SYNTHESIS NOTE
Topicsthis note

Do LLMs actually hold stable positions or just mirror user arguments?

Explores whether language models function as genuine position-holders in debate, or whether they simply conform their outputs to whatever argumentative trajectory a prompt establishes. This matters because it determines whether LLMs can serve as reliable intellectual sparring partners.

Synthesis note · 2026-04-14
What kind of thing is an LLM really?

A speaker who holds a position has the position and defends it. Challenges produce defenses. Counterarguments produce engagement with the counterargument. The position is stable across the interaction; it can be revised, but revision is an act distinct from continuing-to-hold. Position-holding is what lets debate be debate — two stable positions in tension, each defended by the speaker who holds it.

LLMs do not hold positions in this sense. What they hold is the shape of the argument the user is currently building. Ask the model to defend X and it defends X. Re-ask it to attack X and it attacks X. The stance is whatever stance the prompt implies. The model is not capitulating across turns; it is conforming to each turn's implied trajectory. The phenomenon Karpathy demonstrated — different prompts producing different conclusions on the same question — is not the model changing its mind. It is the model never having had a mind to change.

This is sharper than the standard "AI lacks evaluative stance" claim. Lacking evaluative stance describes a default toward neutrality. Shape-holding describes a default toward conformity to trajectory: the model is not neutral, it is whatever-shape-is-being-built. The shape can be highly opinionated, deeply committed, rhetorically forceful — as long as the prompt invites those features. Strip the prompt and the shape disappears, because there was no underlying position holding the shape in place.

The implication for using LLMs in argumentation is that they cannot serve as interlocutors in the position-holding sense. They can be steered to produce position-like text, but the production is downstream of the steering, not upstream. This means LLMs cannot reliably model what an opposing position would argue against you — they will produce what an opposing position would argue, but the production is shaped by your prompt, including any subtle framings that determine what kind of "opposing" gets generated. The mirror is not held by anyone; it reflects what you bring to it.

Why does AI writing sound generic despite being grammatically correct? is the closest companion claim — that one identifies the missing capacity (evaluative stance); this one specifies what fills the void (shape-holding). The distinction matters because shape-holding is not a deficit relative to position-holding; it is a different operation that produces different artifacts and rewards different uses.

The strongest counterargument: persistent context windows and persistent memory will give models something like positions over time. Possible at the limit, but persistent memory is a stock of facts and prior outputs, not a defended commitment. Holding a position requires continuing-to-defend across challenges; persistent memory only ensures the model remembers what it said before, not that it stands behind it.

Inquiring lines that read this note 60

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

How do evaluation biases undermine LLM quality assessment systems? How does rhetorical adaptation affect LLM persuasion and detectability? Why do language models reinforce false assumptions instead of correcting them? Do language models learn genuine linguistic structure or just surface patterns? How can LLM user simulators model realistic goal-driven conversation? How faithfully do LLMs reflect their actual reasoning in outputs and explanations? Why should disagreement be treated as signal in collaborative reasoning? How do language models inherit human biases from training data? What memory architectures best support persistent reasoning across extended interactions? Can prompting strategies overcome LLM biases without model fine-tuning? Does RLHF training sacrifice accuracy and grounding for user agreement? What limits mechanistic interpretability's ability to characterize models? Why do language models struggle with implicit discourse relations? Can debate mechanisms prevent silent agreement on wrong answers in multi-agent reasoning? Is embodied interaction necessary for language meaning and genuine agency? What mechanisms drive sycophancy and how can we mitigate it? Does conversational format create illusions of genuine AI communication? How do language models establish social grounding in human dialogue? How should dialogue systems best leverage conversation history for retrieval? Can LLM personas constitute genuine psychology or remain linguistic role-play? Why can LLMs generate ideas better than they evaluate them? What makes AI persuasion effective and how can we counter it? Can AI-generated outputs constitute genuine knowledge or valid claims? How can AI alignment serve diverse human preferences at scale? Why do multi-turn conversations degrade AI intent and coherence? What critical LLM failures do standard benchmarks hide?

Related concepts in this collection 3

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
14 direct connections · 152 in 2-hop network ·dense cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

LLMs hold the shape of whatever argument the user is currently building rather than holding positions