AI is evolving 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 much more complicated. You can’t turn around without running into ChatGPT, Gemini or Meta AI. We’re drowning in a sea of AI slopeworry data centers and observe the evolution of labor markets in real time.
If all this seems like too much, perhaps it’s because the vocabulary of artificial intelligence is evolving as fast as the code and the dizzying range of products. And if you want to do more than just stare at a blinking cursor, you need to speak the language. You can’t exactly navigate a job interview in 2026 (or even a casual happy hour) if you’re stumped by LLM, hallucinations, or claws.
We have moved beyond the “gee-whiz” phase of AI and have entered the era where it is essentially the new plumbing of the Internet. If you’re tired of just nodding your head when the discussion gets technical, 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 looking like you know exactly where the future is going.
This glossary is regularly updated.
agent, agent: AI that performs a task, often autonomously, is a agentwhile agentic is the generic term for this category of software. An AI agent can use disparate systems to do this work: for example, reading your shopping list in one notes app, then placing an order and paying for it using other apps.
AI Ethics: Principles for preventing AI from harming humans, achieved through means such as determining how AI systems should collect data or handle bias.
AI Psychosis: A phenomenon in which individuals become overly obsessed, enamored, or self-aggrandizing with AI chatbots, leading to delusions of grandeur, deep emotional connections, and a break from reality. It is not a clinical diagnosis.
AI Security: An interdisciplinary field that is interested in the long-term impacts of AI and how it could suddenly evolve into super intelligence that could be hostile to humans.
algorithm: A series of instructions that allows a computer program to analyze data in a particular way, such as recognizing patterns, and then perform a task such as sorting results or making recommendations.
alignment: Adjusting an AI to better produce the desired result. This can refer to everything from moderating content to maintaining positive interactions with humans.
anthropomorphism: When humans attribute human characteristics to inanimate objects. In AI, this might 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, capable of performing tasks much better than humans while improving its own capabilities. Beyond this, hypothetically, lies superintelligence.
artificial intelligence, or AI: Using technology to simulate human intelligence, whether in computer programs or robotics. A field of computer science that aims to build systems capable of performing human tasks.
bias: Errors resulting from an LLM’s training data, such as incorrectly assigning characteristics to certain groups based on stereotypes.
chatbot: An AI program that leverages an LLM to communicate with humans by simulating human conversation in response to text or verbal prompts.
claw: A type of autonomous AI agent that is empowered by users to “rummage” through files and other software on their computers, including web browsers, to accomplish tasks.
cognitive computing: Another term for artificial intelligence.
data augmentation: Remix existing data or add a more diverse dataset to train AI.
dataset: A collection of digital information used to train, test and validate an AI model.
deep learning: An AI method and subfield of machine learning that uses multiple parameters to recognize complex patterns in images, audio, and text. The process takes inspiration from the human brain and uses artificial neural networks to create models.
broadcast: A machine learning method that takes existing data, such as a photo, and adds random noise to it. Diffusion models train their networks to reorganize or recover this photo.
emergent behavior: When an AI model exhibits unintended capabilities.
end-to-end learning, or E2E: A deep learning process in which a model is asked to perform a task from start to finish. It is not trained to complete a task sequentially, but rather learns from inputs and solves it in one go.
mousse : Also known as quick takeoff or hard takeoff. The concept that if someone builds an AGI, it may already be too late to save humanity.
generative adversarial networks, or GAN: 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 if it is authentic.
