What is AI? Understand artificial intelligence through email classification and generated content
Understand artificial intelligence through spam detection and document summaries, distinguish machine learning from large language models, and learn how training differs from using a model and how to check AI outputs.
Understand the concept
AI, or artificial intelligence, refers to technologies and systems that infer outputs from inputs. An output might be a spam classification or a piece of generated text. AI is broader than a chatbot: machine learning is a common way to build AI, while a large language model is a type of model that works with language. Judge a tool by its inputs, outputs, and usefulness for the task rather than by its AI label.
AI covers more than chatbots
The OECD definition emphasizes a system inferring how to produce outputs from inputs, with those outputs able to affect physical or digital environments. Object recognition and content recommendations are examples of AI applications. A system does not have to chat to be AI or continue learning after deployment, and systems vary in their degree of autonomy.
AI also includes knowledge-based and rule-based reasoning approaches, not only machine learning. The presence of training data is therefore not a universal dividing line. For a particular product, identify the task and method it uses; that tells you more about its role than the AI badge on the interface.
Classification and generation need different checks
A spam classifier receives an email and predicts whether it is spam. A generative assistant might instead draft a reply from your instructions. These tasks need different checks: for the classifier, look for important messages that were wrongly blocked; for the draft, also verify facts and any commitments you did not authorize.
The examples below illustrate concepts; they are not product test results. One application can combine several AI methods, such as recognizing text in an image and then generating a summary. A successful step does not establish that every later step is correct. Check the result you actually intend to use.
| Example task | Output | Useful check |
|---|---|---|
| Detect spam | An email category or score | Inspect false positives and missed spam |
| Extract text from a screenshot | Editable recognized text | Compare figures and names against the image |
| Generate a meeting summary | A summary generated from the material | Check decisions against the original record |
Training a model and using it are different processes
In machine learning, training adjusts a model using data so it learns relationships useful for prediction or generation. Inference uses an existing model on new inputs. Putting a document into a chat usually supplies material for the current response; that alone does not establish that the model parameters changed or that it will remember the document later.
Product memory, external knowledge bases, and model training are separate mechanisms. An assistant might store preferences for later use or search documents when answering. The former can be a memory feature, while the latter is common in RAG. Both can change the answer users see without requiring the base model to be retrained.
A fluent answer still needs fact checking
Generative AI can produce confident but incorrect content, often called hallucinations. NIST’s generative AI risk guidance also notes that errors can come with plausible explanations or citations. For example, turning a discussion about whether to postpone a deadline into a statement that it was postponed changes the source meaning.
When trying a tool, use a short source whose answers you already know, check its dates, and ask one question the source does not answer. Recording actual errors is more useful than relying on a demo. This article explains general concepts and proposes checks; it does not claim an accuracy test of any model.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.