Reference

Glossary

The words you keep hearing, in plain English: AI basics, data centers, and privacy and safety. Where an official body or a primary source defines or measures something, we cite it; plain definitions of a term are not cited.

Original entries checked September 24, 2026; super intelligence added Oct 1, 2026; expanded glossary checked Oct 2, 2026

Accelerator: A chip built to speed up one kind of heavy computing. In AI, GPUs and TPUs are the main accelerators. Lawrence Berkeley National Laboratory (LBNL) uses “accelerated servers” for servers fitted with GPUs for AI work.¹¹
AI (artificial intelligence): Computer systems that do tasks we usually associate with human thinking, like understanding language, recognizing images or making predictions.
AI agent: AI that takes actions for you (browsing, filling out forms, running tasks over several steps) rather than only answering questions.
AI avatar: A computer-made person who appears on screen and speaks a script you type, with a matching voice. Companies use them for training and marketing videos, and some services let you make an avatar of yourself. HeyGen and Synthesia are examples.¹² Because avatars can copy a real face and voice, see also deepfake and voice cloning. See AI Systems.
AI browser: A web browser with an AI assistant built in. It can summarize the page you are on, answer questions across your open tabs and try tasks such as filling in forms. Perplexity’s Comet is one; OpenAI retired its Atlas browser in August 2026 and moved those features into ChatGPT.¹³ Hidden text on a web page can trick the assistant (see prompt injection), so keep it away from banking and other sensitive accounts.
AI safety: The work of finding the risks of AI, such as bias, data-security problems and attacks, and putting processes in place to avoid or reduce them.¹⁴
Air cooling: Cooling a data center by blowing cooled air past the equipment. LBNL notes that air-cooled chillers are widely used and are preferred where minimal on-site water use is crucial.¹¹
Air-gapped: A computer or network physically cut off from the internet and from other networks, so data cannot leak out or be reached from outside. Used for very sensitive work. See an example in the news.
Alignment: Making an AI system behave in line with human values and goals so that it is helpful, safe and reliable. IBM describes it as encoding human values and goals into AI models.¹⁵
Artificial general intelligence (AGI): A hypothetical AI with human-level ability across a wide range of tasks. It does not exist today, and researchers do not agree on exactly what would count as AGI.¹⁶ It is a different idea from super intelligence, which would go beyond human ability.
Behind-the-meter: Power that is generated and used on the customer’s side of the electric meter, so it does not pass through the public grid first. A data center built next to a power plant can be described this way. See an example in the news.
Benchmark: A standard test used to compare AI models, such as a set of exam-style questions scored the same way for every model. A high score shows how a model did on that test, not how it will do on your own task.
Bias: When an AI’s results are skewed or unfair to some people or groups, often because the data it learned from reflects existing inequities. IBM and NIST both point to training data as a source of bias.^17,1
C2PA and Content Credentials: An open standard for attaching a tamper-evident record to photos, video and audio that shows where they came from and how they were edited. C2PA, the coalition behind it, calls the result Content Credentials and compares it to a nutrition label for digital content.¹⁸ Many files do not carry one, so lacking credentials proves nothing.
Carbon dioxide equivalent (CO₂e): A way to add up different greenhouse gases as one number: the amount of carbon dioxide that would have the same warming effect.
Chatbot / AI assistant: An app you talk to in plain language, such as ChatGPT, Claude or Gemini, that runs on a large language model.
ChatGPT: OpenAI’s chatbot. OpenAI offers it on the web and in apps for Windows, Mac, iPhone and Android.¹⁹ See Which AI? for how it compares.
Chiller: A machine that cools water or another fluid, the way an air conditioner does, so the fluid can carry heat away from a data center’s equipment. LBNL describes both air-cooled and water-cooled chillers as widely used in data centers.¹¹
Claude: Anthropic’s chatbot. Anthropic says it works on the web, iOS, Android and desktop.²⁰ See Which AI?.
Cloud computing: Renting computing power, storage and software over the internet instead of owning the machines yourself. NIST defines it as on-demand network access to a shared pool of configurable computing resources.²¹
Coding agent: An AI tool that takes a programming task, such as “fix this bug,” works through it on its own, and hands back the finished change for a person to review. Google’s Jules is one example; Cursor, Claude Code and Codex offer similar modes.²² See AI Systems.
Colocation: A data center where companies rent space, power and cooling for their own servers instead of building a facility themselves. LBNL tracks colocation data centers as a separate category from hyperscale and from in-house ones.¹¹
Compute: Short for computing power: the chips, time and electricity used to train and run AI. “More compute” means more chips running for longer, which means more data center capacity.
Consumer vs. business accounts: The same AI company can handle data differently for individual and business accounts. OpenAI, for example, says it may use content from individual services such as ChatGPT to train its models unless you opt out, and by default does not use inputs or outputs from ChatGPT Business, Enterprise, Edu or its API.²³
Context window: All the text a model can look at when it writes a reply, including the reply itself. Anthropic compares it to the model’s “working memory,” which is different from the data the model was trained on.²⁴
Copilot (Microsoft): Microsoft’s AI assistant. Microsoft offers a limited free version and sells paid versions bundled with Microsoft 365 plans.²⁵ See Which AI?.
Copyright and AI training: Whether AI companies may train models on copyrighted books, art, music and news without permission is an open legal question. The U.S. Copyright Office has been studying it and has published a report on generative AI training in pre-publication form, along with separate parts on digital replicas and on whether AI-generated works can be copyrighted.²⁶ Not legal advice.
