What Do We Mean By Artificial Intelligence?

A shop owner wants to find purchases that might be fraudulent. A filmmaker wants to turn a description into an image. A robot needs to pick up a cup without breaking it. All three could use artificial intelligence. Yet the work is different. The shop’s system estimates whether a purchase looks suspicious. The filmmaker’s system creates an image. The robot chooses how to move and controls its gripper.

When someone says a product uses AI, which of these kinds of work do they mean? We’ll start this course by explaining the terms you’ll hear most often. Later lessons will cover the models, computers, data centers, and businesses involved. This lesson starts with what an AI system does. There are two parts. First, we’ll explain artificial intelligence and machine learning, or ML. Then we’ll look at three jobs a system can perform: generating content, predicting something we don’t know, and deciding what to do.

We’ll use the same shop for those three jobs, so you can compare them. At the end, you’ll try identifying them yourself. If the names are unfamiliar, that’s fine. We’ll explain each one with an example before moving on. Artificial intelligence, or AI, is a field of work focused on building systems that can perform tasks involving perception, language, reasoning, planning, and learning. Let’s explain those words. Perception means making sense of information from an image, a sound, or another signal.

For example, a system might inspect a medical image and flag an area for a clinician to examine. Language means working with spoken or written words. A system might interpret a question or write a reply. Reasoning means using information to work toward a conclusion. Planning means working out the steps needed to reach a goal. For the robot holding a cup, those steps might be to approach it, close the gripper, and lift. The robot also needs to control the force so it doesn’t break the cup.

We’ll explain learning using the shop’s fraud problem shortly. These tasks can be built in several ways. Some systems use explicit rules written by a developer. Others use a model that has been adjusted using examples. A model is the part that processes an input and produces an output. An input is the information it receives. An output is what it returns. A system can combine those approaches with search,

optimization, tools, and conventional software. Search examines possible answers or routes. Optimization compares possible solutions against a goal while accounting for limits. For example, a delivery company might look for a route that takes less time but still visits every required stop. A tool adds another capability, such as finding a document or calculating a figure. Conventional software handles tasks such as collecting information, running rules, and passing results between components.

So the term AI covers a range of tasks and ways of building them. It doesn't tell you exactly how a particular product works. Machine learning, or ML, is one common approach within AI. In machine learning, a procedure uses data to adjust a model's behavior. The model learns statistical patterns from examples. It can then use those patterns to produce an output for a new input. Here's what that means for the shop. Suppose the owner writes a rule.

Flag every purchase above a certain amount for investigation. A customer buying several expensive items could be flagged even though the purchase is legitimate. Someone making a smaller fraudulent purchase could go unnoticed. The amount alone doesn't tell us enough. With machine learning, the team can use records of past transactions. Each record might include the amount, where and when the purchase happened, the customer's purchasing history, and whether the transaction was later confirmed as fraudulent. The learning procedure adjusts the model using these examples.

The model may find useful relationships among several details, rather than relying on the amount alone. For a new purchase, it could return a score estimating how suspicious the transaction is. The team still has to check whether it works well. If the model flags too many legitimate purchases, staff could spend hours reviewing customers who have done nothing wrong. Those are false alarms. People also choose the data, the objective, and the architecture.

The objective is the result they want the model to work toward. The architecture is its design. Evaluation means testing how well it performs. Deployment means putting it into use. People decide how a flagged purchase reaches a reviewer and what happens after the review. Learning from data doesn't remove those responsibilities. To recap part one, AI is the field. Machine learning is one approach within it.

An AI system can use learned components, written rules, or a combination of both and other methods. Now for the second part of the lesson. What job is a system performing? Knowing that it uses machine learning tells us how some of its behaviour was developed. It doesn't tell us what the user receives. Our shop sells umbrellas. The owner wants help with three tasks.

Writing a product description, estimating next month's sales, and choosing how many umbrellas to buy. Writing the description produces content. Estimating sales produces an estimate. Choosing a quantity produces a decision about what to do. We'll call these three jobs, generation, prediction, and decision making. We'll explain them in that order. AI can do other work too, such as recognizing speech,

finding documents, and performing calculations. And one product can combine several jobs. For now, these three will help us make sense of the shop's tasks. As you listen to each example, ask yourself, what does the system produce? The first job is generation. Generative AI produces content, such as text, images, audio, or video. The filmmaker gives the system a description of a scene. The system produces an image.

The description is the input, and the image is the output. For our shop owner, the task is to write a product description. She supplies details about an umbrella and asks for a few sentences for the website. The system writes a draft using those details. That draft is generated content. It may save her time, but she still needs to check it. Suppose the product information says the umbrella is water resistant, but the draft calls it waterproof.

The sentence might read well, even though the claim is wrong. A language model learns statistical patterns in language and uses them to produce text. That helps it write sentences that sound plausible. Plausible means they seem believable. It doesn't establish whether a particular claim is accurate. The same problem can occur with a date, a price, or a source. If customers will rely on the description, the owner needs to check those details before publishing it. That's generation.

The system produces content. In this example, it's a draft product description that the owner reviews. The next task is to estimate how many umbrellas the shop might sell. That's prediction. A predictive system estimates something we don't know. Its output could be a number, a category, or a probability. A probability expresses how likely something is. In our fraud example, the system could estimate how likely a transaction is to be fraudulent.

The transaction has already happened, but the answer is still unknown. Prediction can also concern the future. A sales forecast estimates how much a business may sell over a coming period. Suppose the umbrella shop uses past sales and other relevant information to estimate demand for next month. The system forecasts about 800 umbrella sales. The sales records and other information are the inputs. The estimate of 800 is the output.

