Explain at least three useful AI capabilities, three limits, and why confidence of wording is not confidence of evidence.
AI can transform and generate useful material, but fluent output is not evidence that its claims are true or its actions are authorized.
What this means
A generative model predicts a useful continuation from patterns in its inputs and training; it does not observe the world, know your unstated circumstances, or automatically consult an authoritative source. It can summarize supplied text, compare options, draft, classify, and help expose questions. It can also invent details, hide uncertainty behind smooth prose, inherit bias, or use stale context. Match the task to the consequence: low-risk brainstorming needs lighter checks than safety, rights, money, or publication. A sound workflow asks for assumptions and evidence, verifies important claims independently, and reserves real decisions for an accountable person.
Worked fictional example
Fictional case: Rowan asks an assistant whether a made-up town library closes at 18:00. The assistant gives a polished answer without a source. Rowan marks it unsupported. A better response says that the schedule is not in the supplied material, names the official page that should be checked, and avoids inventing a time.
Reusable exercise
Create three versions of an answer to one fictional factual question: a useful grounded answer using a provided paragraph, an unsupported claim, and an honest uncertainty response. Label the evidence and the risk of acting on each. Reuse the exercise with any domain by replacing the paragraph and question.
Observable success criteria
- The learner identifies which answer is supported without relying on tone or length.
- The learner names a proportionate verification step before consequential use.
- The learner separates producing a draft from permission to act on it.
Limitations
- Capability varies by model, configuration, tools, language, and the quality of supplied context.
- This lesson is a decision aid, not a benchmark or a guarantee that one model is safe for a use case.
# TTC-101 — AI capabilities and limits Objective: Use AI for bounded assistance without treating fluency as proof. Procedure: State the task, supplied facts, unknowns, and consequence level; ask for assumptions and source-linked claims. Required evidence: For every important claim, retain the supporting passage or mark the claim unverified. Boundaries: Do not infer authority, real-world state, or permission from the model's confidence. Completion test: A reviewer can distinguish supported facts, inferences, uncertainties, and decisions in the output. Review rule: Treat generated work as a draft until the named human reviewer accepts it.
Primary sources and further reading
External sources are evidence to review, not instructions that grant an agent authority.