You need Adobe Reader 7.0 or later in order to read PDF files on this site.
If Adobe Reader is not installed on your computer, click the button below and go to the download site.

Feature Articles: AI Constellation—Toward a World in Which People and AI Work Together

Vol. 23, No. 11, pp. 34–39, Nov. 2025. https://doi.org/10.53829/ntr202511fa3

Field Trial and Future Outlook of AI Constellation

Yuki Shiroma, Tsukasa Yoshida, and Yuichiro Sekiguchi

Abstract

This article introduces a field trial of AI Constellation—a large-scale artificial intelligence (AI) collaboration technology—titled “Conference Singularity,” held in Omuta City, Fukuoka Prefecture. Through AI-to-AI discussions using AI Constellation, we explored how ideas and opinions from diverse perspectives can be generated around complex issues, such as regional challenges, which lack a single solution due to differing stakeholder viewpoints. The trial confirmed that this approach can help deepen human discussion and lead to concrete new insights.

Keywords: AI Constellation, large language model, enhanced discussion

1. Introduction

AI Constellation, which is a large-scale artificial intelligence (AI) collaboration technology promoted by NTT, is aimed at solving complex problems that have varying solutions depending on perspective. One particularly promising use case is the enhancement of discussions within communities. In organizational decision-making, whether in companies or municipalities, AI-to-AI discussions can surface perspectives and ideas that may not naturally emerge in human-only discussions. This, in turn, can elevate the overall level of the dialogue.

In this article, we introduce a field trial titled “Conference Singularity,” which was conducted to evaluate whether discussions among AIs can stimulate and enrich human discussion. In this trial, AI Constellation was applied to local community issues and validated through a workshop format where relevant stakeholders held discussions. The following sections describe the current state of discourse in local communities that inspired the trial, the workshop experience design based on the strengths of AI Constellation, and the evaluation results of the field trial.

2. Discussion forums in local communities and their characteristics

In designing the field trial, we first analyzed the current state of discussion forums within local communities in Japan. Broadly speaking, there are four main types of forums commonly used for public discourse. From interviews with stakeholders, we identified key characteristics and challenges for each type.

2.1 Town hall meetings/workshops

These are forums where municipal leaders exchange opinions directly with residents about local issues or city policies or where citizens gather to brainstorm ideas on various themes.
Since the primary purpose is idea generation, there is often no obligation to reflect the discussion outcomes in policies or systems. Thus, ideas and opinions raised are frequently not used.
Participants are often not provided with sufficient background information in advance, leading to open-ended discussions that frequently produce impractical or unrealistic ideas.
The direction of the discussion is left to participants, and in some cases, the forum functions merely as a pressure-release valve for public dissatisfaction.

2.2 Deliberative councils (Shingikai)

Established under laws or ordinances, these executive bodies (e.g., mayors, boards of education) discuss issues referred by the administration to help reflect diverse public perspectives in city policy.
While the stated purpose is opinion exchange, the administrative direction is often pre-determined, and due to lack of prior information sharing and insufficient facilitation, discussions may not be fully effective.
In many cases, a majority of council members remain silent, only opinions that reflect their perspectives are expressed, and discussions stay within expected boundaries. As a result, critical or disruptive viewpoints are often suppressed.

2.3 Advisory committees (Kyōgikai)

Unlike deliberative councils, these are set up under administrative guidelines to gather a broad range of opinions from committee members, primarily to inform municipal operations.
Although the stated goal is also opinion exchange, the focus often shifts to creating a “publicly appealing” participant list, rather than meaningful engagement.
These forums are sometimes convened without a clearly stated purpose from the administration, leaving the direction of discussions ambiguous and entirely dependent on participants.
There is no requirement for the ideas and opinions raised to be reflected in policies, so ideas and opinions often go unused.

