💬 NLP

Dyadic: A Scalable Platform for Human-Human and Human-AI Conversation Research

This paper introduces Dyadic, a modular, web-based platform designed to facilitate scalable research on human-human and human-AI conversations through features like multi-modal interaction, real-time monitoring, AI assistance, and integrated survey deployment without requiring coding expertise.

Original authors: David M. Markowitz

Published 2026-03-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David M. Markowitz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to study how people talk to each other. In the past, doing this for research was like trying to film a play using a camera that only recorded in black and white, had a broken timer, and required you to build the camera yourself out of spare parts. It was clunky, expensive, and hard to control.

This paper introduces Dyadic, a new, all-in-one "digital playground" designed to make studying conversations (both between humans and between humans and AI) easy, precise, and flexible.

Here is a simple breakdown of what Dyadic is and why it matters, using some everyday analogies:

1. The "Lego" Research Lab

Think of Dyadic as a Lego set for conversation research.

Before: Researchers had to build their own chat rooms from scratch, which required heavy coding skills (like being a master carpenter).
Now: Dyadic is a pre-built kit. You just snap the pieces together. You can create a "Study" (the whole project), build "Rooms" (the specific chat sessions), and assign "Slots" (the seats where people sit).
The Magic: You can give the person in Seat A different instructions than the person in Seat B without them knowing. It's like a stage director whispering different lines to two actors in the same scene, and they don't realize the other is hearing something different.

2. The "Chameleon" AI Partner

One of Dyadic's coolest features is that it can bring AI robots into the conversation.

The Disguise: The AI can sit in a chat room looking exactly like a human. Unless you tell the participants, "Hey, that's a robot," they won't know the difference.
The Stopwatch: Researchers can control exactly how long the AI takes to reply. Want the AI to be super fast? Set it to 2 seconds. Want it to be slow and thoughtful? Set it to 5 seconds. This helps scientists study how waiting for a reply changes how we feel.
The Voice: It works with both text and voice. You can talk to the AI just like you would on a phone call, right in your web browser.

3. The "Live Control Room"

Imagine a director sitting in a booth watching a live broadcast.

Live Monitoring: Researchers can watch the chat happen in real-time. If things get weird or sensitive, they can step in and send a message to the participants to guide them back on track.
The "Suggestion Box": Dyadic can offer a "smart assistant" feature. If a participant is stuck, the AI can whisper three suggested replies to them. The participant can pick one, edit it, and send it. This lets researchers study how much help people need when they talk.

4. The "Pop-Up Quiz" (In-Action Surveys)

Usually, after a conversation, researchers ask people, "How did that feel?" But by then, people have forgotten the details.

The Innovation: Dyadic lets researchers pop up a survey right in the middle of the chat.
The Analogy: It's like a video game that pauses to ask, "How are you feeling right now?" before you continue playing. You can ask people to rate their mood or answer a question after every 5 messages, or every 30 seconds. This captures their feelings while they are happening, not after they've faded away.

5. The "Black Box" Recorder

Every single word typed or spoken is recorded with military precision.

The Data: It saves who said what, exactly when they said it (down to the millisecond), and whether it was a human or a bot.
The Export: At the end, you can download all this data into simple spreadsheets (CSV files) to analyze. It's like getting a perfect transcript of the entire event, ready for you to study.

6. Safety and Security

The paper also reassures us that this playground is safe.

The Vault: All data is locked down with high-tech encryption (like a bank vault).
Privacy: It doesn't store people's home addresses (IP addresses). Researchers can only see the data from their own experiments, ensuring no one accidentally spies on another researcher's participants.

The Bottom Line

Dyadic is a tool that lowers the barrier to entry.
Previously, if you wanted to study how people interact with AI, or how two people build a relationship over text, you needed a team of computer programmers. Now, a psychologist, a sociologist, or even a curious student can log in, set up a chat room, invite participants, and start collecting high-quality data immediately.

It's like giving everyone a professional-grade camera instead of a broken toy one, allowing us to finally capture the messy, beautiful, and complex reality of human (and human-AI) conversation.

Based on the paper provided, here is a detailed technical summary of Dyadic, a scalable platform for conversation research.

1. Problem Statement

The empirical study of conversation has historically been hindered by a lack of tools that are both modular and adaptive to diverse researcher needs. Existing tools often fail to address:

Complex Dependencies: The difficulty of modeling within- and between-person dependencies in interactive processes.
Measurement Fidelity: The inability to capture dynamic, emergent phenomena at the precise timescales they occur.
Siloed Approaches: A disconnect between fields such as computer-mediated communication, AI-mediated communication, and the psychology of language.
Scalability and Accessibility: The high technical barrier (often requiring coding) to study human-human and human-AI interactions simultaneously, particularly with the recent emergence of Large Language Models (LLMs) as conversational partners.

