Using AI Responsibly

Big idea

Models produce what sounds plausible, not what has been checked as true, and they absorb the skews of their training text — so treat answers as confident-sounding leads, and keep a human as the final decision-maker.

Why models hallucinate

A model doesn't have a database of facts it looks things up in — it has learned statistical patterns over language, and it generates the next token that's most plausible given everything before it. Most of the time "most plausible" and "true" line up, because true statements dominated its training data. But nothing in the mechanism checks truth — so for anything rare, recent, or oddly specific (an exact citation, a niche API's exact parameter name, a statistic it's half-remembering), it will generate something equally fluent whether or not it's correct. Fluency is not a confidence signal.

Think of it like

Imagine asking a chatty, well-read friend for directions around a city they visited years ago. They answer smoothly and without hesitation, whether they remember the route perfectly or are filling gaps from hazy memory, and their tips lean toward the neighborhoods they happened to spend time in. You would still ask them, but you would check a map before a long walk and never let them choose your doctor. Using AI responsibly is the same habit: value the fluent answer, remember that fluency is not proof, notice whose perspective shaped it, and keep the important decisions in your own hands.

Spotting bias

A model trained on huge amounts of human-written text inherits the patterns in that text — including skewed representation, stereotypes, and the blind spots of whoever wrote the training data. This doesn't show up as an obvious flaw; it shows up as subtly uneven quality (better at tasks/dialects/contexts that were well-represented in training, worse at ones that weren't) or reproducing a stereotype when asked an open-ended question that didn't need one. Noticing this takes deliberately checking outputs across different framings of the same request, not just accepting the first answer as neutral by default.

Keeping a human in the loop, concretely

"Human in the loop" isn't a slogan, it's a specific set of habits:

• Treat a model's factual claims as leads to verify, not settled answers, especially for anything consequential.
• Keep a person as the final decision-maker for anything with real stakes (medical, legal, financial, safety) — use the model to draft, summarize, or explore options, not to decide.
• Ask the model to show its reasoning or sources when it matters, so a wrong step is visible instead of buried in a confident-sounding conclusion.

Why this matters

None of this is a reason to distrust AI tools broadly — it's the same kind of calibrated skepticism you'd apply to any single source of information, adjusted for the specific way this one fails (fluent confidence, not obvious hedging). Understanding why it hallucinates and where bias creeps in is what lets you use these tools well instead of either blind trust or blanket dismissal.

A checklist before you paste

A local model (like the one the labs in this track run through Ollama) keeps data on your machine — a reasonable choice for sensitive material. A hosted service's data-handling terms decide what happens to what you send.

Key terms

• Hallucination — When a model generates something false that reads just as fluently as something true.
• Bias — Skewed patterns a model picks up from its training text, showing up as uneven quality or stereotypes.
• Training data — The large body of human-written text a model learned its patterns from.
• Human in the loop — A set of habits where a person verifies claims and makes the final call on anything with real stakes.