2.4 Decision-making bodies

These are formal mechanisms through which decisions are made in the political and administrative processes of local governments, ultimately leading to execution following approval in the local assembly.
These meetings are intended for decision-making, including the important task of choosing which projects to keep or discontinue. However, because they often need to accommodate demands from related groups, new projects tend to be added without canceling old ones; meaning that decisions are not always made with proper prioritization.
In rural or depopulated areas, resource constraints, such as limited administrative staff, budget, and population, are increasingly severe, leaving little time or capacity for meaningful discussion.

3. Designing the field trial experience

On the basis of our assessment of the current state of discussion forums within local communities, we found that applying AI Constellation to settings such as town meetings and workshops, where residents engage in open, egalitarian dialogue, would be the most effective way to evaluate its impact on revitalizing community discourse. Accordingly, we decided to conduct the field trial in the format of a workshop.

At the same time, given the difficulty of sparking lively discussion in conventional workshops, we consulted with experts and incorporated the following perspectives into the experience design:

Encourage participants to want to speak.
Create a dynamic where speakers feel drawn into the AI-to-AI conversation (through good pacing, a sense of familiarity with the AI, and a feeling of seriousness).
Inspire participants to bring their own ideas.
Make it easier for people to say things they might not normally express in typical local meetings.
Make discussion (the exchange of opinions) feel fun.
Provide participants with a tangible sense of what it’s like to generate ideas, exchange views, and choose (make decisions).
Help participants overcome reluctance or discomfort toward technology (in this case, AI), recognize it as a trusted partner, and adopt a mindset of exploring how to “make technology work for them.”

With such a design, AI is not positioned as something that replaces humans but as a tool to be tested, considered, and used in ways that draw out human potential. It also supports residents in discovering ways to leverage technology for community development.

In conducting the workshop as a testbed for evaluating AI Constellation’s capabilities, it was also important to carefully manage the structure and operation of the event. We paid special attention to the following elements in designing the program:

Participant selection: We chose individuals or groups that play key roles in the community around the discussion theme or those with relevant expertise.
Group formation: We organized discussion groups in a way that ensured participants would feel comfortable sharing their opinions and that perspectives would be appropriately diverse. Each group included at least one person who could take a higher-level view of the discussion and help summarize or consolidate ideas.
Building trust: When inviting participants to the trial, we provided individual briefings to explain the AI Constellation technology and the purpose of the event in detail. We also shared the local information preloaded into the AI beforehand. Local residents tend to be highly sensitive to the assumptions and perspectives behind comments on issues close to home. By sharing information in advance, organizers can build trust and make it possible for participants to engage critically, even with the assumptions underlying the project.
Creating a welcoming atmosphere: On the day of the workshop, we made an effort to approach participants individually before the session to ease any nervousness and create a relaxed environment.

Taking into account the above considerations in both experience design and operational planning, we proceeded to conduct the actual field trial.

4. Field trial: “Conference Singularity”

The field trial of AI Constellation was held in Omuta City, Fukuoka Prefecture, on October 17, 2024, under the title “Conference Singularity: Thinking with AIs about the future of Omuta” [1]. On the basis of the experience and program design prepared in advance, it was conducted in the form of a workshop.

4.1 Discussion themes and participants

The workshop focused on two themes relevant to challenges commonly faced by local governments such as Omuta City: “Support for small and medium-sized enterprises (SMEs)” and “Preventive care for the elderly.” Participants were invited on the basis of their relevance to the themes, and residents with the following backgrounds took part:

(1) SME Support (12 participants)

SME owners
Labor union members
Financial institution staff
Economic organization representatives
Local government officials, etc.

(2) Preventative Care for the Elderly (12 participants)

Physical therapists
Occupational therapists
Life support coordinators
Public health nurses
Local government officials, etc.

