Course 0 · Lesson 08/10 · Interactive guide

What AI Is Good and Bad At

AI can repeat pattern-based work at scale, but a confident-looking answer is not proof. The right check depends on what happens if it is wrong.

Lesson
08 / 10
Time
About 8 minutes
Level
Beginner · 0/5
Prerequisite
Lesson 07 · Multimodal AI

01 · Set the right expectation

Good at one task does not mean good at everything

An AI model is trained for patterns in data and a defined objective. A model that separates cat photos does not know how to plan a school trip, and a model that writes fluent sentences can still get arithmetic or facts wrong.

Think of a picture-card sorterAfter seeing varied, clear examples, it can repeat the same sorting task quickly. Give it a wet card, a dark photo, or something never seen before, and its old patterns may send the card to the wrong pile.

A better question than “Is AI smart?” is “What task does this system do, what data shaped it, and what happens when it is wrong?”

02 · Strengths

AI is useful when a task contains many repeated patterns

Pattern

Finding patterns

Notice similarities across images, audio, text, or tables of data.

Rank

Ranking

Order search results or other items by likely relevance.

Predict

Predicting

Estimate a value such as travel time or future demand from past data.

Classify

Classifying

Choose a category such as a document type or spam/not spam.

Transform

Transforming

Transcribe audio, summarize, or change content into a more useful form.

Scale

Repeating at scale

Process many examples with the same steps, while still requiring quality checks.

The strength has a cost: speed and scale can reduce repetitive work, but a small error can also be repeated at scale.

03 · Make a decision

How convincing an answer looks is not how much you should trust it

Choose a situation, then choose a verification level. The consequence changes with the risk and the evidence you add.

System answer

“No package damage” — high confidence score

But glare and the camera angle hide a crushed corner, unlike most training photos.

How much will you verify?

Decision trust: low

Acting quickly on unfamiliar data

A damaged box may move forward. A confidence score says which output the model favored; it does not prove that output is correct.

04 · A deeper layer

Important limits have names

Out-of-distribution (OOD)New input differs from training data—for example in lighting, place, language, or a new kind of event.
Missing contextA system only sees what it receives, so it may lack intent, history, or exceptions a person knows.
Correlation ≠ causationTwo things appearing together does not prove that one caused the other.
Confident errorWording or a confidence score can look certain while the answer is wrong.
Data biasMissing groups or unfair past decisions in examples can carry bias into outputs.
BrittlenessA small input change may cause a large output change, especially beyond tested conditions.

Match the check to the risk: a low-impact task may need sample checks. Work affecting money, rights, safety, or opportunity needs direct evidence, a way to appeal, and a person accountable for the decision.

05 · Remember this

Use the strengths without hiding the limits

  1. 1

    Tie ability to a specific task. Ask what the system ranks, classifies, predicts, or transforms.

  2. 2

    Confidence is not correctness. Check evidence, data conditions, and the cost of a wrong answer.

  3. 3

    Verify according to risk. Higher impact needs source material, testing, and human oversight.

06 · Knowledge check

Answer all three

1. Which task best matches a common AI strength?

2. What does a high confidence score mean?

3. What should happen when an output affects rights or safety?