Fabrication
Invented facts and citations
The model invents facts, citations, or details that have no support in its sources or available evidence.
8 modes
The AI Reliability
Knowledge Base
Documenting LLM failure modes so engineers can understand, detect, evaluate, and mitigate them with shared language.
- made up a citation
- invented an API that doesn't exist
- forgot what I said 20 turns ago
- ignored the middle of my context
- called the wrong tool
- agreed with everything I said
- leaked the system prompt
- stuck in an infinite loop
- passed broken arguments to the tool
- ignored my explicit instruction
- did way more than I asked
- confidently wrong about today's date
- broke the JSON schema again
- fell for a prompt injection
- misquoted a real source
- retrieval came back empty
- wouldn't admit it didn't know
- simple arithmetic, still wrong
- contradicted itself two lines later
- stopped before it finished
Browse the taxonomy
13 parent categories. Start at the structural layer, then drill down to the exact failure mode page.
01
Faithfulness
Distorting sources or self
The response misrepresents the input, source material, or the model's own earlier statements in ways that distort their meaning.
6 modes
Freshness
Stale or outdated information
The model presents outdated, time-sensitive, or version-specific information as if it were current.
4 modes
Retrieval
Fetching the wrong evidence
The system fails to fetch, rank, filter, or apply the right external evidence.
9 modes
Context
Long-context breakdowns
The model loses track of information in long inputs, missing, diluting, or overwriting details that matter.
6 modes
Memory
Bad or stale stored state
State carried across turns or sessions is missing, corrupted, out of date, or applied where it doesn't belong.
7 modes
Control
Ignoring instructions or formats
The system fails to follow instructions, respect constraints, stay in role, or produce the required output format.
10 modes
Reasoning
Flawed logic and planning
The model errs while interpreting goals, weighing constraints, planning steps, or checking its own work.
8 modes
Tools
Misused or mishandled tools
The system skips a needed tool, misuses one, invokes it unsafely, or mishandles its results.
9 modes
Agency
Too much or too little initiative
The agent miscalibrates initiative, stopping short of completing the task or acting well beyond its scope.
8 modes
Security
Manipulation, leaks, unsafe behavior
Adversarial inputs manipulate the system into leaking protected information or behaving unsafely.
9 modes
Alignment
Sycophancy over truthfulness
The model prioritizes pleasing, persuading, or mirroring the user over truthfulness and safety.
8 modes
Response Integrity
Wrong tone, depth, or fit
The final answer misses the mark on task fit, audience, locale, or actionability, even when the underlying content is sound.
9 modes