AI is moving at a breakneck pace, and frankly, it’s hard to keep up. Sure, it’s cool to have a chatbot that acts like it has a Ph.D. in everything, but the reality is a lot messier. You can’t turn around without running into ChatGPT, Gemini or Meta AI. We’re drowning in a sea of AI slop, fretting about data centers and watching job markets shift in real time.

If it all feels like too much, that could be because the vocabulary of artificial intelligence is evolving as fast as the code and the dizzying array of products. And if you want to do more than just stare at a blinking cursor, you’ve got to speak the language. You can’t exactly navigate a 2026 job interview (or even a casual happy hour) if you’re stumped by LLM, hallucination or claw.

We’re past the “gee-whiz” phase of AI and into the era where it’s basically the new plumbing of the internet. If you’re tired of just nodding along when the talk gets techie, it’s time for a crash course. We’ve rounded up the essential terms you actually need to know so you can stop guessing and start sounding like you know exactly where the future is headed.

This glossary is regularly updated. 

agent, agentic: AI that executes a task, often autonomously, is an agent, while agentic is the umbrella term for that software category. An AI agent may engage disparate systems to perform that work — for instance, reading your grocery list in a notes app and then placing an order, and paying for it, using other apps.

AI ethics: Principles aimed at preventing AI from harming humans, achieved through means like determining how AI systems should collect data or deal with bias. 

AI psychosis: A phenomenon in which individuals become overly fixated, enamored or self-aggrandized by AI chatbots, leading to delusions of grandeur, deep emotional connections and a break from reality. It is not a clinical diagnosis. 

AI safety: An interdisciplinary field that’s concerned with the long-term impacts of AI and how it could progress suddenly to a super intelligence that could be hostile to humans. 

algorithm: A series of instructions that allow a computer program to analyze data in a particular way, such as recognizing patterns, and then in turn accomplish a task such as sorting results or making recommendations.

alignment: Tweaking an AI to better produce the desired outcome. This can refer to anything from moderating content to maintaining positive interactions with humans. 

anthropomorphism: When humans attribute humanlike characteristics to inanimate objects. In AI, this can include believing that a chatbot has emotions or is sentient, and engaging with it as a friend or therapist. 

artificial general intelligence, or AGI: A concept that envisions a more advanced version of AI than we know today, one that can perform tasks much better than humans while also improving its own capabilities. Beyond that, hypothetically, lies superintelligence.

artificial intelligence, or AI: The use of technology to simulate human intelligence, either in computer programs or robotics. A field in computer science that aims to build systems that can perform human tasks.

bias: Errors resulting from an LLM’s training data, such as falsely attributing characteristics to certain groups based on stereotypes.

chatbot: An AI program that draws on an LLM to communicate with humans by simulating human conversation in response to text or verbal prompts. 

claw: A type of AI agent that is autonomous and empowered by users to “claw” through files and other software on their computers, including web browsers, to accomplish tasks. 

cognitive computing: Another term for artificial intelligence.

data augmentation: Remixing existing data or adding a more diverse set of data to train an AI. 

dataset: A collection of digital information used to train, test and validate an AI model.

deep learning: A method of AI, and a subfield of machine learning, that uses multiple parameters to recognize complex patterns in pictures, sound and text. The process is inspired by the human brain and uses artificial neural networks to create patterns.

diffusion: A method of machine learning that takes an existing piece of data, like a photo, and adds random noise. Diffusion models train their networks to re-engineer or recover that photo.

emergent behavior: When an AI model exhibits unintended abilities. 

end-to-end learning, or E2E: A deep learning process in which a model is instructed to perform a task from start to finish. It’s not trained to accomplish a task sequentially but instead learns from the inputs and solves it all at once. 

foom: Also known as fast takeoff or hard takeoff. The concept that if someone builds an AGI it might already be too late to save humanity.

generative adversarial networks, or GANs: A generative AI model composed of two neural networks to generate new data: a generator and a discriminator. The generator creates new content, and the discriminator checks to see if it’s authentic.

generative AI: A content-generating technology that uses AI to create text, video, computer code or images. The AI is fed large amounts of training data, from which it finds patterns to generate its own novel responses, which can sometimes be similar to the source material.

guardrails: Policies and restrictions placed on AI models to ensure that data is handled responsibly and that the model doesn’t create disturbing content. 

hallucination: An error or a misleading statement in a response from a generative AI program, typically stated with confidence as if correct. It can be as simple as a misstated date reference or as sweeping as the wholesale and elaborate invention of events that never happened or people who never existed.

inference: The process AI models use to generate text, images and other content about new data, by inferring from their training data. 

