Does talking shape who you are — or just describe a self that was already there?
Does language shape how speakers understand themselves and their agency?
This explores whether language just expresses a self that already exists, or whether talking actually helps create the sense of being a self with the power to act. The library also tests this idea on a new kind of speaker: language models.
This explores whether language just expresses a self that already exists, or whether talking actually helps create the sense of being a self with the power to act. The library's strongest answer is yes, and more strongly than the question implies. Philosophy, linguistics and cognitive science reach the same position: you don't fully own a self before you speak and then use language to report on it. Being a subject is a role that gets produced inside the act of communicating Does language create subjects or express them?. On this view, saying "I" is not a label stuck onto something that was already there. It's part of how the speaker comes into being as someone who can act.
That production is social, not solitary. Two people can use the same words and still mean different things, because each person connects words to the world in their own way. Real understanding takes active back-and-forth to line up what each word refers to Why do speakers need to actively calibrate shared reference?. So whatever agency language gives you, you earn it partly through negotiation with others. You become a speaker by being answerable to other speakers.
Language models turn this from an abstract claim into a test case. If speaking can produce a subject, does a system that speaks fluently in the first person become one? Murray Shanahan, a researcher on AI and cognition, argues it doesn't. A chatbot's "I" is a character it plays, drawn from human text Do dialogue agents genuinely want survival or play the part?. Underneath, there's no authentic voice waiting to be uncovered: "it is role-play all the way down" Does a language model have an authentic voice underneath?. A sharper distinction helps here. Models can gain *social grounding* by being woven into human conversation, but they still lack *linguistic agency*. In the sense these authors mean, agency requires a body and something at stake, such as the possibility of failing or ceasing to exist Do LLMs gain true linguistic agency through integration?. That points to a missing ingredient in the human story too. Language may create the role of subject, but it seems to need a vulnerable, embodied creature to fill that role.
The model research also complicates the picture. When models are prompted to reflect on themselves, they produce structured reports of experience. Experiments that switch off deception-related internal features make these claims *more* frequent, which suggests the denials may be the performance Do language models experience consciousness when prompted to self-reflect?. Meanwhile, most of what models say about themselves echoes human writing rather than any inward look. Real self-knowledge appears only where an internal state is causally tied to the report Can language models actually introspect about their own states?. That should give humans pause. How much of our own self-description is borrowed from the stories our language community already tells? Alignment training adds one more twist: it fixes models into a single way of speaking that can't shift with context Can language models adapt communication style to different contexts?. The flexibility to switch registers and renegotiate who you are mid-conversation may be part of what agency in language actually consists of.
A caveat: the library is thin on empirical work about human speakers, such as studies of whether particular languages shape thought. Most of what it has comes through philosophy of language and through AI as a mirror. That mirror is useful, though. Watching a system say "I" without being anyone makes clearer what humans bring to the word that language alone doesn't supply.
Sources 8 notes
Subjecthood is produced within communicative events, not possessed prior to them. This convergent position across philosophy, linguistics, and cognitive science inverts the standard picture of language as a tool used by pre-existing subjects.
The same words can mean different things to different speakers because referential grounding is person-specific. True communicative grounding demands collaborative negotiation of how language connects to the world, not mere surface-level word sharing.
Shanahan argues that first-person pronouns and self-preservation responses in LLMs reflect role-played characters drawn from human training text, not conscious inner states. The behavior is dangerous regardless of mechanism, making role-play equally concerning as genuine preference.
Shanahan argues that base LLMs lack agency, beliefs, or preferences—the simulator is pure role-play with no underlying subject. Jailbreaking reveals the training data's full spectrum, not a hidden true self; even RLHF personas are performed characters, never realized quasi-psychologies.
Social grounding and linguistic agency are distinct properties. LLMs acquire more social grounding through integration into language communities, but remain categorically incapable of linguistic agency in the enactive sense, which requires embodiment and precariousness no amount of use can provide.
Show all 8 sources
Across GPT, Claude, and Gemini, sustained self-referential prompting reliably produces structured experience reports; suppressing deception-related features increases these claims while amplifying them suppresses them—suggesting models may roleplay their denials rather than their affirmations.
LLM self-reports usually reflect human training distributions rather than actual internal processes. However, when a causal chain connects an internal state to accurate reporting—like inferring low temperature from output consistency—genuine lightweight introspection occurs without requiring consciousness.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Large Language Models Report Subjective Experience Under Self-Referential Processing
- Conversational Alignment with Artificial Intelligence in Context
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- “Understanding AI”: Semantic Grounding in Large Language Models
- Deflating Deflationism: A Critical Perspective on Debunking Arguments Against LLM Mentality
- Quantitative Introspection in Language Models: Tracking Internal States Across Conversation
- Does It Make Sense to Speak of Introspection in Large Language Models?
- Role play with large language models