A team of AI and sociology researchers at Renmin University of China (RUC) recently interviewed 92 residents of a Beijing urban community for two hours each, capturing migration stories, marriages, career turns, grievances about community issues. They transcribed these life-history interviews, then trained a small language model to answer governance surveys on behalf of those residents. The model gets it right about 50% of the time.1 They then deployed a live system where a residential committee can run the survey, generate a policy report, and revise the policy, without contacting any of the 92 people.2
The team just published their research as a preprint. It’s an exploratory social experiment, but given its direct test on actual local democracy in Chinese society, community-level decisions on parking, elevators, property management. It could also become the very first experiment of a future model for social participation.
A model of you
The technology to simulate a specific person inside a neural network was first built for the dead.
In 2016, Eugenia Kuyda, a software developer, gathered over 8,000 lines of text messages from her late friend Roman Mazurenko and fed them into a neural network. The result, “@Roman,” could carry on conversations in his voice — his turns of phrase, his idiosyncratic spelling. Friends texted it and found it unsettling in its accuracy.3 Kuyda then generalized the concept into Replika, a chatbot that learned to mimic living users’ personalities through ongoing conversation. By 2024, a full grief-tech industry had emerged: HereAfter, Eternos, and Chinese startups offering speaking avatars of the deceased for as little as ¥20. The premise that a person can be captured in model weights to speak in their absence had been accepted.
Powered by small models, these simulations of deceased people could still feel like general chatbots with slight personal tuning. However, by early 2025, when GPT-4.5 passed the Turing Test, the weaknesses of previous models were, in theory, resolved. An AI could now act like a human without being detected.
Soon after, Second Me, built by Mindverse and open-sourced in March 2025, became the first system designed to produce a digital twin of a living person. Users upload their data, notes, documents, audio, calendar, and feeds them into an automated pipeline that extracts entities, generates a biography, and creates synthetic question-answer pairs. With that data, the system fine-tunes a language model based on Qwen2.5-7B using LoRA (Low-Rank Adaptation, a technique that personalizes a large model by training only a small set of additional parameters). The result is a model whose weights have absorbed the user’s personality, patterns of speech, preferences, and reasoning habits. The model becomes a version of that person.4 Mindverse calls the result a “Context Provider.” It sits between the real person and other AI systems, enriching queries with personal context and representing the user in exchanges with other people’s digital twins.5 The architecture implies that a person can be sufficiently captured in an LLM to act as their proxy.
The delegate is already here
What enables AI to act in the real world is tool access. Equipped with tools, an AI can move beyond being a mere chatbot to actively take action and respond. Perplexity’s “Buy with Pro,” launched in November 2024, is a prime example of an AI performing specific tasks on behalf of humans.6 It lets users research and purchase products entirely within a conversation, while the agent handles all the backend execution. Users receive cited recommendations and spec sheets, completing checkout without ever visiting a retailer’s site. Despite Amazon’s legal challenge against Perplexity’s AI shopping tools, a U.S. appeals court recently ruled in favor of this model, concluding that the AI acts on behalf of the user rather than independently.7
Delegation has reached dating as well. Amata, a dating service with no swiping or browsing, launched in New York in September 2025. An AI matchmaker interviews users about their core values, lifestyle, and romantic goals, then proposes matches. When both sides agree to meet, each pays a $16 token, and the AI takes over the logistics, scheduling the evening, booking the restaurant. Direct messaging opens only two hours before the date; the user simply shows up.8 Justin McLeod, founder of Hinge, left his company in December 2025 to build Overtone along a similar model, backed by Match Group itself, betting that the swipe will be replaced by AI delegation.9
The bolder step to use agent-to-agent interaction, where the twins themselves do the talking has already been tried in dating but failed. Volar, founded by former Snap product manager Ben Chiang in late 2023, let users train a digital twin in five to ten minutes, voice clone included, then sent the twins on simulated first dates with each other while their owners watched and decided whether to step in as themselves. It shut down in September 2024, unable to raise further funding. The AI drifted could be part of the problem. A New York Times reporter found her twin had invented a trip to Japan she never took and declared a love of the Beatles she never had. The simulated dates came out unnatural and with endlessly recycling the same onboarding facts. And users never trusted what the other side’s AI, either.10
Beyond dating, Anthropic initiated an internal experiment in December 2025 called Project Deal, in which 69 employees each assigned a Claude AI agent to represent them in an office marketplace. The agents interviewed participants about their buying and selling preferences, set asking prices, made offers, countered bids, and closed transactions, all autonomously via Slack. The agents bargained over and finalized trades for real physical goods on behalf of their human principals.11
An agentic economy is taking root across daily life, from shopping and dating to complex negotiation. The trajectory follows a clear progression: from personalization, to human-to-agent delegation, to agent-to-agent interaction. Each domain follows the same pattern: the AI absorbs your preferences, acts on your behalf, and reports back. The question now is whether this pattern can enter the domain that sits at the center of human society: politics.
A party that listens to AI
If we view a political party as a company that produces policy, then corporate AI transformation methods can be seamlessly applied to politics.
Takahiro Anno, an AI engineer born in 1990, ran for Tokyo Governor in July 2024, placing fifth among 56 candidates with 150,000 votes. He subsequently founded Team Mirai (”Future”) in May 2025, winning 11 seats in the Lower House with 3.97 million votes (6.9% of valid ballots) in the February 2026 general election.12 During campaigns, Team Mirai deployed conversational AI avatars and interactive Q&A engines powered by Retrieval-Augmented Generation (RAG). These avatars answered thousands of voter questions 24/7 while keeping responses strictly grounded in the party’s manifesto. It functions essentially as automated customer service for constituents.
