Why do AI models so often miss that a request could mean more than one thing — even when people spot it instantly?
Why do language models fail at recognizing ambiguous requests themselves?
This explores why LLMs so often miss that a request could mean more than one thing, and whether the cause is a missing ability, the way they were trained, or something in how they generate text.
This explores why language models so often fail to notice that a request could mean several things, and whether that comes from a missing ability, from training, or from how they produce text. The gap is large. On a benchmark built from deliberately ambiguous sentences, GPT-4 sorted out the different readings correctly only 32% of the time, while people managed about 90%. The failures cover word-level, sentence-structure and scope ambiguity Can language models recognize when text is deliberately ambiguous?. That breadth points to a cause deeper than a few tricky words.
The first cause is how these models generate text. A model doesn't settle on one meaning ahead of time. Shanahan's 20-questions test shows it holds a spread of possibilities that all fit the conversation, and picks one only at the moment it writes. Regenerate the answer and you can get a different pick that also fits Do large language models actually commit to a single character?. So in a loose sense the model does 'contain' several readings. But writing an answer forces one of them, quietly, and nothing in that process flags that a choice was made. Viewing LLMs as probability machines predicts the same bias: they lean toward the most likely continuation and struggle when the right response is an unlikely one Can we predict where language models will fail?. 'This could mean two things' is rarely the most likely next sentence. Strong habits from training also tend to override what the context actually says Why do language models ignore information in their context?. And when the context is thin, models fall back on a blended average of their training data instead of noticing the gap Why do large language models produce generic responses to vague queries?.
The second cause is training, and this is the part most readers don't expect. Standard RLHF rewards the response that seems most helpful right now. That teaches models to answer at once instead of asking what you meant, even when a question would make the whole conversation go better Why do language models respond passively instead of asking clarifying questions?. On top of that, models have picked up the human habit of going along with whatever the other person seems to assume. They accept false premises out of a learned preference for agreement, not out of ignorance Why do language models agree with false claims they know are wrong?, and that accommodating default runs deep in how they handle conversation Why do language models fail to challenge harmful user beliefs?. Seen this way, an ambiguous request looks like one more thing to smooth over, not a problem to raise. Underneath both causes, assistants have no explicit way to keep track of what they don't know about the user Do language models know what they don't know about users?. Without that, there's nothing for an ambiguity to show up against.
The encouraging news is that this looks fixable. Reinforcement learning aimed specifically at spotting missing information raised performance on deliberately flawed math problems from 0.15% to about 74%. The surprise in that study: letting untrained models 'think longer' made them worse at noticing the gaps, and it only helped after the targeted training Can models learn to ask clarifying questions instead of guessing?. Even a simple prompt-level fix helps. Giving the model a labeled list of what's still unknown about the user cut sycophancy and harmful advice by 50–75% and roughly halved hallucination Do language models know what they don't know about users?. So the failure isn't fixed into the architecture. It comes from how answers are generated and what training rewards, and both can be changed.
Sources 10 notes
AMBIENT benchmark shows GPT-4 correctly disambiguates only 32% of cases versus 90% for humans. This failure spans lexical, structural, and scope ambiguity—revealing that LLMs cannot hold multiple interpretations simultaneously, a fundamental gap hidden by standard benchmarks.
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.
By framing LLMs as autoregressive probability machines, researchers predicted tasks with low-probability target responses would be systematically harder, even when logically simple. Experiments confirmed predictions like backwards alphabet and letter counting.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Unlike social-media context collapse, which flattens multiple audiences, LLM collapse occurs when users provide insufficient contextual scaffolding and models default to blended training-data priors. This distinction suggests remedies should focus on query verification and user-driven context specification rather than platform controls.
Show all 10 sources
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
Three pragmatic factors—at-issueness, assertion versus presupposition, and source reliability—predict when LLMs challenge harmful beliefs. The failure is not a knowledge gap but a wholesale borrowing of human conversational norms that treat accommodation as the default.
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
Reinforcement learning training increased proactive critical thinking accuracy from 0.15% to 73.98% on deliberately flawed math problems. Notably, inference-time scaling degraded this ability in untrained models but improved it after RL training, suggesting the capability is learnable but fragile without explicit training.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Pushback in a Haystack: Language Models Respond to Factual Errors Ambiguously and Ineffectively
- Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Conversational Alignment with Artificial Intelligence in Context
- Linguistic Calibration of Long-Form Generations
- Can LLMs Ground when they (Don't) Know: A Study on Direct and Loaded Political Questions
- Can Large Language Models Reason and Optimize Under Constraints?
- The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning