Glossary: AI / ML / Deep Learning / NLP, Defined Plainly

This isn't a complete AI dictionary — just plain-language definitions for the terms that come up most often on this blog. If you hit a word here that isn't unfamiliar, skip it. If you hit a word on the blog that isn't here, that's a sign this page needs an update — let me know.

Contains AI-Generated Content

This article contains text written by Claude. Per our editorial policy, the ideas are human, but we believe it is important to clearly mark content that AI has even partially generated.

The basics: how AI is organized

Artificial intelligence (AI) — The broad umbrella term for machines doing things that normally require human-like intelligence: recognizing patterns, making predictions, generating language or images, making decisions. Ask ten people in the field to define it and you'll get eight different answers — it's less one technology than a whole field with a lot of overlapping subfields underneath it.

Machine learning (ML) — The largest and most talked-about subfield of AI. Instead of a programmer writing explicit rules for every situation, a machine learning system is fed data and learns patterns from it, then applies those patterns to new situations it hasn't seen before. Most of what people mean when they say "AI" today is actually machine learning.

Model — The end result of machine learning: a trained system that takes an input (a question, an image, a sensor reading) and produces an output (an answer, a label, a prediction). When people say "the model did X," they mean the trained system did X — not a person hand-coding the response.

Algorithm — The set of steps or rules a computer follows to solve a problem. Every piece of software runs on algorithms, not just AI — but in AI discussions, "algorithm" often refers specifically to the decision-making logic inside a model, especially when someone's asking whether that logic is fair or biased.

Training / training data — The process (and the data) by which a machine learning model learns its patterns. A model is "trained" on a large collection of examples — text, images, past decisions — and adjusts itself to get better at predicting or generating things like them. This matters ethically because a model trained on biased or incomplete data will tend to reproduce that bias, even without anyone intending it to.

Deep learning — A subfield of machine learning that uses layered structures called neural networks (loosely inspired by how neurons connect in a brain) to find patterns in large, messy data — images, audio, raw text — that are hard to hand-describe with simple rules. Most of the recent, headline-making AI progress, including large language models, is built on deep learning.

Natural language processing (NLP) — The subfield of AI focused specifically on human language: understanding it, generating it, translating it, summarizing it. Chatbots, translation tools, and language models are all NLP applications.

Large language model (LLM) — A deep learning model trained on huge amounts of text, built to predict and generate language. ChatGPT, Claude, and similar tools are LLMs. They're the reason "AI" has come to mean "a thing you can chat with" for a lot of people, even though language models are just one branch of the broader field.

Forecasting — A machine learning application focused on predicting future values from past data — demand forecasting, weather prediction, that kind of thing. Less flashy than language or image generation, but one of the oldest and most widely used applications of ML.

Terms that come up in the ethics conversation

Bias (algorithmic bias) — A systematic pattern where a model treats some people or groups differently than others, often because the data it learned from reflected historical inequities. "Accurate" (matches the historical data) and "fair" can pull apart here — a model can faithfully reproduce a biased pattern and still be doing real harm.

Deployment — The point at which a model moves from being built and tested to actually being used in the real world — approving loans, screening resumes, driving a car. A lot of AI ethics problems live specifically in deployment: a model can be technically sound and still cause harm because of how, where, or on whom it's used.

Transparency (explainability) — How well people can understand why a model produced a particular output. Some models are simple enough to explain step by step; others (especially deep learning models) are effectively black boxes even to the people who built them, which makes it hard to check their reasoning or catch mistakes.

Consent — Whether people know they're interacting with an AI system, and whether they agreed to have their data used to build or run it. Comes up both on the input side (was this data collected and used with permission?) and the interaction side (does the person on the other end know they're talking to a machine?).

Synthetic media — Images, audio, or video generated or altered by AI rather than captured directly from reality — deepfakes are the best-known and most concerning example, but the category also covers plenty of mundane, disclosed uses.

Sentience / consciousness — Whether a system has genuine subjective experience — whether there's "something it's like" to be it — as opposed to just producing convincing outputs. Nobody has a settled test for this, in AI or otherwise, which is most of why it's a hard question rather than a solved one.