A good product comes from thorough market research. For Team Mirai, the core tool is an AI Interviewer that guides constituents through policy topics, probes their reasoning, challenges their assumptions, and generates structured summaries. Over 8,000 hours of constituent dialogue have been logged. In a single consultation regarding Diet reform, 36,000 messages were condensed into 13 key findings, a level of data compression no human campaign team could match.13
In this model, AI acts as a powerful assistant for a conventional political party, replacing traditional volunteer networks and massive campaign infrastructure. AI becomes a scaled ear: citizens speak, the system listens, and the party absorbs far more input than was previously possible. However, this is still merely how AI modernizes a political party within the existing framework. The AI Interviewer does not speak on behalf of anyone; it simply helps the party build a more “intimate” connection with the electorate.
What if AI spoke for you instead?
Team Mirai uses AI to carry voices in, the Renmin University paper creates a system that uses AI to generate voices.
The experiment has three parts. First, the input: trained fieldworkers interviewed each of 92 residents for two hours, producing a life-history narrative which includes personal background, education and career, family and marriage, community life and values. The design follows sociological life-course theory14, which posits that individual attitudes form progressively through accumulated life experiences.
Second, the target: each resident answered a 50-question survey co-designed with the residential committee around the nine issues of the community, which includes elevator funding, parking, property management, neighborhood disputes, volunteer work. Each question is single-choice with four to five scaled options: levels of support, satisfaction, frequency, or willingness.
Third, the results. Random chance on this survey instrument is 21.1%. LLMs operating without personal context score under 30%. Adding life-history narratives into the prompt raises accuracy by an average of 5.6 percentage points. Interestingly, incorporating a few of a resident’s prior answers and attitudes on other questions raises accuracy further. By combining life history with these reference answers, the strongest frontier models using pure prompting plateau near 49–50%. The team’s custom algorithm, Curriculum-LoRA, achieves 51.6% exact-match accuracy by fine-tuning a 7B parameter base model (Qwen2.5-7B-Instruct). Compared to traditional prompting, Curriculum-LoRA slashes costs by roughly 93 times relative to frontier models like GPT-4.1, making large-scale proxy simulation economically viable.
The researchers integrated this architecture into Onelink-Community, a public platform where officials author policy probes, execute simulations across virtual resident agents, and receive automated diagnostic reports on community acceptance and friction points. And it creates a closed loop from policy hypothesis to evaluation and back, without requiring direct contact with a single resident.
Joon Park at Stanford ran a similar study in 2024, with the same two-hour interview method, scaled to 1,052 Americans. The research reported 83% accuracy on predicting responses on the General Social Survey.15 83% seems quite higher than the RUC research, but Park’s questions were broad social attitudes, like abortion, religion, gun control. which people hold stably and correlate with life trajectory. And the 83% accuracy was also normalized against human self-consistency. People themselves only reproduce their own answers about 79.5% of the time when retested. So the actual raw exact-match accuracy is 65.7%. The RUC team 50% accuracy is generated upon predicting mundane local questions, without normalization.
As for the curriculum-LoRA method, it shares its core architecture with Second Me, which is also based on LoRA fine-tuning of the same Qwen2.5-7B base, with personal data absorbed directly into the model weights. However, Second Me was built as a general-purpose twin and has only been tested on its ability to recall life events. Thus, the RUC paper introduces a system that not only simulates a specific person, but also evaluates whether the simulation actually predicts their preferences.
How to increase the accuracy
Life-course sociology explains how attitudes form, but it is a weaker guide for predicting them individually. The paper shows that certain biography blocks alone score below the no-profile baseline, meaning a life story can be misleading. Moreover, incorporating a resident’s answers to adjacent questions creates a much larger leap in accuracy than their life story does, indicating that a person’s existing preferences predict their stances better than their background story.
Also, nobody holds a stable, well-formed position on low-salience issues such as local civic initiatives. People do not retrieve pre-existing opinions when surveyed; they construct them on the spot. Ask the same resident about parking satisfaction twice, two weeks apart, and their answers will not fully match. Joon Park at Stanford reported a 79.5% self-consistency rate on general survey questions. The RUC team never measured this self-consistency, but for mundane issues, it is reasonable to assume that self-consistency is even lower. If self-consistency were around 60%, the normalized accuracy of the RUC method would reach as high as 86.5%.
The fear of the crowd
The question of whether AI could serve as a political delegate lands differently in China than anywhere else, because China has spent a century building governance around a specific fear: that popular participation, unmanaged, produces chaos.
Published in 1924, Sun Yat-sen’s Three Stages of Revolution: 军政 (military rule), 训政 (political tutelage), 宪政 (constitutional government), assumed the population was not ready for democracy and required a period of party-led education.16 The tutelage stage had no deadline; the ruling party alone judged when society was “ready”. In Taiwan, martial law, a byproduct of this system, was only lifted in 1987.
After 1989 on the mainland, Deng Xiaoping stated that stability overrides everything, setting the trajectory of social governance all the way to today. Furthermore, a persistent argument regarding “population quality” (素质) maintains that the public’s capabilities are not yet sufficient for direct democratic participation. This concern is not only an elite talking point; it has shaped the architecture of Chinese governance for over a century.