2. Methodology and System Architecture

The paper presents Dyadic (https://www.chatdyadic.com/), a web-based research platform designed to overcome these barriers. The methodology is defined by its technical architecture and experimental design capabilities:

Infrastructure:
- Cloud-Based: Hosted on cloud infrastructure to support concurrent sessions across multiple chatrooms.
- Real-Time Communication: Utilizes persistent WebSocket connections to ensure low-latency interaction regardless of geographic distance.
- Security: Implements HTTPS with HSTS, bcrypt password hashing, and AES-256-GCM encryption for API keys. Participant IP addresses are not retained.
Organizational Hierarchy:
- Structure: Researchers $\rightarrow$ Studies $\rightarrow$ Rooms $\rightarrow$ Participant Slots.
- Slotting System: A core feature where participants are assigned fixed, numbered slots (e.g., Slot A, Slot B). This allows for within-room manipulations, where different participants in the same session receive different instructions, roles, or conditions without knowing their counterpart's setup.
Conversation Modalities:
- Text-Based: Minimalist interface with full history, timestamps, and embedded surveys.
- Audio-Based: Browser-based microphone API (no software installation required). Supports "always-on" or "push-to-talk" modes.
- Transcription: Audio sessions are transcribed server-side using OpenAI's Whisper-1 (for human speech) and gpt-4o Realtime (for AI speech), stored with millisecond-precision timestamps.
AI Integration:
- Supports Human-Human and Human-AI interactions.
- LLM Connectivity: Integrates with OpenAI, Anthropic, Google Gemini, and Hugging Face via researcher-provided API keys.
- Control: Researchers can define AI personas via system prompts and manipulate response latency (fixed or random intervals) to simulate human-like or distinct AI timing.

3. Key Contributions and Features

Dyadic introduces several novel features that distinguish it from existing chat platforms:

In Situ Survey Deployment:
- Unlike traditional post-interaction surveys, Dyadic allows researchers to embed surveys (Likert scales, feeling thermometers, open-ended text) directly within the live chat session.
- Trigger Mechanisms: Surveys can be triggered manually, by time elapsed, by message count, recurring intervals, or post-chat.
- Granularity: Responses are linked to specific conversation data preceding them, enabling dynamic systems analysis.
Live Monitoring and Intervention:
- Live Monitoring: Researchers can observe chats in real-time (similar to a lab control room) to ensure participants remain on task, especially during sensitive topics.
- Message Injection: Researchers can inject instructions or feedback directly into the chat stream during the session.
- AI Suggestions: The system can provide an LLM-generated list of suggested responses to a specific participant (visible only to them), allowing for the study of AI-assisted communication.
Data Export and Metrics:
- Exports are available in flat-file formats (CSV) at both the Room and Study levels.
- Data Points: Includes full transcripts, millisecond timestamps, sender slot IDs, bot/human flags, response latency, typing behavior, and mouse clicks during composition.

4. Results and Outcomes

As a tutorial and platform introduction, the paper does not present empirical results from a specific study but rather demonstrates the functional capabilities of the platform through its design and feature set:

Operational Success: The platform successfully supports the creation of complex experimental designs (e.g., between-subjects experiments via room sets, within-subjects via multiple rooms) without requiring participants or researchers to write code.
Flexibility: It successfully bridges text and audio modalities and integrates AI agents indistinguishably from human participants.
Data Integrity: The system successfully captures high-fidelity data (transcripts, metadata, behavioral metrics) suitable for linking verbal behavior to broader social and psychological outcomes.

5. Significance

The significance of Dyadic lies in its potential to democratize and advance conversation research:

Lowering Barriers to Entry: By removing the need for coding, it allows scholars from various disciplines (psychology, communication, linguistics) to conduct rigorous conversation studies.
Enabling New Research Paradigms: It facilitates the study of Human-AI interaction, a rapidly growing field, allowing researchers to manipulate AI behaviors (latency, persona) systematically.
Methodological Precision: The ability to collect in situ data (surveys during the chat) and high-resolution behavioral metrics (typing speed, latency) allows for a more precise understanding of interaction dynamics than retrospective self-reports.
Interdisciplinary Bridge: It serves as a unifying tool for siloed fields, providing a common infrastructure for studying interaction across computer-mediated communication, AI research, and social psychology.

Limitations

The authors acknowledge current limitations:

Mobile Optimization: The platform is currently best suited for desktop; mobile interfaces are not fully optimized.
Audio Dependencies: Audio rooms rely on browser microphone permissions, which can introduce technical friction.
AI Costs and Replicability: While API access is possible, using commercial LLMs incurs costs. Furthermore, changes in underlying AI models can affect replicability, necessitating careful documentation of model versions.