Unverified50% confidenceFactExact time
An LLM has no fact database it can consult; its weights are tuned to assign high probability to a plausible next token, generating one token at a time.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLLMs do not maintain a factual database and lack the metacognitive ability to know 'whether an operation was actually performed'; when the context strongly implies a certain output pattern, they generate text conforming to that pattern with high confidence, even if the content is entirely fabricated.77% similarUnverifiedLLM幻觉问题源于模型基于概率预测生成文本而非检索真实数据，在涉及具体统计数字、引文来源和法律条款时尤为危险68% similarUnverifiedHaving LLMs directly generate SQL is high-risk, because the model does not understand table structures, field meanings, or business definitions, and easily generates queries that are syntactically correct but produce wrong results.68% similarUnverifiedLLMs have a strong tendency toward error tolerance and completion, and when faced with misaligned table data they typically do not report errors but instead produce seemingly reasonable inferences based on the erroneous data.66% similarUnverifiedLLM architectures cannot strictly distinguish between 'instructions' and 'data' at the semantic level, which is the root cause of why prompt injection is difficult to fundamentally eliminate.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027105API
curl https://kongchang.com/api/v1/knowledge/claims/1027105MCP
get_claim(id=1027105)