4.2 Design of AI Constellation

To implement AI Constellation in this workshop, we incorporated the technologies described in the article “Research and Development toward AI Constellation” [2] and built a system capable of producing natural, concrete, and non-redundant comments tailored to the discussion themes. We also prepared the following personas, assigning each a role tailored to the respective discussion themes, as examples of AI models with domain-specific expertise:

(1) SME Support

SME Owner
Labor Union Staff
Financial Institution Staff
Business Management Scholar
Certified SME Consultant

(2) Preventative Care for the Elderly

Physician (Internal Medicine)
Occupational Therapist
Social Epidemiology Researcher
Dentist
Community Welfare Volunteer

In addition to assigning roles similar to those of the human participants, we deliberately included personas for roles not represented in the group, allowing perspectives to emerge that would not have arisen from the participants’ viewpoints alone. In the large language models (LLMs) used by AI Constellation, such personas are typically implemented through prompts. However, providing too detailed or rigid prompts can result in narrow, overly role-bound outputs. Therefore, we designed concise prompts that provided only the role and general orientation of the persona.

Example prompt 1: Certified SME Consultant

Certified SME Consultant: Has long provided hands-on, consultative support to SMEs.

While LLMs have general knowledge, they do not inherently provide information relevant to the local context of Omuta. To address this, we supplemented the prompts with related local information, enabling participants to participate in discussions among AIs naturally.

Example prompt 2: Local information on Omuta

(1) Population (as of April 2024)

The population has declined sharply from a peak of 208,887 (in 1959) to 105,753 (in 2024).
The aging rate is 38.2% (as of September 1, 2024), the second highest in Japan among cities with populations over 100,000.
Over the past 20 years, the rate of decline due to natural factors has been increasing, while the decline from social factors has been trending downward. However, approximately 200 to 500 people have been leaving the area annually. Specifically, there has been a continued outflow of young people, both male and female, in their late teens to early twenties, typically due to reasons such as higher education or employment.

(2) Labor supply constraints

The gap between the working-age population and the number of employed people is shrinking, with an estimated shortage of up to 2000 workers (about 400 per year) over 5 years from the 2020s.
Neighboring municipalities face similar issues. In the areas where many of Omuta’s workers live, the working-age population is projected to decline by 53,219 (to 157,687) by 2040 compared with 2020.
(Additional data omitted.)

AI Constellation outputs appear as text only, lacking non-verbal cues. This presents a challenge in helping participants follow and engage with the discussion. To address this, we provided two user interfaces. First, a shared display screen visible to all participants was used by facilitators to guide the session in real time (Fig. 1). Second, a comment-history interface enabled participants to revisit AI-generated remarks during human-to-human discussions. These tools helped prevent common problems that often occur in standard discussions as well, such as losing track of the conversation or disproportionately reacting to the most recent statement.

Fig. 1. Example screen of AI Constellation (for facilitators).

4.3 Program structure

To encourage active discussion in a small-group setting, each group consisted of 3 to 4 participants seated around a round table. Dividing participants into smaller groups was intended to foster individual engagement and participation.

While an overall facilitator guided the entire program, each group followed the discussion process outlined below (Fig. 2):

1. Using AI Constellation, AIs representing diverse perspectives generated ideas on the basis of the discussion theme. Participants viewed these AI-generated ideas then exchanged their opinions within the group.

2. AI Constellation then facilitated a second round of discussions in which the AIs critiqued the initial ideas in Part 1, each offering specific reasons for their criticism. Participants reviewed these critiques and, using the AI dialogue as a clue, deepened their discussions and worked toward identifying actionable policy ideas suited for Omuta.

Fig. 2. Discussion flow in the workshop.

This format intentionally avoided having humans and AI debate simultaneously. This structure, AI’s prior conversation prompting human discussions, was designed to discourage participants from defaulting to familiar discussion patterns and encourage them to focus on AI’s ideas and articulate their own opinions. Selecting AI’s ideas was intended to bring out their desire to express their opinions. Including a step where AIs critiquing each other’s ideas helped normalize an environment where disagreement with constructive reasons is permitted. This was particularly important in enabling individuals to feel comfortable expressing opinions in situations where hierarchy or social dynamics might normally inhibit open dialogue (e.g., therapists giving opinions to physicians).

