Source-linked AI summary
One Is Not Enough: How People Use Multiple AI Models in Everyday Life
Seunghwa Pyo, Donggun Lee, Jungwoo Rhee, Soobin Park, Youn-kyung Lim
TL;DR
Prior HCI research has offered limited empirical understanding of how people coordinate across competing MLLM platforms. This study combines diary documentation with follow-up interviews to examine everyday multi-MLLM use. It finds that users maintain context-sensitive model hierarchies, with implications for tools that support cross-platform workflows.
Problem
Method
The study paired a four-day diary of MLLM use with follow-up interviews and thematic analysis.
Source: §2.1 · p. 2, §2.3 · p. 2Results
Takeaways & Limitations
The findings point to design opportunities for transferring structured task context between models and selectively partitioning interaction histories.
Source: §4 · p. 5The ten participants were mostly in their twenties and technically proficient, and the four-day diary did not capture longer-term hierarchy dynamics.
Source: §5 · p. 5
Abstract
from arXiv · showhide
People increasingly use multiple Multimodal Large Language Models (MLLMs) concurrently, selecting each based on its perceived strengths. This cross-platform practice creates coordination challenges: adapting prompts to different interfaces, calibrating trust against inconsistent behaviors, and navigating separate conversation histories. Prior HCI research focused on single-agent interactions, leaving multi-MLLM orchestration underexplored. Through a diary study and semi-structured interviews (N=10), we examine how individuals organize work across competing AI systems. Our findings reveal that users construct primary and secondary hierarchies among models that shift over usage context. They also develop personalized switching patterns triggered by task aggregation to adjust effort and latency, and output credibility. These insights inform future tool design opportunities, supporting users to coordinate multi-MLLM workflows.
1 Introduction
People increasingly combine MLLMs for their perceived complementary strengths, but coordinating across platforms remains underexplored. This study examines users’ cross-platform organization and coordination strategies.
1 Introduction
- People combine MLLMs to use their perceived complementary strengths within and across tasks.The introduction describes selecting models for capabilities such as reasoning or creative generation.Source: §1 · p. 1
- Prior research leaves limited empirical understanding of how people allocate tasks, reconcile outputs, and maintain coherent workflows across MLLM platforms.Existing HCI work largely examines trust and collaboration in single-agent interactions.Source: §1 · pp. 1–2
- A four-day diary study and semi-structured interviews examine users’ model hierarchies and cross-platform coordination strategies.The study asks how users distribute tasks and coordinate interactions across multiple MLLM platforms.Source: §1 · p. 2
2 Methodology
The qualitative study paired four days of diary entries with follow-up interviews, then used inductive thematic analysis to develop themes about model hierarchies and coordination.
2.1 Diary Study and Post-study Interview
- Participants logged each MLLM use in a web diary for four days, followed by interviews about model choices and workflow strategies.Diary entries recorded the model, selection rationale, prompt, satisfaction, and emotional state.Source: §2.1 · p. 2
2.2 Participants
- Ten adults aged 23–29 who regularly used at least two MLLM services and used models complementarily took part.Recruitment required practical cross-platform coordination experience.Source: §2.2 · p. 2
2.3 Data Collection and Analysis
- Researchers analyzed 129 diary entries and 10 interviews using inductive thematic analysis, consolidating codes into two overarching themes.The themes were Individual MLLM Hierarchy Structures and Cross-Platform Coordination Strategies.Source: §2.3 · p. 2
3 Findings
Participants organized multiple MLLMs into context-sensitive hierarchies and used personalized strategies to coordinate work. They switched models to match task needs or credibility demands, but sometimes stayed with one model to preserve context or refine prompts.
3.1.1 Different Hierarchy Across Personal and Work Contexts.
- Four participants used one favorite model across personal and professional contexts, valuing familiarity and accumulated history while occasionally consulting secondary models.They favored a general-purpose assistant that required less re-explanation.Source: §3.1.1 · pp. 2–3
- All ten participants used a primary model for frequent core tasks and secondary models for verification, refinement, or specialized needs; hierarchy patterns varied by context.The three recurring patterns are shown in Figure 1.Source: §3.1.1 · p. 2, Figure 1 · p. 3
- Three participants reserved a primary model for professional work and used secondary models to validate, compare, or debate work outputs.They did not feel a strong need to use the work-focused primary model in personal contexts.Source: §3.1.1 · p. 3
- Three participants maintained separate personal and professional primary models, matching lighter or faster options to low-stakes tasks and stronger models to demanding work.Their choices could reflect the perceived cost of correcting errors or task-specific capabilities.Source: §3.1.1 · p. 3
