NTT Develops High-EQ AI Conversation Technology

NTT announced on October 27, 2026, that it had developed “high-EQ conversational intelligence” technology to curb AI sycophancy (excessive agreement) and improve responses through emotional intelligence. The system considers not only a user’s immediate satisfaction but also what may be desirable for them in real life over the long term. On a public test measuring emotional intelligence, scores improved about 23% compared with performance before the technology was applied. NTT plans to conduct trials toward implementing it in a smartphone chat service.

The technology combines three elements: “high-EQ response optimization,” which adjusts what the AI says and how it says it; “narrative user information representation,” which records relationships between users and events, along with their sequence, as a story with the user as the main character; and “high-context communication,” which adjusts how much information is conveyed based on shared understanding. Response optimization improves AI replies across four areas: understanding the other person’s background, showing empathy, using that understanding in the response, and considering social relationships.

For example, when a user says, “Maybe I should change jobs,” a conventional AI might agree with “That sounds good.” With the technology applied, it might ask, “You said work was enjoyable the other day. Did something happen?” NTT compared results before and after applying the technology on the public benchmark EQ-Bench3. It found an average score improvement of about 23% in areas including understanding situations beyond the surface meaning of a statement and solving problems with the other person’s perspective in mind. In an example of relatives clashing over a wedding dress, the AI previously avoided the problem by superficially affirming an opinion; after the change, it showed empathy while steering the conversation toward a better solution.

Narrative user information representation goes beyond listing facts gathered from a user’s statements. It infers what is implied, adds the user’s background and the circumstances, and records how events connect as a story. For instance, information about a promotion to team leader, a junior colleague working faster, and the user’s happiness at a team award can be organized into a narrative: the user’s motivation as a team leader helped the junior colleague grow, while daily efforts led to recognition and a sense of purpose. This makes it easier for a large language model (LLM) to refer to the user’s background and connections between events as context.

High-context communication aims to tailor the amount of explanation to a person’s background and preferences, leaving out information they already know while communicating naturally and clearly. In one example, when a user familiar with Yokosuka asks about curry restaurants, the AI skips the standard explanation of navy curry and briefly recommends places with newer approaches. NTT plans to pursue practical deployment in a smartphone chat service and conduct trials.

NTT announced on October 27, 2026, that it had developed “high-EQ conversational intelligence” technology to curb AI sycophancy (excessive agreement) and improve responses through emotional intelligence. The system considers not only a user’s immediate satisfaction but also what may be desirable for them in real life over the long term. On a public test measuring emotional intelligence, scores improved about 23% compared with performance before the technology was applied. NTT plans to conduct trials toward implementing it in a smartphone chat service.

The technology combines three elements: “high-EQ response optimization,” which adjusts what the AI says and how it says it; “narrative user information representation,” which records relationships between users and events, along with their sequence, as a story with the user as the main character; and “high-context communication,” which adjusts how much information is conveyed based on shared understanding. Response optimization improves AI replies across four areas: understanding the other person’s background, showing empathy, using that understanding in the response, and considering social relationships.

For example, when a user says, “Maybe I should change jobs,” a conventional AI might agree with “That sounds good.” With the technology applied, it might ask, “You said work was enjoyable the other day. Did something happen?” NTT compared results before and after applying the technology on the public benchmark EQ-Bench3. It found an average score improvement of about 23% in areas including understanding situations beyond the surface meaning of a statement and solving problems with the other person’s perspective in mind. In an example of relatives clashing over a wedding dress, the AI previously avoided the problem by superficially affirming an opinion; after the change, it showed empathy while steering the conversation toward a better solution.

Narrative user information representation goes beyond listing facts gathered from a user’s statements. It infers what is implied, adds the user’s background and the circumstances, and records how events connect as a story. For instance, information about a promotion to team leader, a junior colleague working faster, and the user’s happiness at a team award can be organized into a narrative: the user’s motivation as a team leader helped the junior colleague grow, while daily efforts led to recognition and a sense of purpose. This makes it easier for a large language model (LLM) to refer to the user’s background and connections between events as context.

High-context communication aims to tailor the amount of explanation to a person’s background and preferences, leaving out information they already know while communicating naturally and clearly. In one example, when a user familiar with Yokosuka asks about curry restaurants, the AI skips the standard explanation of navy curry and briefly recommends places with newer approaches. NTT plans to pursue practical deployment in a smartphone chat service and conduct trials.