Affective Dialog
Affective Dialog refers to conversations between humans and computers in which the system detects the mood of the person it's talking to and adjusts its responses accordingly. So it's not just about what is said, but also about how it is said.
When you talk to a person, you hear more than just the words. You can tell from the voice whether someone is annoyed, sad, or excited. A normal computer program doesn’t notice any of this. Affective Dialog describes conversational systems that try to do exactly that: they estimate the user’s emotional state and tailor their response accordingly. One example: someone who writes angrily to a customer service program doesn’t get the same standard phrase as someone who is simply asking out of curiosity. The English term “affective” means roughly “related to feelings.”
Why machines should pay attention to moods
A conversation rarely fails because the information was wrong. It fails because the tone wasn’t right. If someone writes “This still doesn’t work” after three failed attempts, the right response isn’t the same instructions for the fourth time. Systems that ignore this difference come across as stubborn to users and are quickly abandoned.
For companies, this is a very concrete issue. Call centers and chat support cost money, and companies want to hand off as large a share of that as possible to software. This only works if the software prevents conversations from escalating. That’s why some systems automatically hand off to a human as soon as they detect strong anger. This handoff is often the actual economic benefit, not the empathetic wording.
At the same time, this is a sensitive area. A system that measures emotions collects very personal data about people. The European Union’s AI regulation, known as the AI Act, therefore largely bans emotion recognition in workplaces and schools. Anyone using such technology is thus operating within a legally narrow framework.
From mood detection to the appropriate response
Technically, such a system consists of two tasks. First, it has to assess the mood. With typed text, this happens through word choice, sentence structure, and punctuation. With spoken language, pitch, volume, and speaking rate are added. Some systems additionally analyze facial expressions via the camera.
The result is usually not a clear emotion, but a probability. The model might output, for example: 70 percent anger, 20 percent disappointment. It is trained on thousands of conversation excerpts that people have previously labeled by hand with emotion tags. This is exactly where a weakness lies: such labels are subjective, and cultures express emotions differently. A loud voice does not mean anger everywhere.
In the second step, the detected mood flows into the response. In modern language models, i.e. AI systems that generate text word by word, this often happens via an additional instruction in the background. It might say, in effect: “The user is frustrated, respond briefly, without promotional phrases, and offer a solution.” The model then adjusts its word choice and length accordingly. A common misconception is that the machine actually feels something in the process. It merely recognizes patterns and selects an appropriate language behavior.
Where such systems are in use today
The technology is most widespread in customer service. Major providers of call center software analyze calls in real time and display a mood indicator on the employee’s screen. Chatbots from banks and online shops also use simple variants of this. In cars, there are assistants that try to detect fatigue or stress from the voice.
A second field is companion apps and health applications. Voice assistants with speech output now sound noticeably more lively and adjust their tone. In mental health apps, the conversation is meant to have a calming effect. Experts warn, however, against confusing such programs with therapy.
In news articles, the term usually appears in two contexts: in product announcements, when an assistant is supposed to become “more empathetic,” and in regulation, when it comes to the limits of emotion recognition. It is useful to distinguish this from sentiment analysis. Sentiment analysis only assesses whether a text is positive or negative. Affective Dialog goes further because it feeds the insight directly back into the ongoing conversation.