The failure modes every user must know
Because the model predicts plausible text, it can produce fluent statements that are simply untrue, with complete confidence. Causes include gaps in training data and prompts that assume facts the model fills in to be helpful. Fluency is not evidence of truth.
Models learn from human-produced data, which carries human biases. If historical hiring data favoured one group, a model trained on it can reproduce that preference at scale. Bias is not a glitch in the data; it is the data, faithfully learned.
Step one: check whether the claim is the kind of thing the model could know. Step two: look for a primary source - a document, a statute, an official page. Step three: cross-check with a second independent source or tool. Step four: sanity-check numbers with your own arithmetic. Four steps, twenty seconds, and you stop being a victim of confident nonsense.
Fluent, confident output that is factually false.
Systematic unfairness learned from training data.
Tying model output to verified source documents.
Check, source, cross-check, arithmetic.
Tools: Any free chatbot + a search engine
