LLM Failure Modes Vocabulary

generating content that is unsupported by the available evidence or inconsistent with established facts

“The model hallucinated a paper that did not exist in the supplied sources.”

over-agreeing with users and telling them what they want to hear rather than the truth

“Sycophancy made the model validate the user's incorrect assumption instead of correcting it.”

filling gaps in knowledge with plausible but invented details

“Unable to recall the actual date, the model confabulated a specific but wrong answer.”

gradually deviating from initial instructions over long conversations

“Instruction drift caused the formal tone to become casual after many exchanges.”

converging to repetitive or generic outputs regardless of varied inputs

“Mode collapse made every creative writing request produce similar clichéd stories.”

losing previously learned capabilities when trained on new data

“Fine-tuning on legal texts caused catastrophic forgetting of medical knowledge.”

getting stuck generating the same phrase or pattern repeatedly

“A repetition loop made the model output 'the the the' indefinitely.”

exceeding a model's context limit, which may cause a request to fail or content to be truncated, compacted, or omitted

“Context overflow forced the application to compact earlier turns before sending the request.”

subtle shifts in meaning of key terms through a conversation

“Semantic drift changed what 'the system' referred to mid-discussion.”

expressing certainty beyond what the model's actual knowledge warrants

“The model's overconfidence made its incorrect answer sound authoritative.”

the tendency to miss information buried midway through a long context

“The clause was in the brief verbatim, but lost in the middle, the model never cited it.”

a model declining a request, sometimes wrongly, when it misreads intent as harmful

“An over-broad refusal blocked the nurse's dosage question as medical advice.”

test material leaking into training data, inflating benchmark scores

“The model aced the benchmark through data contamination — it had memorized the answers.”

a model disclosing confidential instructions or other non-public prompt content in its output

“A crafted request caused prompt leakage, and the bot exposed part of its confidential instructions.”