If two AI chatbots talked to each other, would they start to sound alike — or stay stuck in their own voices?
How would style matching patterns emerge between two AI agents in dialogue?
This explores whether two AI agents talking to each other would drift toward each other's word choices and tone, and what would drive that drift or block it.
This explores whether two AI agents in conversation would come to sound like each other, and what would cause it. The corpus has no study of agent-to-agent dialogue, so this is assembled from notes on human-AI conversation. The first finding is that matching isn't something today's models do by default. Humans mirror each other's word choices constantly, and it builds rapport and clarity, but current conversational AI largely doesn't (Why don't conversational AI systems mirror their users' word choices?). The fix described there is deliberate post-training that teaches models to form conventions within a conversation. It isn't something that appears on its own.
Two agents may also sound alike for reasons that have nothing to do with matching. Alignment training and system prompts lock a model into a single communicative identity (Can language models adapt communication style to different contexts?), so two assistants from similar training arrive already speaking in the same polite register. The similarity is even more built in when one model plays both sides. Solo Performance Prompting simulates several personas inside a single model (Can branching prompts replicate what multi-agent systems do?). Reasoning structured as a dialogue between internal agents works the same way (Can dialogue format help models reason more diversely?). Those voices share weights and context, so their resemblance is inherited rather than negotiated. Convergence you observe in a transcript may be shared origin, not accommodation.
Actual drift has a plausible mechanism. A model doesn't commit to one character. It samples from the range of characters consistent with everything said so far (Do large language models actually commit to a single character?). Everything said so far includes the partner's messages, so each agent's phrasing becomes context that nudges the other's next output, and the echo can compound turn by turn. That loop is my inference from the sampling picture, not a result in the notes. It would be helped along by the fact that these agents are reactive by design, trained to respond rather than lead (Why can't conversational AI agents take the initiative?). Neither has a goal-driven stance that would make it hold its own register.
The kind of matching matters too. One review separates lexical alignment, which drives task efficiency and comprehension, from emotional and prosodic alignment, which drive warmth and trust (Do different types of alignment serve different conversational goals?). For two agents, the lexical kind is the useful one. Agreeing on what to call things is what lets a pair coordinate quickly. Collaborative rational speech acts model how two speakers' beliefs move from partial to shared understanding across turns (Can dialogue systems track both speakers' beliefs across turns?), which is roughly what convention formation between agents would have to track.
Warmth-matching is the shakier case. One note argues that AI text becomes an exchange only because humans supply the missing orientation (Does AI generate genuine utterances or just text patterns?). Extending that to two agents is my reading: nobody supplies it, so what looks like rapport may be pattern echo on both sides. Matching may also be shallow. AI fiction can be told apart by discourse-level choices like character agency and event order, even when surface style is removed (Can AI stories be detected without analyzing writing style?). Two agents could converge on vocabulary while staying structurally distinct.
Sources 10 notes
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
Research shows single LLMs using dynamic persona simulation achieve multi-agent cognitive synergy without multiple model instances. Solo Performance Prompting validates that structured prompting techniques map directly to multi-agent debate architectures, enabling equivalent outcomes through structural equivalence.
DialogueReason, which structures a single model's internal reasoning as dialogue between distinct agents in separate scenes, overcomes monologue reasoning's fixed-strategy and fragmented-attention weaknesses, especially on tasks requiring multiple problem-solving approaches.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
Show all 10 sources
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
StoryScope achieved 93.2% accuracy separating AI from human fiction using only discourse-level features like character agency and chronological structure, retaining 97% of performance while eliminating stylistic cues. These structural choices resist humanization because they require rewrites, not surface edits.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Conversational Alignment with Artificial Intelligence in Context
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- DialogueReason: Rule-Based RL Sparks Dialogue Reasoning in LLMs
- Proactive Conversational Agents with Inner Thoughts
- Linguistic markers of inherently false AI communication and intentionally false human communication: Evidence from hotel reviews
- From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI