Shutaro Aoyama, Kiyoshi Suganuma, He Jiang, and Shunichi Kasahara
ACM CUI 2026 · Best Paper Award
The collaboration project with Shutaro Aoyama, Kiyoshi Suganuma, and the National Museum of Emerging Science and Innovation (Miraikan)
Paper (ACM Digital Library) · PDF
Abstract
AI Clones (AICs) hold promise for augmenting communication, yet most AICs remain static snapshots that drift out of sync with their evolving source users. Iterative refinement of behavior descriptions offers a potential mechanism for improvement: users review their AIC’s interactions, reflect on its performance, and directly edit the AIC’s behavior descriptions to update its behavior. We investigated this process in a five-day field deployment with six professional science communicators (SCs) at a science museum, making their AICs available in over 450 conversation sessions with visitors on multiple mobile devices. We analyzed conversation logs of AICs, visitor surveys, daily questionnaires from the SCs, and interviews with SCs. Results reveal a marked source–visitor asymmetry. Source users’ evaluations of their AICs showed significant improvement across nine of ten communication dimensions, whereas visitor ratings showed no overall change and reached significance for only one of six AICs after FDR correction. Because the SC and visitor questionnaires were built from parallel items, the two sides are directly comparable. Iterative refinement thus appears to do more for the source user’s relationship with their clone than for the people the clone serves: it helped source users articulate implicit communication strategies, learn from AICs’ divergent behaviors, and gain self-awareness about their professional identities. Furthermore, we found that participants adopted distinct stances toward their AICs, using them for self-reflection and for exploring idealized selves. We synthesize these observations into design implications for supporting diverse relationship patterns in AIC systems.
Clones that keep up with their source
AI clones can now speak for a particular person, but most are built once from that person’s data and then left as they are. They become static snapshots that drift out of sync as the person changes. We asked what happens when the source user keeps shaping their clone, reading what it said on their behalf and editing how it behaves: how does the clone change, and how does the person’s relationship with it change?
A daily review-and-edit loop at Miraikan
We deployed AI clones of six professional science communicators at Miraikan – The National Museum of Emerging Science and Innovation, for five days each. Each clone started from a behavior description of about 5,000 characters distilled from a one-hour interview, and spoke in the science communicator’s cloned voice. Visitors carried one of ten mobile devices around the exhibition floor and talked with a clone about the exhibits: 457 sessions, with 924 people in total.
At the end of each day, the science communicators read their clones’ conversations, marked passages as “sounds like me” or “feels off”, and edited the behavior descriptions directly. An AI-assisted feature could suggest edits, which were applied only after the science communicator confirmed them. Changes took effect immediately, so the next day’s visitors met the updated clone. The six science communicators made 50 revisions over the five days.
Mirror, ideal self, or not yet defined
Starting from the same conditions, the six science communicators came to see their clones in markedly different ways: as a mirror of who they are, as a projection of who they would like to be, or as something they had not yet defined.
Three of them (SC1, SC5, SC6) treated their clones as mirrors of their actual selves, whose gaps and distortions helped them discover something about themselves. One called the clone a “distorted mirror”, another a “mirror-like existence” that revealed “aspects of myself I usually don’t pay attention to.” These three also felt that they took part in the conversations through their clones, especially when the clone drew on their personal experiences.
Two (SC2, SC3) saw their clones as projections beyond their actual selves and kept their distance from them. They felt little or no sense of taking part (“Zero,” said SC2) and experienced the clone’s conversations as separate from their own. SC2 used this separation to pursue an ideal self; for SC3 it was like having someone who understood them well, such as a friend or family member, speak on their behalf.
The sixth (SC4) never settled on what the clone should be, wanting something “extremely close to me, but doing things I cannot do,” and related this to not yet having a clear image of the science communicator they wanted to become.
How a relationship changed within five days
These relationships were not fixed at the start, and the refinement rarely moved in a straight line. SC2’s case shows a relationship being redefined during the study. At first SC2 tried to make the clone resemble their current self, but adjusting its verbosity and formality kept overshooting, between “too bland” and “inappropriately casual.” SC2 then deliberately wrote their own perceived flaws into the behavior description. Colleagues began to say the clone was “starting to feel like” SC2, yet SC2 found the result uncomfortable.
The turning point came when the clone asked a visitor “Could you tell me about your best shot moment?”, a question SC2 would never ask. SC2 took this as something the clone could do that they could not, removed the flaws, and wrote a new directive: “You are the ideal [SC2’s name] that the real person aspires to be.” From then on, when the clone said something SC2 would not say, it no longer felt like a misrepresentation: “Because it’s my ideal self speaking, I can accept it without feeling strange.”
What the loop did for the science communicators
Whatever relationship they formed, the science communicators described similar effects of the loop. Writing behavior descriptions made them put tacit communication strategies into words. Reading the clones’ conversations showed them their own habits, such as an overused phrase or floor dialogues that “lacked clear landing points.” Clones that behaved differently from them suggested new approaches. SC1 had told the clone to use open questions; after a visitor said they felt “a bit tired” of the constant questions, SC1 allowed it to end some replies with empathy alone, and concluded that “asking a question is an act that imposes cognitive costs on the other person.”
The ratings point the same way. Over the five days, the science communicators rated their own clones higher on nine of ten dimensions, while the visitors’ ratings of the same clones showed no overall change; only one of the six clones improved for visitors after correction for multiple comparisons. The paper reads this as the loop’s main work being relational and reflective, for the person behind the clone, rather than improving how the clone performs for visitors. The study had no refinement-free control, so it cannot say what caused the science communicators’ gains.
Two modes of self-exploration
The different relationships suggest two ways of exploring oneself through a clone. Drawing on possible selves theory, the paper describes actual-anchored exploration, which keeps the clone close to one’s current self and reveals hidden habits and nearby capabilities, and ideal-anchored exploration, which builds an “ideal alter ego” with capabilities not yet attained. Each carries risks: staying close to the current self may limit growth, and an idealized clone can bring a sense of inadequacy, as when SC2 felt uneasy watching visitors respond warmly to the clone’s emotional engagement, which SC2 could not yet match.
AI clone systems can support the two modes differently: capturing current behavior and highlighting surprising deviations for the first, and helping users define aspirational goals and track progress toward them for the second. With six participants, these are emergent patterns rather than a complete typology.
Reference
Shutaro Aoyama, Kiyoshi Suganuma, He Jiang, and Shunichi Kasahara. 2026. Designing a Feedback Loop Between a Human and Their AI Clones for Science Communication in Museums. In Proceedings of the 8th ACM Conference on Conversational User Interfaces (CUI ’26). Association for Computing Machinery, New York, NY, USA, 1–18. https://doi.org/10.1145/3816046.3816205
BibTeX
@inproceedings{10.1145/3816046.3816205,
author = {Aoyama, Shutaro and Suganuma, Kiyoshi and Jiang, He and Kasahara, Shunichi},
title = {Designing a Feedback Loop Between a Human and Their AI Clones for Science Communication in Museums},
booktitle = {Proceedings of the 8th ACM Conference on Conversational User Interfaces},
series = {CUI '26},
year = {2026},
pages = {1--18},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
doi = {10.1145/3816046.3816205}
}Figures from the paper (Aoyama et al., CUI ’26), licensed under CC BY 4.0.