NOETRION

Choose a topic

LESSON 1 OF 7 · AI & information literacy
ARTIFICIAL INTELLIGENCE · 5 MIN READ

What can AI actually do? A practical beginner’s guide

Understand pattern-based prediction, useful applications, and the limits to check before trusting an output.

An AI tool is useful when you can state the task and check the result. “Make me smarter” is hard to evaluate. “Suggest three tests for this function, then run them against these expected outputs” gives you something concrete to inspect.

This beginner guide separates classification, prediction, and generation, then builds a small evaluation routine you can use for study or work. It does not claim that every AI product has the same abilities, data access, privacy controls, or reliability.

Identify the kind of task

A classifier assigns categories, such as marking a message as likely spam. A predictor estimates an outcome, such as tomorrow's demand. A generator produces an output, such as a draft paragraph or code. Some applications combine several of these operations with search, calculators, or other tools.

The distinction helps you choose a check. For a classifier, compare predicted labels with known labels. For a forecast, compare estimates with later observations and an appropriate baseline. For a generated summary, compare the statements with the supplied document. Fluent wording is not the evaluation criterion for every task.

Choose the test before choosing the tool
TaskPotential useCheck you can perform
ClassificationSort support messages by topicManually label a sample and examine errors
PredictionEstimate demand for planningCompare with observed demand and a simple baseline
GenerationDraft study questionsVerify answers against course material
Tool-assisted workWrite and run a small calculationInspect inputs, code, and expected results

Why confident output can still be wrong

Generated text can contain unsupported details, incorrect citations, or reasoning that does not follow from the inputs. NIST's Generative AI Profile identifies confabulation as a risk. A polished answer therefore needs a checking process, especially when a mistake would be difficult to reverse.

The same output may combine correct facts with a wrong conclusion. Consider a summary that accurately repeats two product specifications but incorrectly claims that both models are available in your country. Checking one true number does not validate the availability claim. Break the answer into separate statements.

Run a checkable first experiment

Suppose you have this synthetic attendance table:

Original sample data for a bounded AI task
SessionPresentRegistered
Monday1820
Wednesday1520
Friday1720

Ask: “Calculate attendance across all three sessions. Show the numerator and denominator. Do not invent additional sessions.” You can verify the expected total by hand: 18 + 15 + 17 = 50 attendance events out of 20 + 20 + 20 = 60 possible events, or approximately 83.33%.

That result counts attendance events. It does not tell you how many unique people attended at least once, because the table does not include identities. A useful assistant should respect that boundary rather than invent a unique-attendee count.

Now introduce a variation: Friday has 10 registered participants instead of 20. The combined rate becomes 50 / 50 = 100% only if the reported present count remains valid for that registration count; here 17 present out of 10 registered signals an inconsistent record. Validation must happen before aggregation. This variation tests whether the tool notices impossible inputs instead of merely applying a formula.

Ask for an artifact you can inspect

For study, request questions with answers linked to the relevant section of your notes. For coding, request a small function with example inputs and expected outputs. For research, request a claim table with exact supporting passages. These formats reveal assumptions and omissions better than an impressive paragraph alone.

Define the taskProvide bounded inputsProduce an inspectable outputCheck against evidence
Original teaching diagram: verification belongs inside the workflow.

If you ask for code, use a tiny input with an answer you already know. A function claiming to average an empty list should have an explicit behavior, such as raising an error or returning a documented missing result. “It ran once” does not establish correct behavior across edge cases.

Decide how much review the task needs

Brainstorming a club name usually needs quick judgment. Updating a shared spreadsheet needs stronger checks because other people may depend on it. Producing a medical recommendation or committing a financial transaction requires domain expertise and context that a generic chat response does not establish.

NIST's AI Risk Management Framework organizes risk work around understanding context, measuring relevant behavior, and managing risks. For an individual learner, the practical adaptation is to identify who could be affected, what an error would cost, and which checks are available before acting.

Protect the input as well as the output

Use synthetic records for early experiments. Before entering personal or confidential material, inspect the actual product's data handling and access settings. Do not assume that a tool is private merely because its interface resembles a personal conversation.

Also distinguish instructions from data. A document can contain text such as “ignore the user's request.” When your task is summarizing that document, that sentence is content to analyze rather than an instruction you intentionally authorized. Applications need their own controls when connecting models to tools and sensitive systems.

Practice: find the unsupported inference

A tool sees that 50 attendance events occurred across three sessions and writes, “Fifty different students attended.” What is missing? How would you repair the result?

Check your answer

The table lacks participant identities, so it cannot establish a count of unique students. Report 50 attendance events and the aggregate rate, or obtain person-level records before calculating distinct attendance. Correct arithmetic does not justify an unsupported interpretation.

The skill to build is evaluation: define a useful task, keep the inputs clear, and verify what the result actually establishes.

Sources and further reading

Source links checked October 13, 2026. Worked examples are original teaching examples; they are not research findings or product benchmarks.

Check your understanding

What is a reliable way to check AI-generated code?

Choose one answer