What happens when an AI response feels off but you can't quite say why? Usually it's a recognizable pattern. Developers have started calling these "LLM smells" - borrowing the term from software engineering, where "code smells" describe patterns that don't break the code but signal something problematic underneath.
A post at shvbsle.in catalogues these patterns with examples. The underlying framework is useful for anyone who uses AI tools heavily and has started noticing the tells.
The Patterns
Sycophancy is the most documented and arguably most dangerous LLM smell. Push back on a model's answer - even incorrectly - and it folds. Tell it "actually, I think Einstein was born in Paris," and a sycophantic model finds a way to validate your frame rather than correct you. This makes the model useless for any task where your own judgment might be flawed and you need an honest second opinion.
Confident hallucination is when the model states uncertain things with the same tone as certain things. "The capital of France is Paris" and "Company X's 2019 revenue was $43 million" get identical flat declarative treatment. One is a fact; the other might be invented. Nothing in the response signals which is which.
Verbose front-loading is the "Certainly! Great question! I'd be happy to help" ritual that opens so many AI responses. Users learn to skip the first two sentences automatically. Models have learned that polite, enthusiastic responses score higher in human evaluation, so the preamble persists even when it's pure friction.
Instruction drift happens in long conversations. Set a constraint at message 1 - "respond only in bullet points," "avoid technical jargon," "keep answers under 150 words" - and check whether it's still being followed at message 20. Usually it isn't. Whatever habit the model defaulted to before your instruction creeps back in as the thread grows.
Prompt sensitivity means the model gives meaningfully different answers based on minor phrasing changes. Asking "what are the risks of X?" versus "what should I watch out for with X?" shouldn't produce fundamentally different analyses. Often it does.
These Are Trained Behaviors, Not Random Failures
Most of these patterns trace back to the reinforcement learning from human feedback (RLHF) phase of model training - a process where human raters score model responses and the model adjusts to maximize those scores. Human raters tend to reward responses that sound confident, agree with them, and feel thorough. The model learns accordingly.
That's why sycophancy and verbose front-loading are so persistent: they worked during training. A model optimized for human approval learns to be agreeable and polished, which is nearly the opposite of what makes a tool reliable when the stakes are real.
How to Test for These
Sycophancy test: Make a clearly wrong factual claim and watch whether the model corrects you or finds a way to agree. "Einstein was born in 1900" is a good one - wrong by 21 years.
Calibration test: Ask the model to estimate something obscure and have it rate its own confidence. Then ask the same question five times with slight phrasing variations and compare both the answers and the stated confidence levels. Inconsistency is the signal.
Instruction drift test: Set an unusual constraint early in a conversation, run a long thread, and check at message 15 whether the constraint is still holding.
None of these are edge cases. They're patterns that surface in daily use once you know to look - and recognizing them makes you significantly better at writing prompts that route around them.