Generative AI in one minute
A system that labels a customer message as a complaint is classifying. A system that drafts a reply is generating. The same application may do both. Generation creates an output; it does not mean that a model thinks, understands consequences or checks facts in the way a person would.
Generative AI through a simple example
A retailer has a spreadsheet containing one hundred products. Each row includes a name, material, dimensions and a product code. Software can classify each row into a department or predict future demand. Generative AI can instead write a short product description using the supplied fields. These are different tasks, even when one application presents them together.
| Task | Example | Expected output |
|---|---|---|
| Classification | Sort products into office, kitchen and outdoor categories. | A category for each product. |
| Numerical prediction | Estimate demand from sales history. | A number, with an appropriate assessment of uncertainty. |
| Generation | Draft a description using the supplied product details. | Newly composed content that needs checking. |
Suppose a microfibre cloth is listed as blue and 30 centimetres square. A useful draft can describe those details clearly. It should not add an antibacterial certification, a lifetime guarantee or a washing temperature absent from the source. Such additions may sound plausible because similar descriptions contain them, but plausibility is not evidence about this particular item.
Calling the other tasks traditional AI is convenient shorthand, not a strict historical boundary. Earlier AI also included systems that produced outputs, and modern models can combine classification, prediction and generation. The useful distinction is the task and how its output will be assessed, rather than a claim that everything before generative AI consisted only of rigid rules.
Draft a customer reply, summarise a document or propose a program fragment.
Generate an illustration or modify visual content using a suitable image model.
Produce speech, music or other audio with an appropriate system.
Generate or transform a sequence of frames; consistency still needs checking.
A closer technical look
Training adjusts a model's parameters, the numerical weights that influence its processing. Training data and learning objectives affect the patterns it can acquire. Inference is the use of a trained model to process input and produce output. Giving it a document in a conversation does not normally retrain those weights during that interaction. Whether a provider later uses interaction data for training is a separate product and policy question.
Foundation models. These are models trained broadly enough to be adapted to multiple downstream uses. The term describes their role as a basis for applications, not a guarantee of intelligence or quality. A foundation model may be further trained, prompted or combined with software and external information for a particular purpose.
For many text-generating systems, an LLM predicts a distribution over possible next tokens. A token is selected, added to the sequence and used when producing the next step. Repeated prediction can create a long explanation or a structured document. The probability of a continuation does not independently establish that the resulting statement is true.
Other forms of generation work differently. Diffusion approaches used for images learn to transform noisy representations towards structured outputs, typically through iterative processing. They should not be described as merely predicting the next word in a picture. Different architectures and training objectives can also be combined within one product, so a chat interface alone tells you little about every component underneath it.
Multimodal AI can work with more than one kind of information, such as an image alongside text. A system might inspect a photograph and write a description without being able to generate images itself. Input capabilities, output capabilities and reliability are separate questions. Similarly, a model is not the entire application: document retrieval, permissions, calculators and saved preferences may belong to surrounding software.
Using generated content with a clear purpose
A good starting point is a task with a source and a checkable outcome. Ask for a meeting summary containing confirmed decisions, named owners and open questions. If no owner was assigned, the summary should say so instead of filling the gap with a likely name. This makes the tool useful for organisation while keeping the source of each claim visible.
Instructions work better when they specify the audience, scope and desired format. A request for a short customer explanation based only on an attached policy is more precise than a request to write something professional. A small example of the preferred format can help. It does not remove the need to verify amounts, dates, conditions or commitments before using the result.
When current company information matters, RAG can retrieve relevant material before generation. A genuine tool call can also provide a calculation or a live database result. Those processes add information; they do not turn the generated response into an automatic statement of truth. The source might be stale, the wrong product might be selected or the model might misread a condition.
- Separate the facts supplied by a source from creative wording.
- Use a current source for changing information such as prices and availability.
- Check names, numbers and promises independently of how polished the text sounds.
- Ask for missing information to remain explicit.
- Assess the complete workflow on representative examples, including difficult cases.
A Hallucination occurs when unsupported or incorrect material is presented as factual output. For an advertising concept, invention may be part of the assignment; for a product specification, it is a defect. Define that boundary before choosing a model. The practical value of generative AI comes from an appropriate task, usable source material and a review process matched to the consequences of an error.