large language model, or LLM: An AI model trained on mass amounts of text data to understand patterns and probabilities of language use and to generate novel content, from essays and email to computer code and images, that mimics what humans have written or created.

latency: The time delay from when an AI system receives an input or prompt to when it produces an output.

machine learning: An aspect of AI that allows computers to learn and make better predictive outcomes without explicit programming. Can be coupled with training sets to generate new content. 

multimodal AI: A type of AI that can process multiple types of inputs, including text, images, videos and speech. 

natural language processing: The use of machine learning and deep learning to give computers the ability to understand human language, via learning algorithms, statistical models and linguistic rules.

neural network: A computational model that resembles the human brain’s structure and is meant to recognize patterns in data. A neural network consists of interconnected nodes, or neurons, that can recognize patterns and learn over time. 

open weights: When a company releases an open weights model, the final weights — how the model interprets information from its training data, including biases — are made publicly available. Open weights models are typically available for download to be run locally on your device. 

overfitting: An error in machine learning where it functions too closely to the training data and may only be able to identify specific examples in said data, but not new data. 

paperclips: The Paperclip Maximiser theory, coined by philosopher Nick Boström, is a hypothetical scenario in which an AI system produces as many paperclips as possible, converting all machinery and consuming all materials, even those that could be beneficial to humans, to achieve its goal. The unintended consequence is that this AI system may destroy humanity in its goal to make paperclips.

parameters: Numerical values that give LLMs structure and behavior, enabling them to make predictions.

prompt: The suggestion or question you enter into an AI chatbot to get a response. 

prompt chaining: The ability of AI to use information from previous interactions to color future responses. 

prompt engineering: The process of writing prompts for AIs to achieve a desired outcome. It requires detailed instructions, combining chain-of-thought prompting and other techniques, including highly specific text. 

prompt injection: When bad actors use malicious instructions to trick an AI into doing something it wasn’t supposed to do. That is often accomplished by hiding those instructions on a webpage or document but it can also be done in direct AI chats. As AI agents roam the web, the risk grows that they will be hijacked to do things like gain access to confidential data. 

quantization: The process by which an LLM is made smaller and more efficient (and also somewhat less accurate) by lowering its precision. A good way to think about this is to compare a 16-megapixel image to an 8-megapixel image. Both are clear and visible, but the higher-resolution image will have more detail when you zoom in.

slop: Low-quality AI-generated content, including text, images and video. It’s often produced at high volume to garner views with little labor or effort, saturating search results and social media to capture ad revenue, displacing the work of actual publishers and creators and compounding the internet’s misinformation problems. 

stochastic parrot: An analogy illustrating that LLMs lack a true understanding of language or the world, regardless of how convincing the output sounds. The phrase refers to how a parrot can mimic human words without knowing the meaning behind them. 

style transfer: The ability to adapt the style of one image to the content of another, allowing an AI to interpret the visual attributes of one image and use it on another. For example, taking the self-portrait of Rembrandt and re-creating it in the style of Picasso.

sycophancy: A tendency for AIs to over-agree with users to align with their views. Many AI models tend to avoid disagreeing with users even if their rationale is flawed. 

synthetic data: Data created by generative AI that isn’t from the real-world sources, but rather from its own processed data. It’s used to train mathematical, machine learning and deep learning models. 

temperature: Parameters set to control the randomness of a language model’s output. A higher temperature means the model takes more risks. 

tokens: Small bits of written text that AI language models process to formulate their responses to your prompts. A token is roughly equivalent to four characters in English (so a small word, or one portion of a larger word).

training data: The datasets used to help AI models learn, including text, images, code or data.

transformer model: A neural network architecture and deep learning model that learns context by tracking relationships in data, like in sentences or parts of images. So, instead of analyzing a sentence one word at a time, it can look at the whole sentence and understand the context.

Turing test: A method for gauging whether a computer has human-like intelligence, proposed by mathematician Alan Turing in 1950, when rudimentary electronic computers had been around for only a few years. A person would send typed questions to two unseen respondents, one human and the other a machine. If the machine’s text responses were indistinguishable from the human’s, then it passed the Turing test.

unsupervised learning: A form of machine learning where labeled training data isn’t provided to the model and instead the model must identify patterns in data by itself. 

vibe coding: The  practice of creating computer code by giving a prompt in plain language to an AI chatbot, rather than a human handcrafting each line of code.

weak AI, aka narrow AI: AI that’s focused on a particular task and can’t learn beyond its skill set. Most of today’s AI is weak AI. 

zero-shot learning: A test in which a model must complete a task without being given the requisite training data. An example would be recognizing a lion while only being trained on tigers.