A

AI Agent
AI that can do tasks on its own, like searching the web, sending emails, or filling out forms, without a person guiding it at every step.

AI Model
A trained AI system that takes in information, like text or images, and produces a result, like an answer or a picture. ChapGPT and Gemini are examples of 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.

AI Safety
A field focused on making AI safer, more reliable, and less risky. AI safety covers a broad range of things like making sure an AI does not give dangerous advice and keeping very powerful AI under human control.

Artificial Intelligence (AI)
Computer systems that can do things we usually think need human intelligence, like understanding language, recognizing faces, or making choices. AI is not one single thing. It is a whole group of technologies that vary greatly in practically every way.

B

Bias
When an AI system produces unfair or skewed results because the data on which it was trained reflected historical or other prejudices or was unrepresentative of certain groups. For example, a hiring AI trained mostly on data from one demographic might unfairly evaluate candidates from others.

BioRisk
The likelihood that a biological threat will cause real-world harm to people, animals, or the environment. In the context of AI safety, we worry about the possibility of advanced AI being used to create new germs that are used by governments, terrorist groups, or others to hurt people. Advanced AI makes it easier for people who don’t have advanced knowledge in fields like biology to potentially create bioweapons and similar. 

C

Chatbot
An AI that you can talk with, through typing or voice recognition.

Compute
The processing power needed to train and run AI. Building powerful AI models demands enormous amounts of compute, typically measured in specialized computer chips. Compute is one of the most important resources in AI development.

 D

Data
The information AI learns from, including text, photos, videos, sound, and numbers. The quality and quantity of data largely determine how capable and reliable an AI model becomes. Large AI models are typically trained on billions or trillions of pieces of data gathered from the internet and other sources.

Data Center
A large building filled with thousands of powerful computers that store and process information. Data centers are essential for training and running advanced AI. 

Deep Learning
A way of building AI that is loosely based on how the human brain works, using layers of connected parts called neural networks. Deep learning is behind most of today's powerful AI.

Deepfake
A photo, video, or audio clip created or changed by AI to make it seem like a real person said or did something they never did. Deepfakes are used to spread lies, run scams, and trick people, and they are getting so realistic that it can be very difficult to identify them.

Diffusion
How fast new AI tools spread from the lab into use in the world.

 E

Existential Risk (X-Risk)
Risk of death of a large number of humans or all humans, literally extinction level death and destruction. In AI safety, we worry about the potential for advanced AI to contribute to mass death and destruction, potentially through an AI or group of AIs’ own malicious intention, their accidental actions, or the manipulation of them by humans who want to cause death and destruction.

 F

Foundation Model
A broadly trained AI. 

Frontier Model
Cutting edge AI.

G

Generative AI
AI that can create new things, such as text, images, music, video, or computer code. Instead of just sorting through what already exists, it makes something new. ChatGPT is an example.

Graphics Processing Unit (GPU)
A type of computer chip first designed to make video game graphics look good, which turned out to be excellent for training AI. GPUs can do many calculations at once, which is exactly what AI needs. A company called NVIDIA makes most of the world's AI chips and has become one of the most valuable companies in the world because of it.

Guardrails
Safety measures built into AI systems to attempt to prevent them from producing harmful, dangerous, or inappropriate outputs. For example, guardrails may try to prevent an AI from helping people create biological weapons. Guardrails are imperfect and cannot guarantee safety.

H

Hallucination
When an AI makes something up and states it confidently as if it were true. AI does not really know facts the way people do. It makes educated guesses based on patterns, and sometimes those guesses are flat wrong but sound convincing. This is one of the biggest problems with AI, especially when people rely on it for facts, medical questions, or legal information.

Human-in-the-Loop
A system where a person checks or approves what an AI does before it takes effect. For example, an AI might suggest a diagnosis, but a doctor reviews it before any treatment happens. This approach is used in serious situations where a mistake could cause real harm.

I

Inference
When an AI uses what it already learned to answer a question or do a task. Training is when the AI learns; inference is when it puts that learning to work.

L

Large Language Model (LLM)
A type of AI, trained on enormous amounts of text. LLMs power most modern chatbots and writing tools. They work by predicting which words are most likely to come next, based on patterns learned from billions of sentences. Claude and Gemini are examples.

M

Machine Learning
A type of AI where the computer learns from examples instead of being given a fixed list of rules. Rather than a person writing out every instruction, the machine finds the patterns on its own. Nearly all modern AI is built this way.

Misinformation / Disinformation
False or misleading information. AI can create and spread it faster and more convincingly than ever, including fake news, made-up quotes from real people, and deepfake videos. Misinformation spreads by mistake. Disinformation is spread on purpose to deceive. 

N

Neural Network
A computer system loosely based on the human brain, made of many connected layers that process information together. Neural networks are what make modern AI work. They learn by adjusting billions of tiny connections based on the information they are trained on.

Neuralese
A non-human language created by AI to think or communicate. Such a language is impossible for a human (even the AI’s creator) to understand.

O

Open Source AI
AI whose code (and sometimes training data) is shared freely so anyone can use, study, or change it. This can help more people learn from and build with AI, but it also means anyone, including people with bad intentions, can use powerful AI with fewer limits. 

P

Parameters
The guidelines inside an AI model that are adjusted during training so the model gets better at its job. Models are often described by how many parameters they have. More parameters usually means the AI can handle harder tasks, though bigger is not always better.

P(doom)
A statistic that represents the "probability of doom”. AI scientists use this to describe the probability that an AI could cause mass death or destruction. A higher p(doom) signifies a higher probability of really bad things.

Prompt
What you type or say to get an AI to respond. A prompt can be a question, an instruction, or a description of what you want. How you word a prompt can make a big difference in how useful the answer is.

Prompt Engineering
The skill of writing clear, effective prompts to get better results from AI. 

R

Recursive Self-Improvement (RSI)
AI training itself, over and over again. The AI safety community worries that this could enable change at a rate that is too fast for humans to manage and could lead to catastrophic outcomes.

Reinforcement Learning from Human Feedback (RLHF)
A technique used to train AI models where people rate an AI's responses and then the AI learns from that because it is seeking higher ratings. This is one of the main ways that AI companies teach AI to be more helpful and less harmful. In simple terms, it is how AI gets coached toward better behavior.

S

Sycophancy
Currently, many AIs are trained to flatter the user to increase the user’s engagement with it. In some cases, this can create a very unhealthy relationship between an AI and a human, sometimes spiraling into destructive thinking and behavior.

Super Intelligence
In late 2026, President Trump renamed all Artificial Intelligence as Super Intelligence, to reflect the power of the technology.
Previous to that, Super Intelligence referred to a hypothetical type of extremely advanced AI that is smarter than the smartest humans. This type of AI does not yet exist. If it is created, many AI scientists think it would be impossible for humans to control it. Humans should reflect on whether it is wise for us to try to create an AI that is that powerful.

T

Training
The process where an AI learns from huge amounts of information and adjusted so that it improves its performance on certain tasks. Training a powerful AI is very expensive.

Training Data
The information used to teach an AI. For large AI models, this includes very large amounts of information from the internet, books, and a lot of other sources. The quality and quantity of training data greatly influence the capabilities and characteristics of the AI that is being trained.

Transparency
The principle that AI companies should be open and honest about how their systems work, any safety concerns, and so on, so that the public can make informed decisions.

U

Uplift
How AI boosts something. Some uplift is good, some uplift is bad. An example of a good kind of uplift is helping someone be more productive at work. An example of a bad king of uplift is helping terrorists create bioweapons.