Hallucination
Hallucination is when a language model produces text that sounds confident and coherent but is factually wrong. The model invents details, citations, or function arguments that have no basis in its training data or the conversation context.
Why it matters
Hallucination is the dominant failure mode of modern LLMs and the reason "the model said it" is not a substitute for verification. A hallucinated fact is indistinguishable in tone from a verified one; the model does not know it is wrong.
For chat products, hallucination is annoying. For agent products that take real action, hallucination is dangerous. An agent that hallucinates a customer's email address sends the email to the wrong person. An agent that hallucinates a payment amount sends the wrong invoice. Bounded blast radius for consequential actions is the defense; verification through real tool calls is another.
Hallucination is reduced (not eliminated) by grounding the model in real data, calling tools rather than describing what tools would do, requiring citations for factual claims, and approval-gating consequential actions.
The 2026 frontier models hallucinate meaningfully less than 2023 models did, but the floor is not zero. Any agent product running in production has to assume the model can hallucinate at any point and design the operational guardrails accordingly.
How Ajeris handles this
Three layers of defense:
- Tool calls over assertions. When the user asks "what is my balance," the agent calls the bank API rather than asserting a number. The result is real or the agent says so.
- Approval gating on consequential actions. Money movement, public posting, and similar actions queue and wait for the user to authorize before firing. A hallucinated action does not execute silently.
- Memory citations. When the agent recalls something from long-term memory, the source is the memory, not the model's assertion. Memory misses are visible (the agent says "I do not remember") rather than masked by hallucination.
For most users, these layers are invisible until something goes wrong, at which point the bounded blast radius is the difference between "annoying" and "catastrophic."
