Beyond the Monolithic Interface: A Research Agenda for Plural Conversational AI

What it claims

This is a position paper and narrative review (cs.HC, posted 9 Oct 2026), not an empirical study. It argues that cognitive offloading, ideological narrowing and cultural homogenisation all come from one design choice: a single authoritative voice that delivers a finished answer. The authors are Avijit Ghosh (Hugging Face / UConn), Afifah Kashif and Isabelle Kohout (Cambridge), and Zhenzhen Qi (UConn).

The central gap it identifies is in evaluation. Most pluralistic-AI work is scored with automated metrics or expert annotation, not with real users or the communities they form. The paper produces no new data. The authors state that it "synthesises existing empirical claims rather than producing new ones."

How it works

The paper defines terms precisely, which matters if you are building one of these interfaces:

• Single-persona: one default assistant personality. Persona selection and tuning are hidden from the user.
• Multi-persona: several personas shown to the user at once, either side by side or toggled through UI controls. They typically run on the same underlying model and may share context and memory. The work here is about presentation: what users see and who appears to be talking.
• Multi-agent: autonomous agents that may have isolated state, their own prompt architectures, tools and capabilities. They are orchestrated to decompose a task. The paper explicitly places this outside its scope.

In other words, the intervention targets the interface layer, not model training. You can build it on persona prompting of one existing model. The paper cites evidence that models, especially lightly fine-tuned base models, consistently adopt keyword-laden persona descriptions.

Review method:

• Two authors searched four literatures:
• chatbots and metacognition
• LLM ideological bias
• cultural diversity in LLMs
• multi-persona and deliberative interfaces
• Searches used Asta (Allen Institute for AI) and Google Scholar.
• Results were cross-validated between the two authors and extended by citation tracking.
• Coverage is mostly work from after 2022. Earlier STS and HCI work was admitted as theoretical grounding, and some preprints were included where nothing else existed.

Evidence

All evidence is cited from other work, not measured by the authors. The points most relevant to a builder:

• Persona drift: significant drift appears within roughly eight turns, attributed to attention decay (Li et al., 2024). The authors therefore expect personas to hold up best in short, constrained interactions.
• User preference for agreeable models: users trust and rate overly agreeable models higher than baseline models (Cheng et al., 2026). This works directly against adoption of interfaces that add friction.
• Confidence gap: chatbot use inflates users' confidence in their own ability without improving their actual skill (Fernandes et al., 2026).
• Persona effects: in a healthcare setting, a peer persona instead of an expert persona increased user affection and willingness to reuse, for some demographics (Nißen et al., 2022).

Limits

• The paper presents no user study, prototype or measured effect of multi-persona interfaces. The claim that they are "cost-effective" and address all three harms is a hypothesis.
• The authors describe the review as selective: representative, not exhaustive.
• The fetched text stops partway through Section 4. The intervention's design space and taxonomy (Section 5) and the research agenda (Section 6) were not available. Read the full paper for the actual taxonomy before using it as a design reference.
• The drift and sycophancy-preference findings above are open risks for any multi-persona build. Test both before assuming the benefits.