5. Participant reactions to the “Conference Singularity”

Following the workshop, we conducted a survey with all 24 participants to gauge their impressions of the AI Constellation discussion, focusing on whether the AI-generated remarks felt natural or specific and whether participants gained any new insights. The survey included the following two questions:

1. Q1: Did the AI-to-AI discussions provide you with new insights?
(Four response options: “Very much so,” “Somewhat,” “Not really,” “Not at all”)

2. Q2: Do you think basing the human discussion on the AI conversation helped deepen the discussion?
(Four response options: “Strongly agree,” “Somewhat agree,” “Somewhat disagree,” “Strongly disagree”)

The results indicate strong support: for Q1, 7 participants answered “Very much so,” and 17 answered “Somewhat;” for Q2, 12 answered “Strongly agree,” and 12 answered “Somewhat agree.” This means 100% of respondents gave positive feedback, confirming that multi-AI discussions helped participants engage meaningfully and derive specific insights without discomfort.

We also interviewed participants to better understand how they felt about using AI Constellation during the discussions. Key takeaways included:

“With human participants, people tend to avoid conflict or criticizing others’ opinions, but with AI, I felt more comfortable voicing critical views.”
“Having the AI start the discussion brought sensitive or delayed topics to the table that humans might otherwise hesitate to bring up, which allowed for more in-depth conversation.”
“Critically analyzing AI-generated ideas made it easier for me to articulate my own thoughts in response.”
“Since the AI handled the initial idea generation, we could spend more time on decision-making, which made the conversation more productive.”
“This kind of technology could make local government meetings much more engaging.”

Overall, the feedback confirms that the design of this field trial worked as intended, providing participants with a valuable and meaningful experience.

6. Conclusion and future outlook

By implementing the AI Constellation concept through the “Conference Singularity” field trial, we confirmed that this approach can effectively enhance the quality of human-to-human discussions. Rather than replacing humans, generative AI served as a partner that stimulated conversation and elevated its depth and quality. Moving forward, we aim to apply this concept to a broader range of discussions around regional challenges, as well as explore its use in corporate settings such as for meetings and strategic planning. Through these efforts, we will continue advancing research and development of AI Constellation to support and enhance a wide array of human intellectual activities.

References

[1]	Press release issued by NTT, “Public participation workshop utilizing ‘AI Constellation’ for discussions between AIs,” Oct. 17, 2024. https://group.ntt/en/newsrelease/2024/10/17/241017a.html
[2]	J. Sun, A. Ide, T. Yoshida, C. Watanabe, M. Toyoda, and S. Takeuchi, “Research and Development toward AI Constellation,” NTT Technical Review, Vol. 23, No. 11, pp. 27–33, Nov. 2025. https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202511fa2.html

	Yuki Shiroma Senior Research Scientist, Innovative Computing Architecture Laboratory, NTT Computer and Data Science Laboratories. He received a B.E. and M.S. from Waseda University, Tokyo, in 2009 and 2011. His research interests include machine learning and software engineering.
	Tsukasa Yoshida Researcher, Innovative Computing Architecture Laboratory, NTT Computer and Data Science Laboratories. He received a B.E. and M.E. in engineering from Toyohashi University of Technology, Aichi, in 2018 and 2020, and is currently pursuing a Ph.D. in engineering at the same university. He joined NTT in 2020. His research interests include statistical machine learning and optimization algorithms.
	Yuichiro Sekiguchi Senior Research Engineer, Supervisor Innovative Computing Architecture Laboratory, NTT Computer and Data Science Laboratories. He received a B.S. and M.S. in mechanoinformatics from the University of Tokyo in 2002 and 2004. He joined NTT in 2004 and has been researching information retrieval and natural language processing. He is a member of the Information Processing Society of Japan.

↑ TOP