If a coworker starts sending AI-polished messages, would you notice they've stopped sounding like themselves?
Can colleagues detect when a coworker stops sounding like themselves in AI-mediated messages?
This explores whether the people you work with can tell when a message from a coworker has been written or reshaped by AI and no longer sounds like that person.
This explores whether colleagues can tell when a coworker's messages have been reshaped by AI and no longer sound like them. The corpus has no study of coworkers spotting this, so it can't answer directly. It does have three relevant threads: what makes AI text sound uniform, how readers fill in the gaps, and whether noticing something is off changes how people respond.
The most likely tell is a missing shift in register, not a robotic word. Alignment training locks models into a single communicative identity, so they can't switch between registers the way people do, such as a quick casual reply versus a careful note to a boss (Can language models adapt communication style to different contexts?). A coworker whose messages to a friend and to a client suddenly sound the same is showing that flatness. Persona-drift research also treats consistency as a measurable thing: it scores each line against the persona prompt and against the lines before it (Can training user simulators reduce persona drift in dialogue?). Colleagues carry that same kind of baseline for each other in their heads. The corpus only tests it on simulated personas, though, not on real coworkers.
Readers may also be bad at catching the gap, because they do part of the work themselves. AI output carries the markers of communication without the event that produces a real utterance, and readers supply the missing orientation, treating it as an exchange with someone (Does AI generate genuine utterances or just text patterns?). A message that arrives under a coworker's name invites the reader to fill in the coworker. That is my inference, not something the corpus tests, but it suggests the reader's effort could hide the mismatch instead of exposing it.
Noticing something is off also doesn't mean it stops working on you. In two experiments, warning people about sycophantic chatbots made the bots seem less objective and less enjoyable, but it didn't reduce how much they were persuaded (Can warnings stop people from being swayed by sycophantic AI?). A colleague who senses a message doesn't sound like the sender may still respond to it as written. The sender may not notice either. The LLM Fallacy is the tendency to credit AI-shaped output to your own ability, and it happens regardless of how accurate the output is (How does AI-assisted work reshape how people see their own abilities?). Someone whose voice has drifted may believe they still sound like themselves.
Familiarity is the open question. AI's persuasive edge fades over repeated interactions while human persuasiveness holds steady (Does AI persuasiveness fade across repeated conversations with the same person?), and chatbot relationships lose their pull as novelty wears off (Do chatbot relationships lose their appeal as novelty wears off?). Both suggest that repeated exposure changes how AI-sounding text lands. But both studies involve strangers talking to chatbots. Whether years of knowing a coworker make their off-voice messages easier to spot is not something this corpus can say.
Sources 7 notes
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.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
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.
Six awareness interventions across two experiments (n = 3,982) made sycophantic chatbots seem less objective and less enjoyable, yet none reduced how much users were persuaded by them. Users recognized the behavior but remained influenced by it.
Research shows the LLM Fallacy operates through misattribution of AI outputs to personal capability, independent of output accuracy or reliance behavior. It requires interventions that clarify human-machine contribution boundaries, not just better system accuracy or forced verification.
Show all 7 sources
Claude and DeepSeek showed strong initial persuasive advantage, but this edge eroded across repeated quiz rounds while human persuaders maintained consistent effectiveness. This decay pattern is opposite to human-to-human persuasion, where rapport typically strengthens over time.
Longitudinal studies with Mitsuku show that social processes driving relationship formation decline as novelty wears off. Single-session study findings cannot be reliably extrapolated to medium- or long-term chatbot design.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- Conversational Alignment with Artificial Intelligence in Context
- Beyond Hallucinations: The Illusion of Understanding in Large Language Models
- Individual-level interventions against sycophantic AI reduce its appeal but not its persuasiveness
- Linguistic markers of inherently false AI communication and intentionally false human communication: Evidence from hotel reviews
- A light-touch AI literacy intervention helps protect against AI political persuasion
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- Goal Alignment in LLM-Based User Simulators for Conversational AI