Generative AI: A content generation technology that uses AI to create text, videos, computer code, or images. The AI is given 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.
guardrail: Policies and restrictions placed on AI models to ensure that data is processed responsibly and that the model does not create disturbing content.
hallucination: An error or misleading statement in a response from a generative AI program, usually confidently stated as if it were correct. It can be as simple as a reference to an incorrect date or as radical as the comprehensive and elaborate invention of events that never happened or people that never existed.
inference: The process that AI models use to generate text, images, and other content on new data, by deduct from their training data.
large language model, or LLM: An AI model trained on massive amounts of text data to understand patterns and probabilities of language use and to generate novel content, from essays and emails to computer code and images, that mimics what humans have written or created.
latency: The time between when an AI system receives an input or prompt and when it produces an output.
machine learning: An aspect of AI that allows computers to learn and achieve better predictive results 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 input, including text, images, video, and speech.
natural language processing: The use of machine learning and deep learning to give computers the ability to understand human language, through learning algorithms, statistical models and linguistic rules.
neural network: A computer model that resembles the structure of the human brain and is intended to recognize patterns in data. A neural network is made up of interconnected nodes, or neurons, that can recognize patterns and learn over time.
open weights: When a company releases an open weighting model, the final weights (how the model interprets information from its training data, including biases) are made public. Open weight models are usually available for download to run locally on your device.
overfitting: An error in machine learning where it works too closely with the training data and may only be able to identify specific examples in said data, but not new data.
paperclips: The Paperclip Maximizer theory, coined by philosopher Nick Boström, is a hypothetical scenario in which an AI system produces as many paperclips as possible, converting all machines and consuming all materials, even those that could be beneficial to humans, to achieve its goal. The unintended consequence is that this AI system could destroy humanity in its goal of making paperclips.
parameters: Numerical values that give the structure and behavior of LLMs, allowing them to make predictions.
fast: The suggestion or question you type into an AI chatbot to get a response.
fast chaining: The ability of AI to use information from previous interactions to color future responses.
rapid engineering: The writing process motivates AIs to achieve the desired result. This requires detailed instructions, combining thought chain prompts and other techniques, including very specific text.
rapid injection: When bad actors use malicious instructions to trick an AI into doing something it wasn’t supposed to do. This is often done by hiding these instructions on a web page or document, but it can also be done in direct discussions with the AI. As AI agents roam the web, the risk increases that they will be hijacked to, for example, access confidential data.
quantification: The process by which an LLM becomes smaller and more efficient (and also somewhat less precise) by decreasing 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.
slope : Low-quality AI-generated content, including text, images, and videos. It is often produced in high volume to garner views with little work or effort, saturating search results and social media to capture ad revenue, supplanting the work of actual publishers and creators, and worsening misinformation problems on the Internet.
stochastic parrot: An analogy illustrating that LLMs lack a true understanding of language or the world, no matter how compelling the outcome. The expression refers to the way a parrot can imitate human words without knowing their meaning.
style transfer: The ability to match the style of one image to the content of another, allowing an AI to interpret the visual attributes of one image and use them on another. For example, taking Rembrandt’s self-portrait and recreating it in Picasso’s style.
sycophancy: A tendency for AIs to agree too much 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 does not come from real-world sources, but rather from its own processed data. It is used to train mathematical, machine learning and deep learning models.
temperature: Parameters defined to control the randomness of a language model’s output. A higher temperature means the model takes more risks.
tokens: Small pieces 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 part of a larger word).
training data: 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, such as in sentences or parts of images. So instead of analyzing a sentence word by word, it can look at the entire sentence and understand the context.
Turing test: A method of determining whether a computer has human-like intelligence, proposed by mathematician Alan Turing in 1950, when rudimentary electronic computers had only been around for a few years. A person would send typed questions to two invisible respondents, one human and the other machine. If the machine’s textual responses were indistinguishable from those of the human, then I failed the Turing test.
unsupervised learning: A form of machine learning in which labeled training data is not provided to the model and, instead, the model must identify patterns in the data itself.
ambiance coding: The practice of creating computer code by giving a plain language prompt to an AI chatbot, rather than a human manually creating each line of code.
Weak AI, aka Narrow AI: An AI that focuses on a particular task and cannot learn beyond its skill set. Most current AIs are weak AIs.
Learning without shooting: Testing in which a model must complete a task without receiving the required training data. An example would be recognizing a lion while only being trained on tigers.