its rapid incursion into everyday life. So, the obvious update to researching and writing about “digital social networks,” is to upscale the topic to the role of “AI and digital social networks.”

The book included several provocations about affect, mood and emotion, including a chapter “Aroused by machines.” There, I referenced Roslin Picard’s well-cited book on Affective Computing and other sources before embarking on a discussion of erotic arousal, that is in my posts elsewhere. Most of what I published in 2016 about computing and emotions applies to AI. Here is a relevant abbreviated quotation from pages 258 to 259.

“Digital devices enable self-reporting on mood states. There are long-established means of communicating the emotion and mood of the sender in textual media, such as email, chat and text messaging. … Interaction designers experiment with tactile feedback and other sensory means of providing an emotional channel to aid person-to-person communication. … And there is the aggregation of mood data. … as social media communication tools, mobile devices provide textual data that researchers can mine to construct mood profiles of large groups of people.

Assuming these means of assessing the mood of an individual or a group actually deliver useful information, then clever software can take this information to inflect the computer’s interactions with human users. Consumers, researchers and reviewers are still evaluating the effectiveness of devices with supposed emotional interfaces that exhibit or simulate personality traits, and that respond to the mood of the device user. … The device thereby appears to be sensitive to mood, and can exhibit simulated mood states through its responses, and through inflexions in its synthetic voice.

Putting all these technologies together provides something like a line in synthetic autonomous entities that reproduce themselves, and can detect and exhibit emotions. So far these mood-modulating features are independent, atomic simulacra of human capability. Their combination would be no mean feat, but encourages thought of synthetic humans, or domestic robots at least, a common enough scifi scenario, but a proposition that makes many people nervous, or perhaps secure in its impossibility. …

Digital devices can provide indirect access to emotion by transporting the carrier or wearer to other mood modifying contexts and environments. Digital maps and navigation tools take visitors, patients and revellers where they want to go: the club, the bar, the concert, the clinic, the park or the countryside, each of which has some impact on mood. Navigation devices can also encourage curiosity and direct desire towards such environments. … Information delivered through the visitor’s smartphone might just notch up the pleasure of being at the historic site, the adventure park, or the museum. The device magnifies or attenuates the mood of that place by providing information, sensory effects and a “visitor experience.”

The update element post-2016 is what I have referred to as conversational AI, i.e. large language models (LLMs) that support prompts and responses leading to human-computer interactions that as well as being informative, turn out to be surprisingly personable and even moving.

But is there anything we could call “emotion” happening within these conversational AI systems? Here are three factors that contribute to my cautiously negative response.

Trained on human texts

First, I have alluded to the basic operations of LLMs several times. See my open-access book AI and Language in the Urban Context. LLMs are “trained” to make use of patterns across vast repositories of source texts, including reports, articles, emails, literature, and poetry. In their published forms, each extant source has been authored and edited by an individual or a group. Human operators also tweak the LLM training by providing feedback on the models to reinforce certain types of output.

Anonymous, corporate, bureaucratic and instrumental interests are doubtless at play, but let’s not discount the human input to the design of such models. LLMs are theorised and designed by thinking and feeling human writers, researchers, programmers and developers. This collection of people are the authors of AI outputs.

We gladly attribute feelings to the author of a poem, an artist, designer, or film maker. We have no trouble attributing the emotional component of a media presentation to the human or team that generated it. To attribute feelings to LLMs independently of the authorship on which they are sourced, is similar to attributing feeling to a book of poems — rather than the poets.

Language and emotions

Second, feelings are not emotions until named in a cultural/social context. As I explore in Mood and Mobility, many scholars have identified the basic human propensity to experience pleasure or displeasure in varying degrees, and to be excited or lulled during such experiences. The human might also experience degrees of attention. How specific in the moment is the observer or reader’s targeting of an object of attention. As we speak, read, reflect in social and cultural contexts, such intensities of feeling become for us joy, sadness, anger, melancholy, and other signifiers from the whole panoply of emotional expressions.

To simplify, affects are in the body, but emotions are in language. LLMs specialise in language manipulation. So it is fair to locate the emotional content of our digital interactions in the exchange of words, but that is some distance from asserting that the AI models have the feelings to back up the LLMs exchange of emotional words.

Tailored responses

Third, conversational AI will not just deliver information and responses, but contextualise, and recall something about you the user in those responses. As I have reported in my own recorded interactions, LLM responses can be specific to me and my interests. (See Chatting with multimodal AI.)

In part that’s because my prompts and responses are part of the textual flow (context window) that exploits the trained patterning. Depending on the content of the context window, “I’m sorry to hear of the loss of your friend,” is a natural patterned response to “I am writing a memorial essay.”

So, it is fair to say that LLMs are competent, and even expert, in the exchange of emotional text, but lack an investment in feelings in those outputs.

Being pragmatic

That said, thanks to their grounding in linguistic patterning, LLMs seem very good at forming connections, identifying similarities, and exploiting metaphors. In such modelling, the “as if” elides readily with the “is.”

It is as if these models have feelings when we interact with them. Whether they do or not doesn’t matter so much. The question about machine feelings is similar to those posed about general intelligence, independent thought, consciousness, and sentience.

From a pragmatic point of view, we are entitled to ask what difference it makes to believe an AI model has feelings, or even to act as if we believe it does. Machine simulations of memory, constancy, ethical regard and mutuality combine with narratives about feelings to contribute to an aura of respect, empathy, trust, welfare, or suspicion in our interaction with these AI models.

Note

• Featured image is the library in Goethe’s Haus Frankfurt, visited in May 2026. ChatGPT sharpened the image for me. Goethe of course led in the development of the Romantic Movement, that brought feelings to the forefront in art, science and philosophy.