Learn / AI Literacy Bridge / Unit 1: What AI is — and what it isn't

Strengths and blind spots

Lesson 3 of 3 · 15 min read

Modern AI is remarkably good at tasks where there is a lot of example data and where being approximately right is useful: transcribing speech, translating, summarising, drafting, spotting patterns in images.

It struggles in predictable places:

  • Out-of-distribution inputs. A model trained on daytime road photos may fail at night. When the situation looks different from the training data, performance can drop sharply — often without warning.
  • Exact facts and arithmetic. Language models predict likely words; they do not look facts up unless connected to a tool that does. They can state wrong dates, invent citations, or miscount.
  • Causation. Learning that two things occur together is not the same as knowing one causes the other.
  • Accountability. A model cannot be responsible for a decision. People and institutions remain responsible for how it is used.

A useful habit: before trusting an AI output, ask “How would I know if this were wrong?” If the answer is “I wouldn't”, the task needs a check — a source, a test, a second opinion — before the output is used for anything that matters.

Throughout this university you will see this habit built in: every course ends with work that is checked, not just produced.