What Counts as AI?

When an application feels smart, is it always AI?

Not always. It is more useful to understand AI as a system capability that produces predictions, recommendations, decisions, or content from its inputs. A feature may combine AI with ordinary rules and human review. Judge the mechanism and its limits, not the “AI-powered” label.

After this lesson

• Explain AI as a system capability rather than a marketing label.
• Distinguish a manual tool, fixed automation, and a pattern-learning system.
• Examine an AI claim through its inputs, outputs, adaptation, and human role.

Start with everyday life

We use AI before we have a chance to define it.

In the morning, a map suggests a different route from yesterday. Your inbox holds back a message it suspects is spam. As you write a reply, another feature proposes the next sentence. All three are often called AI, yet they do not perform the same job.

The map estimates a journey from location and traffic conditions. The spam filter classifies a message using patterns it has encountered. The writing tool proposes a plausible continuation from context. AI is an umbrella over many capabilities, not the name of one object that always works in the same way.

Three stories, three capabilities

Look at what enters, what comes out, and who checks it.

Mental model

Use a spectrum, not a yes-or-no box.

A manual tool performs an action selected by a person. Fixed automation follows written rules. An adaptive system uses patterns from data on new inputs. A multi-capability system may combine perception, prediction, search, or generation, but still operates within human-created objectives and constraints.

menCARIE Open Learning working definition

AI is a system that produces outputs for an objective with varying levels of autonomy and adaptivity.

This definition is a working tool, not a claim that every researcher agrees on one sentence. Two terms help us examine its behavior.

Autonomy

How far a system can perform steps or produce outputs without a person instructing every step.

Adaptivity

How far system behavior can change through data, context, feedback, or updates after deployment.

Do not look for AI as one object

An AI product is usually a stack of several layers.

Consider a news recommendation feature. A model may score every story. Business rules remove expired items. Profile data helps order the remaining choices. The interface controls what people finally see, while editors and readers produce new signals. The experience comes from the whole stack, not the model alone.

Separating the layers matters because failures have different causes. A poor prediction can come from unrepresentative data. Repetitive recommendations can come from a narrow objective. Harm can grow when there is no human correction path. Calling all of these an ‘AI problem’ makes diagnosis vague.

Compare the behavior

Fixed rules and pattern-based systems can look identical from the outside.

Many products combine both. A better question than ‘is it AI?’ is ‘which part learns or estimates, and who is responsible for the outcome?’

A believable misconception

If it talks like a person, it must understand like a person.

Why this can seem reasonable

Fluent language encourages us to assign human qualities to a system. Yet plausible output can be produced from patterns without human experience, understanding, or responsibility.

How to check it

Inspect the specific task: what are the inputs, how is quality measured, when does it fail, and which decisions must not be delegated?

Pause and predict

A chatbot only displays answers from a prewritten menu. Is it AI?

Choose your answer before revealing the explanation.

Four-question field tool

Test an “AI-powered” claim in two minutes.

1. What input does the system read or receive?
2. What prediction, recommendation, decision, or content does it produce?
3. Does its behavior follow fixed rules or patterns that can change?
4. Who reviews the output, corrects mistakes, and carries the impact?

If the first three questions cannot be answered, the AI claim is not informative enough. If the fourth has no answer, the risk is not yet managed.

Quick practice

Choose two digital features you used this week. Answer the four inspection questions for each one, then record one important fact you still do not know.

Summary

• AI is a family of capabilities and methods, not a magical label inside a product.
• Judge a system through its inputs, outputs, adaptation, objective, and limits.
• An AI label does not prove quality, safety, or value; the human role still needs to be explicit.

Continue from here

• Predictive and Generative AI: Continue the idea from AI Around Us with a closely related example.
• Task Inventory: Connect this lesson to Applied AI and test the idea in another context.
• Ambiguity: See how the same decision changes when viewed through Prompting.

Sources and further reading

• Explanatory memorandum on the updated OECD definition of an AI system: OECD.AI · official-guidance. Primary reference for the definition, evidence, or limits discussed in “What Counts as AI?”.
• How should we define AI?: Elements of AI · University of Helsinki · open-course. Further evidence and context for checking the explanation in “What Counts as AI?”.
• Artificial Intelligence Risk Management Framework: NIST · official-framework. Further evidence and context for checking the explanation in “What Counts as AI?”.