3.1.2 Factors Shaping Model Hierarchies.
- Model hierarchies were shaped by first impressions, expert or social signals, and use costs as participants reassessed their options.Cost could outweigh perceived performance when participants balanced task suitability against recurring fees or usage costs.Source: §3.1.2 · pp. 3–4
3.2.1 Switch Model with Purpose.
- Participants switched models to assign task stages to perceived strengths, match effort and latency to task difficulty, or cross-check credibility-critical information.Examples include ideating in one model and polishing in another, using a faster mode for trivial edits, and seeking evidence from a second model.Source: §3.2.1 · p. 4
3.2.2 Don’t Migrate, Iterate Instead.
- Some participants avoided migrating because a familiar model already held their context, making the cost of re-establishing it greater than a likely performance gain.This was especially relevant when background or personal context would need to be rebuilt elsewhere.Source: §3.2.2 · p. 4
- When outputs disappointed, some participants revised prompts and test cases with the same model, attributing the shortfall to their input rather than model capability.They iterated until the output met their requirements.Source: §3.2.2 · p. 4
4 Discussion
Users coordinate MLLMs as a changing ecosystem, not as fixed tool choices: their hierarchies and switching reflect context, task demands, and credibility. The discussion identifies cross-model context continuity and memory partitioning as design opportunities.
4 Discussion
- Users treat MLLMs as an evolving ecosystem, maintaining priority hierarchies, negotiating context boundaries, and routing tasks using dynamic signals rather than fixed preferences.Model roles and hierarchies remain fluid as users respond to peer recommendations, benchmarks, interface changes, and subscription constraints.Source: §4 · p. 4
- Switching involves re-establishing conversational context, reformulating prompts, and selectively transferring exchanges; model hierarchies also reflect first impressions, expert cues, and subscription costs.The observed selection logic is neither static nor purely performance-driven.Source: §4 · p. 4
- Users use divergent model perspectives to detect hallucinations, extending trust calibration from a user–system relationship to networked, multi-agent verification.This approach treats inter-model disagreement as a reliability signal.Source: §4 · p. 5
- A workflow-aware interface could transfer a structured task representation between models, preserving goals, constraints, and prior outputs across a switch.The proposed task card addresses the need to reconstruct task framing when conversations are isolated across interfaces.Source: §4 · p. 5
- Model-level memory partitioning could segregate histories by domain or task, supporting selective context portability and reducing context leakage across models.The proposed topic containers respond to users’ limited choice between retaining all history and erasing it entirely.Source: §4 · p. 5
5 Limitations and Future Work
The study’s ten participants were mostly in their twenties and technically proficient, and its diary lasted four days. The authors call for broader sampling and longer-term study of how model preferences change.
5 Limitations and Future Work
- The ten participants were predominantly in their twenties and technically proficient, and the four-day diary did not capture longer-term hierarchy stability.The authors note that this sample may inflate benchmark-driven switching and deliberate prompt iteration; they propose sampling across ages and expertise and tracking preferences over several months.Source: §5 · p. 5
6 Conclusion
The paper presents a qualitative account of how people structure model hierarchies and coordinate across MLLM platforms. It frames these findings as empirical complement to prior work centered on single-system interaction and identifies design opportunities for evolving AI ecosystems.
6 Conclusion
- The qualitative study characterizes participants’ model hierarchies across personal and professional contexts and their strategies for coordination across platforms.The authors position the account as complementing prior human–AI interaction research focused primarily on single-system use.Source: §6 · p. 5
A Appendix: Diary Study Questionnaire
The diary questionnaire recorded model use and selection rationale, task purpose, and optional prompt content. It also captured satisfaction, feedback, and post-interaction emotion.
A Appendix: Diary Study Questionnaire
- The questionnaire recorded the MLLM used, reasons for choosing it, purpose of use, and optional prompt content.Selection criteria could include accuracy, speed, cost, feature support, or comparison with other models; purposes included work, consultation, and study or research.Source: A Appendix: Diary Study Questionnaire · p. 6
- Participants rated satisfaction on an 11-point Likert scale and could briefly describe why they were satisfied or dissatisfied.Source: A Appendix: Diary Study Questionnaire · p. 6
- The diary captured participants’ post-interaction emotion and the reason for their selected emotional state.Options included happy, calm, sad, angry, or other.Source: A Appendix: Diary Study Questionnaire · p. 6