All the AI terms you need to know

Machine learning and neural networks: An approach that took off in the 2000s and became the foundation of today's AI. Instead of being programmed in exhaustive detail to "know" bodies of knowledge, these systems feed on hoards of data and gradually refine their ability to make sense of the information, sometimes guided by human feedback.

Artificial general intelligence (AGI): A label computer scientists apply to the still-unachieved goal of creating AI that can reason and learn in broad ways; apply those skills to new realms it hasn't encountered before; and grow in unpredictable ways.

Autonomy: The capacity of AI to act on its own to achieve a goal without specific human direction at every step — in the physical world (self-driving cars), in virtual environments (non-player characters in games) or on computer networks (personal assistants).

Self-awareness or sentience: The ability of AI to "know" that it exists and has continuity in time.

Alignment: The ways in which the goals of the people building AI do, or don't, match the goals of the systems they are creating.

Generative AI (or genAI): Machine-learning based AI that trains on sets of real-world data — most commonly images and text — to learn to predict or "generate" the next word or pixel in a sequence, creating the capacity to "write" new texts and "make" new images.

Transformers: A machine-learning programming approach introduced by Google researchers in 2017 that turbocharged the ability to create generative AI. (Not to be confused with Hollywood's robot/cars.)

Natural language: How AI researchers describe the languages humans speak. "Natural language processing" means making human language intelligible to machines.

Large language model (LLM): An AI program with a mathematical map — across a very large number of dimensions — of the relationships among a large number of words, usually broken down into tokens.

Token: Technical term for the unit generative AI models use to create their mathematical maps. People use words and sentences, but AI breaks them down into more uniform-sized tokens — chiefly for reasons of computing efficiency.

Training data: The data initially provided to an AI model for it to create its map of relationships.

Supervised and unsupervised training: If the training data has been labeled by humans in advance, giving the AI signposts and hints for how to organize it, the training is considered supervised. In unsupervised training, the model is simply turned loose on raw data, and the model gradually draws connections among tokens based on proximity.

Generative pre-trained transformer (GPT): A particular kind of LLM design, introduced by OpenAI, that uses a hybrid training approach, with an initial "pre-training" that is unsupervised and then a supervised "fine-tuning" phase.

Compute: AI industry shorthand for the costly computer time required for all this training. The larger the model, the more "cycles" needed — and the greater the value of the most advanced, speediest processors (chips).

ChatGPT: OpenAI's runaway hit service that brought LLMs and generative AI into the mainstream by packaging it as a simple chatbot that could converse with users, compose stories and poems, answer questions and more.

Chatbot: Any program that can take on one half of a conversation with a user. Chatbots have been around since the 1960s, when Eliza — a simple chatbot that mimicked a therapist — first showed people were eager to personify computers.

Turing test: A thought experiment proposed by computing pioneer Alan Turing in 1950. The Turing test measured whether a computer program could fool a human user via a blind onscreen chat into believing it was human too.

Prompt: The text users enter to describe what they want from a generative AI program like ChatGPT or its image-making equivalents, like OpenAI's Dall-E.

Multi-modal: An AI system that can take input and produce output across different categories of media, typically text, images, audio and video.

Prompt engineering: The practice, often hit or miss, of trying to optimize prompts to produce exactly the output a user desires.

Prompt injection: Like prompt engineering, but with the goal of defeating restraints AI makers have built in to limit the production of potentially harmful content — for instance, instructions for bomb-makers.

Context window: The short-term memory of a generative AI. The larger the context window, the more information you can "feed" the AI along with a prompt, giving it key ingredients upfront or new data that it might not have access to on the open internet.

Hallucination: An answer provided by generative AI that sounds plausible but is made up and incorrect. If the program does not have good information to go on, it will still try to answer a question by guessing "next words" that seem to fit.

Deepfake: Any images, photo or video produced by AI tools designed to fool people into thinking it's real.

Frontier model: An AI model that pushes the limits of what the most advanced programs can do today, and therefore poses the most potential risk.

Existential risk: Danger that a powerful AI system might threaten humanity's future because of malfunction, unintended consequences or failures of alignment.

Explainable AI: Systems that make it possible for programmers and researchers to trace how they arrived at any particular output or response. Many AI systems today are "black boxes" built without the kind of tooling that allows for such explanations.

Go deeper: How AI works, in plain English