A

Agentic AI
An AI that can do a series of complex tasks over a longer period of time. Agentic AI is generally more powerful than an AI agent.

AI Agent
An AI that can complete a certain type of task without needing step-by-step instructions from a person.

AI Bias
When an AI system produces unfair or skewed results because the data it was trained on reflected historical prejudices or was unrepresentative of certain groups. Addressing AI bias is a major challenge for developers and policymakers.

AI Model
An AI system that has been created and trained. For example, Claude and ChatGPT are two different AI models.

AI Psychosis
When a human becomes so engaged with a chatbot that they begin exhibiting very unhealthy behaviors that can include delusions, self harm, and harm to others.

Algorithm
The instructions that a computer follows. Software is built out of a series of algorithms.

Artificial General Intelligence (AGI)
A very powerful type of AI that may or may not quite exist yet, where the AI is advanced enough to perform lots of different types of complicated tasks similar to how a human can.

Artificial Intelligence (AI)
A computer program that does one or many different things. AI is an incredibly broad term and that is part of the challenge in communicating about it. It means very different things in different contexts.

Artificial Super Intelligence (ASI)
A super powerful type of AI that could be smarter than the sum of all people. This does not exist yet. It may or may not be possible to create. Humans need to reflect on whether it is something that we should even try to create.

C

Chatbot
A computer program designed to simulate conversation with a human user, typically through text.

Compute
The processing power used to train and run AI systems. Building powerful AI models demands enormous amounts of compute.

 D

Deep Learning
A type of machine learning inspired by the structure of the human brain, using layered networks called 'neural networks' to process information.

Deepfake
A piece of audio, video, or image content that has been manipulated or entirely generated by AI to make it appear that a real person said or did something they never actually did. Deepfakes are a growing concern for misinformation, fraud, and political manipulation, and are increasingly difficult to detect with the naked eye.

 E

Existential Risk (X-Risk)
The risk that a powerful AI can cause death to a large number of humans, or all humans, through the AI’s own action or the manipulation of it by humans who want to cause death and destruction.

 F

Foundation Model
A type of AI that was trained on tons and tons of data. Some examples are GPT-4, Claude, and Gemini. This is very different from AI systems before that which were built for specific tasks.

G

Generative AI
AI systems that can create new content, text, images, audio, video, code, or other media, rather than simply analyzing or categorizing existing content.Generative AI has rapidly expanded what computers can create and has raised new questions about copyright, authenticity, and creative work.

Guardrails
Safety measures built into AI systems to prevent them from producing harmful, dangerous, or inappropriate outputs. Guardrails may prevent an AI from giving instructions for creating weapons, generating illegal content, or making false claims about real people.

H

Hallucination
When an AI system confidently states something that is factually incorrect or completely made up. AI models do not 'know' things the way humans do; they generate responses based on statistical patterns, which can lead them to produce convincing-sounding but wrong information. Hallucination is one of the most important limitations of current AI systems.

Human-in-the-Loop
A design approach where a human reviews, approves, or corrects AI outputs before they take effect. For example, an AI might draft a medical diagnosis, but a doctor reviews it before any action is taken. Human-in-the-loop systems are often favored in high-stakes settings where errors could be costly or dangerous.

L

Large Language Model (LLM)
A type of AI trained on massive amounts of text to understand and generate human language. They work by predicting what words are likely to follow other words.

M

Machine Learning
A branch of AI in which systems learn to improve their performance by analyzing data, rather than being explicitly programmed. Most modern AI is built on machine learning.

N

Neural Network
The system through which AI processes information.

Neuralese
A non-human language created byAI to think or communicate.

O

Open Source AI
AI models and systems whose underlying code and sometimes training data are made freely available for anyone to use, inspect, or modify. This can facilitate malicious actors harnessing advanced AI to do bad things without safeguards.

P

Prompt
The input a person gives to an AI system, typically a question, instruction, or piece of text, that the AI uses to generate a response.

R

Recursive Self-Improvement
A process through which an AI or group of AIs can train themselves. This can create a cycle of rapid change that can be good or bad or a mix.

T

Training
The process by which an AI model learns from data by processing large amounts of information through certain parameters set by developers.

Transparency
The idea that AI companies should be open about how their AI works and the risks and capabilities it has.