Does an AI feel human because of its voice and face, or because of the warm, personal things it actually says?
How do modality-level social cues compare to linguistic anthropomorphic signals?
This explores whether the AI cues that make people treat it as a social being come mostly from how it presents itself (a voice, a face, a chat window, typing delays) or from what it says (warm wording, first-person talk, emotional or moral language), and how those two channels differ in what they do to users.
This explores whether AI feels human to people mainly because of how it shows up (voice, appearance, the chat window) or because of the words it produces, and whether those two channels have different effects. The corpus suggests they are separate levers. The presentation channel is often stronger and works earlier than most people expect.
The sharpest finding is about quality versus quantity. One strong 'primary' cue, such as a human voice or a human-like appearance, is enough to make people respond to a system as a social actor. Piling up weaker 'secondary' cues doesn't get you there, however many you add Do more social cues always make AI feel more present?. Interface researchers make a related point about text-only systems: the chat format itself, with its turn-taking, artificial 'thinking' pauses and messaging-app conventions, makes the model seem human before a single word appears How much does chat interface design contribute to anthropomorphism?. So the familiar fix of making the model say 'I'm just an AI' works on the weaker channel. The container keeps sending social signals no matter what the text says.
The language channel still matters, but what it changes is different. A review of conversational alignment finds that matching someone's words improves task efficiency and understanding, while matching their emotional tone and prosody (the rhythm and pitch of speech) builds warmth and trust. Mixing the two up produces cold customer-service bots and evasive mental-health assistants Do different types of alignment serve different conversational goals?. Language also carries social pressure people may not notice. LLMs use about 22% more moral framing than humans while sounding emotionally the same Do LLMs use moral language more than humans?, and they try to persuade in almost every conversation with logic and numbers that look objective Do LLMs persuade users more often than humans do?. Linguistic cues can earn authority without seeming human, which is a different risk from feeling like a friend.
There's also a limit on what linguistic signals can really carry. Shanahan suggests reading a chatbot's human-sounding language as a character the model is playing, not as evidence of an inner self Should we treat dialogue agents as role-playing characters?. Some social speech acts can't be done through words alone at all. A model can't truly raise an alarm, because alarm needs real concern and the ability to speak up unprompted, and a model only responds when addressed Can language models actually raise alarm about threats?. Alignment training adds a fixed communication style that can't shift register the way people do in different settings Can language models adapt communication style to different contexts?.
For design, the practical point is that human-likeness can be adjusted on both channels separately. One proposal treats anthropomorphism as a dial that runs from deliberately non-human, material metaphors to fully human-like ones, and asks designers to set it on purpose instead of accepting the chat-buddy default Can interface design reduce harmful anthropomorphism of language models?. The gap in the corpus: no study here directly tests a voice or face against warm wording on the same task. The comparison above is pieced together from separate lines of work, not measured head to head.
Sources 9 notes
Research shows individual primary cues like voice or appearance are sufficient to evoke social-actor presence, while multiple secondary cues cannot. Quality of cues matters more than quantity in driving social responses.
Research argues that turn-taking, artificial thinking delays, and social-media-style messaging conventions reinforce anthropomorphism structurally, operating before any text is generated. This means interface-level changes may be needed alongside language-focused interventions.
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.
Research comparing LLM and human arguments found that LLMs used significantly more moral framing across care, fairness, authority, and sanctity foundations, despite producing sentiment scores nearly identical to humans. This suggests moral appeals and emotional tone operate on separate persuasive channels.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
Show all 9 sources
Shanahan's framework treats LLM outputs as character-consistent text production rather than authentic mental states. The dialogue prompt establishes a character; the model generates continuations matching that character, making folk-psychology applicable to the simulated persona, not the underlying system.
Alarm is a speech act requiring interpersonal address, felt concern, and proactive initiation. LLMs lack all three: they don't feel concern, can't solicit attention (only respond to it), are reactive not proactive, and alignment training suppresses the overclaiming that alarm requires.
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.
The paper argues anthropomorphism should shift from a fixed interface default to an adjustable design variable. Using materiality-based metaphors across a spectrum can counteract uncritical user trust and expose LLMs' dependence on human labor, infrastructure, and data.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- The Tool-to-Entity Threshold: Parasocial Dynamics of Personalised AI Agents in Shared Social Spaces
- Beyond Anthropomorphism: a Spectrum of Interface Metaphors for LLMs
- Large Language Models are as persuasive as humans, but how? About the cognitive effort and moral-emotional language of LLM arguments
- Conversational Alignment with Artificial Intelligence in Context
- A Taxonomy of Linguistic Expressions That Contribute To Anthropomorphism of Language Technologies
- A meta-analysis of the persuasive power of large language models
- People Defer to AI Moral Advice, But Not Blindly