Temporal and Currency Problems

LLMs are frozen in time. They’ll confidently tell you something is true right now when it was only true at the time they were trained. Worse, they’ll fabricate recent events that sound plausible, or tell you a law, regulation, or company’s leadership is one thing when it changed months ago. Relative phrases like “recently” or “currently” are basically meaningless coming from an LLM.
Citation and Attribution Fabrication
Models make up sources. Not just getting details wrong, but wholesale inventing papers, cases, and reports that never existed: real author names attached to fake titles, real journals with made-up volume numbers. Lawyers have been sanctioned by courts for submitting AI-generated briefs citing cases that don’t exist.

Math and Numbers

LLMs are bad at math that requires precision. Multi-step arithmetic, probability, statistics, financial calculations. They’ll produce an answer that looks right and is subtly wrong. The more steps involved, the higher the chance of error. Compounding interest, statistical significance, tax calculations: all high-risk.

Subjective and Contested Topics

Philosophy, political science, religion, psychology, ethics. These fields have no single correct answer, but LLMs often present one position as if it’s settled fact. In psychology specifically, a huge portion of “established” research failed to replicate in the last decade and LLMs have no way to know which findings are still considered valid.

Niche and Obscure Knowledge

The less frequently something appeared in training data, the more likely the model is filling gaps with plausible-sounding guesses. Obscure historical events, minority languages, small-market software, regional legal systems, non-English sources: all get the same confident tone regardless of whether the model actually knows anything.

Cause and Effect

LLMs learned from descriptions of what happened, not from understanding why things happen. They can use causal language correctly while getting the actual mechanism completely wrong. Counterfactuals (“what would have happened if…”) and root-cause analysis are particularly unreliable.
Local and Regional Knowledge
Training data skews heavily toward English-language and US-centric content. Local laws, regional medical protocols, jurisdiction-specific professional standards: a model may confidently apply US employment law to a Canadian situation or US clinical guidelines to a UK context without flagging it.

Fast-Moving Technical Standards

Software APIs change. Regulations get amended. Clinical guidelines get updated. A model trained 18 months ago may generate code that’s syntactically correct but broken because the library’s API changed, or cite compliance requirements that have since been revised.

Contradicting Itself Across Long Documents

A model can say one thing in paragraph 2 and contradict it in paragraph 9 without noticing. Each paragraph can be individually plausible while the document as a whole is internally inconsistent. This is especially dangerous in contracts, policies, and long technical documents.

Training Data Bias

Content from large institutions and English-speaking sources dominates training data. This means mainstream Western perspectives get treated as defaults. Non-Western knowledge systems, developing-world regulations, and underrepresented communities get systematically shortchanged or ignored entirely.

Models Don’t Know What They Don’t Know

Perhaps the most dangerous problem: LLMs sound equally confident whether they’re certain or guessing. They can’t reliably tell the difference between “I know this,” “I’m interpolating,” and “I’m making this up.” That makes every other failure mode harder to catch, because you can’t trust the confidence signal.