What Makes a System AI?

Describe a system through its goal, inputs, method, output, and effect before choosing any label.

Start with the system

People often begin with a label such as chatbot, smart camera, or AI assistant.

We will begin with five questions instead.

Imagine that we hide the product name and logo. We can only study how the system works. What should we ask?

Five questions

1. What is the goal?

What is the system meant to help us do?

It may filter spam, estimate travel time, write a reply, suggest a product, or find payments that need review.

If the goal is not clear, say that it is not clear.

2. What goes into the system?

These are the inputs.

Inputs may include text, pictures, audio, sensor readings, account activity, instructions, or documents found by a search tool.

A system cannot use information that it does not receive. It may also fail to use some information that it does receive.

3. How does it produce the result?

This is the method or mechanism.

Look for facts about:

• Rules written by people.
• Search or document retrieval.
• A model trained from examples.
• A neural network.
• A model that creates new content.
• Tools such as calculators or databases.
• Several steps and actions.
• Review or approval by a person.

A real product may use several of these parts.

4. What comes out?

This is the output.

An output may be a score, category, prediction, suggestion, new piece of content, decision, or action.

There is an important difference between a suggestion and an action.

This payment looks unusual is a suggestion or warning. Freezing the account is an action.

5. What can change because of the result?

The result may change what a person sees or does. It may update a record or control a machine. It may also affect an important decision.

Some changes are easy to undo. Others are not.

Worked example: a spam filter

Let us describe a spam filter.

Goal: keep unwanted or harmful messages out of the inbox.

Inputs: message text, sender details, links, and attachments.

Method: a model trained from examples may give each message a spam score. A fixed rule may also block known harmful senders.

Output: a score or category.

Effect: the message may go to the inbox, spam folder, or a review queue.

Now we have useful facts. If training from examples is confirmed, we can call that part machine learning. We can also see that the full product may use both a trained model and fixed rules.

A friendly calculator

Imagine a calculator with an animated face. It says:

I have thought carefully. The answer is 42.

The calculator still follows a set method. Its face and words do not prove that it uses machine learning.

The way a product speaks is not proof of how it works.

Your turn

Read this description:

A university service receives an essay. It finds similar passages in its document collection. It returns the matching passages, links, and similarity scores.

Let us answer the five questions.

Goal: find passages that may need a closer look.

Inputs: the essay and the document collection.

Method: comparison and retrieval are clear. Machine learning is not confirmed.

Output: matching passages, links, and scores.

Effect: a student or reviewer may study the matches. The score alone is not proof that someone cheated.

The careful answer is:

This system compares and retrieves documents. We do not have enough information to say that it uses machine learning.

Remember

• Ask about the goal, inputs, method, output, and effect.
• A product's look and tone are weak proof.
• One product may contain many different parts.
• It is fine to say that there is not enough information.

Next, we will compare the main types of AI systems and tools.