Artificial intelligence is a broad name for computer systems that produce outputs we associate with intelligent work: a prediction, a recommendation, a decision, a picture, or a written answer. The term covers many different methods. It does not mean that every system thinks like a person or understands the world as a person does. NIST uses a similarly broad definition of an AI system.
If you are new to AI, three distinctions will help you use it well.
AI is a field; machine learning is one way to build it
Some software follows rules written directly by people. A tax calculator, for example, can apply a specified formula to the numbers you enter. Other systems are trained on examples so they can find patterns that would be difficult to write as a complete list of rules.
That second approach is machine learning. A team might train a model on labeled images so it can estimate whether a new image contains a damaged part. During training, the model’s adjustable values change as it works through examples. Later, during inference, the trained model uses those values to produce an output for new input. Google’s machine learning course explains training and inference in more detail.
AI is wider than machine learning, and machine learning is wider than the chatbots people often mean when they say “AI.”
A conversation is not automatically training
When you ask a chatbot a question and then correct its answer, the next answer may improve because it can use the conversation as context. That does not necessarily mean the underlying model has learned a lasting new skill or changed its trained values. Many systems are trained or updated separately from the conversations in which people use them. Some products also use memory, retrieval, or additional training, and their behavior depends on the product’s settings and design. Google distinguishes prompting from changes to model parameters.
This matters because a convincing answer can come from a model that is applying patterns, not checking the claim against a reliable source. Ask where an important statement came from, then inspect that source yourself.
Generative AI makes new output from a prompt
Generative systems can draft text, make images, produce code, or transform existing material. A language model works with pieces of text called tokens and estimates what output fits the input and context. That ability makes it useful for outlining, rewriting, summarizing, and exploring options. It can also produce fluent errors, invented citations, or details you never supplied.
Consider a simple task: drafting an invitation to a neighborhood cleanup. You can give the tool the confirmed date, place, audience, and desired length, then ask for a draft. You still need to check every event detail before sending it. If you ask the tool to choose a medical treatment, approve a loan, or make a claim about a person, the cost of a wrong answer is much higher and the review needs to be much stronger.
A useful way to start
- Choose a task with a checkable result. Draft a short email, summarize a document you can reread, or organize notes you already understand.
- Provide the relevant facts and limits. Say who the result is for, what it should include, and what the system should leave uncertain.
- Inspect the output. Check names, dates, calculations, quotations, and source links. Correct the result or do the task yourself when the stakes call for it.
You can practice the second step in our guide to writing clearer AI requests. For decisions that affect other people, use the Rescue Mission Test to think through dignity, agency, and oversight.
AI can be a useful tool for work and learning. Its value depends on the task, the information it receives, and the person who checks what it produces.
Loading comments…