Module 1 Check · How AI Works — The Guardrailed Classroom

01At its core, what does a large language model do when it generates text?

Looks the answer up in a database of stored sentencesRepeatedly predicts the most likely next token given everything so farRuns a fixed set of if/then rules written by engineersSearches the live web and pastes the top result

02You raise a model's temperature. What happens?

It runs fasterIts answers become more random / variedIt always becomes more accurateIt stops using its training data

03A benefits model uses a decision threshold. You raise the threshold. What's the typical effect?

More people are approvedPrecision goes up but recall goes down (fewer false positives, more false negatives)Both precision and recall always improveFairness is guaranteed

04In a fairness check, what is a 'false negative' for a benefits model?

A veteran who is correctly approvedA genuinely eligible veteran who is wrongly denied / sent to slow reviewA model that runs too slowlyAn ineligible person who is wrongly approved

05Training a model to maximize accuracy can still produce an unfair model. Why?

Accuracy and fairness are the same thingThe most-accurate boundary can still treat two groups unequallyModels can't be unfairFairness only matters during deployment, not training