Artificial intelligence works by processing information through algorithms that produce predictions, decisions, or new content.
Some AI systems follow explicit rules; many modern systems learn patterns from data.
- During training, a learning algorithm adjusts a model’s internal settings.
- During use, that model applies what it learned to new inputs.
Its usefulness depends on the training examples, the objective it was trained to meet, and the situation in which it is used.
AI, machine learning, and deep learning are different layers
Artificial intelligence is the broad field of building systems that perform tasks associated with intelligence, such as recognizing speech, planning, or interpreting language. It includes approaches based on rules and search as well as approaches that learn.
Machine learning is one approach to AI: developers use data to fit a model rather than manually specifying every decision. Deep learning is a branch of machine learning that uses neural networks with multiple processing layers.
Generative AI describes what a system does—creates content—not a separate rung in that hierarchy. A text generator can be both generative AI and a deep learning system. A spam classifier can use machine learning without generating anything.
How a model learns from examples
Consider an illustrative email filter. Its job is to estimate whether a message is spam.
1. Turn the input into usable numbers
A model cannot calculate directly with an email as a person sees it. Software converts relevant information into numerical inputs, called features.
For a simple filter, these might describe whether the sender is familiar, how many links appear, or whether certain words occur. More sophisticated models can learn useful representations of the text itself.
The developers also need a training signal. In this example, each training email has a label identifying it as spam or legitimate mail.
2. Produce a prediction and measure the error
The model combines the inputs using adjustable values called parameters. A weight controls how strongly an input influences a calculation; a bias provides an adjustable offset.
At first, the model’s predictions may be poor. A loss function assigns a numerical penalty based on the prediction and the training target. A confidently wrong prediction can receive a larger penalty than a less confident mistake.
3. Adjust the parameters and repeat
An optimization method such as gradient descent calculates which direction to move the parameters to reduce loss, then makes a small update.
For neural networks, backpropagation efficiently calculates how changes to parameters throughout the network would affect the loss. The optimizer uses those calculations to update the parameters.
These are related but distinct jobs: backpropagation computes gradients; the optimizer applies updates.
Repeated across batches of examples, the process can produce a model that recognizes useful patterns. It does not guarantee that every prediction improves or that the model discovers the best possible solution.
4. Check whether the learning transfers
A filter that recognizes its training emails is not necessarily useful on tomorrow’s inbox.
Developers therefore separate data into training, validation, and test sets. Training data adjusts parameters, validation data helps guide development choices, and a held-out test set provides a final evaluation.
Overfitting occurs when a model fits peculiarities of its training data too closely and performs poorly on new examples. A filter might learn to associate a newsletter’s formatting with spam instead of identifying a more dependable pattern.
The evaluation also needs to resemble actual use. Strong results on one collection of emails do not establish equal performance for every language, organization, or future spam campaign.
What happens when the trained AI receives a new input?
Using a trained model is called inference. The email becomes numerical inputs, the model calculates a score, and the surrounding application decides what to do with it.
Here is an illustrative example; these are invented scores, not measurements from a real filter:
| Message | Model’s spam score | Action at a 0.80 threshold |
|---|---|---|
| A | 0.93 | Send to spam |
| B | 0.72 | Keep in inbox |
| C | 0.18 | Keep in inbox |
The classification threshold separates the model’s estimate from the application’s action. If the threshold drops to 0.70, message B moves to spam even though the model and its score have not changed.
Lower thresholds generally catch more spam but also risk hiding more legitimate messages. Higher thresholds generally reverse that trade-off. A model score is an estimate, not proof of the message’s true category.
This distinction explains why two applications can behave differently while using the same underlying model: their decision rules may differ.
What happens inside a neural network?
A neural network transforms numbers through connected layers. In a basic layer, each unit combines incoming values with learned weights and adds a bias. The resulting values become inputs to another layer.
Hidden layers sit between the original inputs and the output. “Hidden” simply identifies their position inside the network.
Crucially, networks also use nonlinear transformations called activation functions. Without nonlinear operations, stacking basic linear layers would still amount to one linear transformation. Nonlinearity lets the network represent more complex relationships.
For the email filter, that means combinations of signals can matter. A phrase might be suspicious in one context and harmless in another.
“Neural” is an analogy to biology. These units perform mathematical operations; describing them as neurons does not establish that the system thinks or experiences the world like a person.
Where does the learning signal come from?
Not every model learns from human-labeled examples. Common learning approaches include:
| Approach | Learning signal | Example |
|---|---|---|
| Supervised learning | Provided target answers | Emails labeled spam or legitimate |
| Unsupervised learning | Structure within unlabeled data | Grouping similar documents |
| Self-supervised learning | Targets derived from the data itself | Predicting missing or subsequent text |
| Reinforcement learning | Rewards associated with behavior | Learning a strategy through repeated interaction |
These methods can appear at different stages of one system’s development. The objective matters because training rewards success on that objective, which may only approximate what people actually want.
How does a language model generate an answer?
A text-generating language model divides text into tokens—units that can be words, word fragments, or other character sequences.
Many use a transformer architecture. Its attention mechanism lets representations of tokens incorporate information from other relevant tokens in the available context.
In a typical autoregressive generator, the model calculates probabilities for possible next tokens. A decoding method selects one, adds it to the sequence, and repeats. Each new token becomes part of the context for subsequent generation.
That produces connected sentences, but fluent text does not establish factual accuracy. A plausible continuation can contain a false date, invented reference, or unsupported explanation.
Additional training, including fine-tuning, can adapt a model to particular tasks. Giving it instructions or examples in a prompt is different: that changes its immediate input without necessarily changing its parameters.
Is generating an answer the same as searching the web?
No. A language model can generate text without retrieving a live page.
A system using retrieval-augmented generation adds a retrieval component: it finds relevant material and conditions generation on that material. The original research combines a trained language model with a searchable external index.
This separates two sources of information: patterns encoded during training and documents retrieved for the current task. Retrieval can provide useful evidence, but selecting an irrelevant document or misinterpreting a passage can still produce a bad answer.
Does AI keep learning whenever you use it?
Ordinary inference does not necessarily update the model. A chatbot can respond to a correction because the correction is now in its context, even while its parameters remain unchanged.
Likewise, storing a preference and supplying it in a later prompt is different from retraining. Whether interactions are collected for future training depends on the system and its policies.
A deployed AI service also needs more than a model. Production systems include data processing, serving infrastructure, testing, and monitoring. Developers may update models when performance changes, but that requires an actual update process.
To understand an AI result, ask three questions:
- What information did it receive?
- What was it trained to optimize?
- How did the application turn its output into an answer or action?
Those questions reveal both the source of its usefulness and the conditions under which it can fail.