The owner should ask how accurate similar forecasts have been, and what information this one uses. If circumstances change, the forecast may become less reliable. Perhaps next month turns out much drier than expected. At this point, the shop has a forecast. It hasn't chosen how many umbrellas to buy or sent an order to the supplier. The forecast can help the owner make that choice. It doesn't make the choice on its own.

The fraud score works the same way. An estimate of how suspicious a purchase looks can inform whether someone investigates it. Choosing which purchases to investigate is a separate step. We've now covered two jobs. Generation produces content, such as the product description. Prediction produces an estimate, such as expected sales. Next, how does the shop use that estimate to choose an order quantity? A decision system uses information to choose an action.

It considers what we're trying to achieve and the limits we're working within. Those are its goals and constraints. For the shop, the goal could be to have enough umbrellas for customers without spending too much money on stock that doesn't sell. The owner has limited cash and shelf space. She also has to consider delivery times and the risk of running out. These affect what she can buy. The sales forecast helps with the decision,

but the system also needs information about those other factors. Take a simple example. The shop expects to sell 800 umbrellas and already has 300 in stock. It could choose to buy 500 more. 800 is the prediction of sales. 500 is the quantity selected for purchase. That quantity might change if the shop can't afford 500 umbrellas, has nowhere to store them, or can receive another delivery partway through the month.

A decision process considers those conditions when choosing what to do. Its output is a selected action. Here, the action is to buy a particular quantity. Even a good forecast won't help much if the buying rule ignores it or uses the wrong costs. To improve the result, you need to check how the estimate is used as well as how accurate it is. There's another step to keep separate. Recommending an order and sending it to a supplier. The system could recommend 500 umbrellas and wait for the owner's approval.

Or the owner could allow it to send orders automatically under agreed conditions. Either way, choosing the quantity and placing the order are two different events. A recommendation on the screen doesn't tell us whether the supplier has received an order. Sending an order can commit the shop to spending money, so the workflow needs suitable checks and authorization. Also, a decision system doesn't have to use machine learning for every step. It might take a learned forecast and apply ordinary rules to choose a quantity. Let's compare the three jobs again.

Generation writes the product description. Its output is content. Prediction estimates umbrella sales. Its output is an estimate. Decision making chooses how many umbrellas to buy. Its output is a selected action. If the shop then sends the order, it has acted on that decision. The owner could use one assistant for all three tasks.

It gets the latest sales records, forecasts demand, writes an explanation, and proposes an order quantity. The owner reviews the proposal. Once she approves it, the order goes to the supplier. This sequence is a workflow. A series of steps that takes a task through to completion. You can now identify the jobs within it. The forecast is prediction.

The explanation is generated content. The proposed quantity is a decision recommendation. Getting the sales records, checking the proposal, and sending the order are other steps in the workflow. They each have a role, even though they don't all fit into our three job descriptions. Consider a research assistant as another example. It might find documents, rank useful passages, write a summary, calculate a number, and ask a person before taking an action.

Finding documents is retrieval. Ranking puts passages in order of usefulness. Writing the summary is generation. Calculation produces a figure. Approval determines whether the next action is allowed. Someone might describe the whole assistant as generative AI because it writes a summary. That description leaves out much of what it does.

If the result is wrong, knowing the steps helps you decide what to check. Did it choose an outdated document? Did the calculation use the right information? Did it take an action without the required approval? We'll look inside products in more detail in lesson two. For now, practice naming the jobs and the other steps around them. That gives you a better starting point for understanding a failure. So far, we've explained AI and ML,

then compared generation, prediction, and decision-making. All of these can appear in products that don't use a chat window. A shopping app can recommend items. A computer vision system can examine an image. Speech recognition can turn spoken audio into words. A forecasting system can estimate sales. Scientific models can help researchers study complex information. A robot can use a physical control system to move a gripper.

Chatbots are familiar because we can type a question and read a response. But they are only one way of using AI. When you encounter an AI product, ask three questions. What information does it receive? What does it produce? And who or what acts on that result? Try identifying the three jobs. Listen for what the system produces.

First, it estimates next month's sales. Is that generation, prediction, or decision-making? Prediction. The output is an estimate of sales. Second, it writes a description for a new product. Which job is that? Generation. The output is content.

The owner may still need to edit it and check the facts. Third, it considers the forecast, available cash and storage space, then chooses how much inventory to buy. Which job is that? Decision-making. The output is a selected action. Sending the order would be a further step. You can have all three jobs in one business and even within one product.

Let's finish with the two questions from the beginning. First, what are AI and ML? AI is the field of building systems for tasks involving capabilities such as perception, language, reasoning, planning, and learning. Machine learning is one approach within that field. It uses data to adjust a model's behavior. People still choose the data, goals, design, testing, and how the system is used. Second, what job is a system doing?

Look at the output. Generation produces content. Prediction produces an estimate. Decision-making selects an action while considering goals and constraints. A product can combine these jobs with retrieval, calculation, approval, and other steps. Before lesson two, choose a business you're familiar with. Find one task that predicts, one that generates, and one that decides.

Say what each produces. Then, explain whether one output helps with another task. You could use your workplace, a local shop, or a project you're working on. Use an example you know well enough to explain in your own words. In lesson two, we'll look at the difference between a model and the application that uses it. You'll see what else a product needs to do for someone to use the model's output.