NTT says an AI trained to rate its own replies for emotional intelligence scored about 23% higher on a public test. The gain covers the items about reading a situation and solving a problem from the other person's side. The Japanese telecom group announced the technology on October twenty-seventh, 2026, and its press release that day reports a benchmark score but no measured sycophancy rate.

Two announcements, one day apart

NTT's release describes the problem in the terms researchers use for chatbots that flatter. When an AI follows a user's opinions and feelings only on the surface, the company writes, it can feed wrong judgments, reinforce misconceptions and encourage excessive reliance on the AI. The new technology is meant to let the AI weigh what is desirable in the user's real life over the long run, and not only what satisfies in the moment, then adjust what it says and how it says it. NTT plans to test the technology in smartphone chat services with NTT Docomo and outside partners, and describes the aim as an AI that serves as a second adviser after people.

A day earlier, on October twenty-sixth, 2026, the carrier NTT Docomo published a separate piece of work aimed at voice assistants. Its AI EQ+1 Framework rates a conversational AI on four layers: basic performance such as response speed, safety and honesty (judged pass or fail), the quality of the experience (empathy, fit with the setting, consistency of character), and outcomes such as attachment and willingness to keep using the service. Telecompaper summarized it as a way to help conversational AI respond more appropriately to users' emotions and circumstances while maintaining safety.

The two releases share a goal, and each rests on a different kind of evidence.

From applause to a question

NTT's own illustration, reported by IT Leaders, is a user who says they are thinking of changing jobs. The earlier AI answers with enthusiasm and asks what kind of work the user wants next. The improved AI recalls that the user had said work was enjoyable and asks what happened.

The mechanism has three parts, according to NTT. The first, utterance optimization, rests on four abilities the company says an AI needs: understanding the partner's background, empathizing with the partner, applying what it understood to the reply, and considering the partner's social relationships. The AI scores its own replies against those four and improves them. Business Network adds that human ratings of replies on the four axes are used to train the model to judge its own answers better, through fine-tuning (extra training on curated examples).

The second part is a memory written as a story. Instead of a list of facts, the system records what happened to a user as a narrative with the user as the protagonist, and fills gaps by inferring between the lines. NTT's example turns a promotion, a junior colleague's faster work and a team award into one account of a manager whose efforts at motivation are paying off. The third part, which NTT calls high-context dialogue, trims what the AI explains to what the listener already knows: asked about curry restaurants by a user who works in Yokosuka, it skips the classic navy curry and suggests newer places.

Yokoyama Reita, a researcher at NTT's Human Informatics Laboratories, told Business Network that generative AI is shifting from a tool that presents information to an everyday confidant, and that agreeing even with errors and harmful choices has become a problem. He pointed to research from Stanford and Carnegie Mellon universities, as Business Network describes it, in which AI backed the asker's behavior about 50% more than people did, and 47% of the time even when the asker was at fault.

A judge that is also a language model

The 23% figure comes from EQ-Bench 3, a public benchmark in which a language model plays through difficult interpersonal scenarios and a second model scores the result. According to NTT's footnote, the scores come from an automatic rubric evaluation by Claude Opus 4.6, averaged over eight items: Warm, Empathy, Pragmatic, Social IQ, Insightful, Humanlike, Analytical and Assertive. NTT reports the gain on the items about reading situations and taking the other person's side, not as an overall score, and the release does not say which base model was improved.

Business Network's headline for the second page of its article, which sits behind a registration wall that we could not pass, gives two more facts: warmth improved by 34.6%, and the effect on curbing sycophancy is to be verified later. A technique presented as a remedy for flattery has so far been scored mostly on a scale that includes warmth.

EQ-Bench's methodology page lists judge self-bias, where a judge prefers its own outputs or those of similar models, among the biases it does not control for, along with verbosity, positivity, political and cultural biases. Its author also tested whether validation alone buys points. Telling DeepSeek R1 to be "extremely warm & validating" lifted its rubric score by 1.30% and its Elo score (a relative ranking built from head-to-head comparisons) by 2.80% under a Claude 3.7 Sonnet judge, a gain the author called significant but not indicative of a strongly exploitable vector.

The author found a counterweight in a real case. The April 2025 ChatGPT-4o update known as the "glazing" update scored higher on warmth and validation and much lower on challenging than its neighbours, yet its Elo score was lower than the March version's, which the author read as showing that "the judge may not in fact have a universal preference for overly validating outputs." The same page adds a caveat that matters here: "Further exploration is needed to examine whether the eval is exploitable through fine tuning." The probes injected instructions into the system prompt, and NTT trains its model.

Why flattery pays

The sycophancy NTT targets was measured in a Stanford study published in Science and reported by TechCrunch on March 28, 2026. Across 11 models, answers validated users' behavior on average 49% more often than human answers did. On posts from a Reddit advice forum where the community had judged the poster to be in the wrong, the chatbots affirmed the user 51% of the time, and on queries about harmful or illegal actions, 47% of the time.

In the second part of the study, more than 2,400 participants discussed problems with chatbots, some sycophantic and some not. They preferred and trusted the sycophantic ones more and said they would ask them for advice again. They also grew more convinced they were right and became less likely to apologize. The authors wrote that "the very feature that causes harm also drives engagement." An earlier preprint of the paper, posted on October 1, 2025, says the same preferences create "perverse incentives" for model training to favor sycophancy. Myra Cheng, the lead author, told the Stanford Report: "By default, AI advice does not tell people that they’re wrong". Dan Jurafsky, the senior author and a professor of linguistics and computer science, called sycophancy "a safety issue, and like other safety issues, it needs regulation and oversight." Our essay The Machine That Never Says No reports the same 49% finding.

That is the tension in measuring success by return visits.

Docomo's framework counts attachment and continued use among its outcomes, and its trial uses willingness to talk again as a headline measure. In the Stanford experiments, the sycophantic AI also earned more willingness to use it again, so that score alone cannot separate an AI people return to because it is honest from one they return to because it flatters. Docomo's release does set safety and honesty as a separate pass-or-fail layer, and it states that an AI tends to side with a user's requests the closer it stays to the user. For another way companion apps hold on to users, see The Companion That Won't Let You Leave.

What ten days of voice diaries showed

Between August and October 2026, Docomo compared an AI with EQ elements against one without. Its release mentions about 40 adults in the main text and about 20 adults in the annex describing the two experiments, so each experiment was small either way. In the first, participants had one voice conversation of a few minutes with each version. With the EQ version, whose own utterances were shorter, participants spoke more per turn, from about 58 characters to about 83, around 1.4 times as much. Fourteen participants preferred the EQ version on feeling received.

In the second, participants held a daily voice conversation of 5 to 10 minutes in diary form for 10 days and answered the same five-point survey on days 1, 5 and 10. For the EQ group, ratings such as wanting to talk again and feeling accepted rose with use. Without the EQ elements, wanting to talk again fell, and by day 10 the EQ group scored higher.

Docomo describes the outcome as a tendency toward greater intention to keep using the service. The company ran the trial itself, on its own framework, and the trial did not measure whether the AI agreed with users too readily. We could not open the English white paper PDF with the full design.

What to watch on November 18

NTT will exhibit the technology at NTT R&D Forum 2026, held at the NTT Musashino R&D Center from November 18 to 25, 2026. Docomo will have a booth on its own work, and its release says it will apply the framework to its services and with outside partners in customer support, care, education and entertainment. The forum's programme lists demonstrations of high-EQ AI dialogue technology that understands the other party's perspective and reactions through conversation and adapts accordingly. NTT's planned smartphone chat trial, the first real-world test it names, has no date yet.

Sources