Corpus / corpora: A large collection of text or audio gathered for study or for AI training. “Corpora” is the plural of corpus. Language models are trained on very large text corpora, and speech tools on corpora of recorded audio.
CPU (central processing unit): The general-purpose chip that runs a computer’s everyday tasks. AI work leans on GPUs and other accelerators because they handle many calculations at the same time.
Cursor: An app for writing software, made by Anysphere. It looks like a regular code editor, but you can describe what you want in plain English and its AI writes, changes or explains the code, or runs an “agent” that works on a coding task by itself. It is built mainly for programmers. There is a free plan; paid plans are $8, $30, $100 and $200 a month, and team plans start at $30 per person a month. Prices as of October 19, 2026. Paid Cursor plans are now sold together with Grok and X features.²⁷ See AI Systems.
Data center: A building full of computers. AI runs in data centers, which is why AI's electricity use is a growing issue.⁷
Data retention: How long a company keeps what you typed, said or uploaded. It varies by company, plan and settings. See Chat privacy for how to check yours.
Deep learning: A kind of machine learning that uses neural networks with many layers. IBM says deep learning models power most state-of-the-art AI today.²⁸
Deepfake: Fake video, audio or images made with AI to look or sound like a real person.
Direct and indirect water use: Direct water use is water a data center uses on site, mostly for cooling. Indirect water use is water used to generate the electricity it runs on. LBNL estimated U.S. data centers directly consumed about 66 billion liters in 2023 and indirectly nearly 800 billion liters through the electricity they used.¹¹ See also the myth.
Edge computing: Running computing close to where data is created or used, such as on your device or a nearby server, instead of in a faraway data center. Cloudflare says the goal is to reduce delay (latency) and bandwidth use.²⁹
Electric grid: The network of power plants, high-voltage lines and local wires that delivers electricity to homes and businesses. Large data centers draw their power from it.
Embeddings: Lists of numbers that represent the meaning of text, so a computer can tell which pieces of text are related. OpenAI says they are commonly used for search, grouping similar items and recommendations.³⁰
EU AI Act: The European Union’s law on AI, Regulation (EU) 2024/1689. The European Commission calls it the first comprehensive legal framework on AI worldwide and says it sets risk-based rules for AI developers and users.³¹ See Rules & jobs.
Evaporative cooling: Cooling by letting water evaporate, which carries heat away, as in a cooling tower. It consumes water. LBNL notes that the evaporation in cooling towers has raised concerns about data center water use and availability locally.¹¹
Fair use: A U.S. legal rule that lets people use some copyrighted material without permission, decided case by case by courts. The Copyright Office keeps an index of court decisions to help the public follow it.³² How it applies to AI training is still being worked out.
Fine-tuning: Training an existing model a little further on specific examples so it gets better at one job or style. OpenAI describes it as training models to excel at a task.³³
Foundation model: A large, general-purpose model trained on broad data that can be adapted to many different tasks, instead of being built for just one.
Gemini: Google’s AI assistant. Google offers a free version with paid plans above it.³⁴ See Which AI?.
Gemini Notebook (formerly NotebookLM): Google’s research tool that answers questions only from the files you give it, such as PDFs, documents, web links and videos. Each answer comes with numbered citations you can click to see the exact passage it used. Example: upload a lease or a service contract and ask “When can I cancel, and what does it cost?” It points to the clauses so you can check them yourself. It can also turn your files into audio summaries, slide decks, mind maps and flashcards. It was renamed from NotebookLM in July 2026, and the old address still works. It is free with a Google account (up to 50 sources per notebook); more room comes with Google AI plans: Plus about $4.99, Pro about $19.99, Ultra $99.99 or $199.99 a month. Prices as of October 19, 2026. It is not a lawyer: check important clauses with one.^35,36 See AI Systems.
Generative AI: AI that creates new content (text, images, audio, video or code) instead of only sorting or labeling existing content. ChatGPT and image generators are generative AI.
Gigawatt (GW) and megawatt (MW): Units of electric power. One megawatt is 1,000 kilowatts, and one gigawatt is 1,000 megawatts. Data center sizes are usually quoted this way. These describe power at one moment, not energy used over time. See an example in the news.
GPU (graphics processing unit): A chip that can do thousands of calculations at the same time. It was first built for video game graphics and is now one of the main chips used to train and run AI.
Grid interconnection: Hooking a new power source, or a very large new customer, up to the electric grid. FERC explains that for new power plants it involves studies of how the connection affects grid reliability and what upgrades it needs.³⁷ Waiting for these studies is one reason projects take time.
Grok: An AI assistant made by xAI, available on the web, in phone apps and inside X.³⁸ See Which AI?.
Guardrails: Rules, filters and checks that keep an AI system within safe, intended limits. IBM describes them as policies, technical controls and monitoring that govern what models produce.³⁹
Hallucination (confabulation): When AI states something false with confidence. NIST calls it confabulation; OpenAI calls hallucinations "plausible but false statements."^1,2
Hyperscale / hyperscaler: A hyperscaler is a company that runs cloud services at very large scale. A hyperscale data center is one of its very large facilities. LBNL counts hyperscale as its own category of data center.¹¹
Immersion cooling: A form of liquid cooling in which servers sit submerged in a liquid that does not conduct electricity. LBNL lists immersion alongside direct-to-chip and rear-door cooling as liquid cooling options.¹¹
Inference: Using a trained model to produce an answer. Every time you send a prompt and get a reply, that is inference. Training is the earlier, much larger job of building the model.
Jailbreak: A prompt crafted to make an AI ignore its safety rules. OWASP describes it as a form of prompt injection in which the input causes the model to disregard its safety protocols.⁴⁰
Kilowatt-hour (kWh) and terawatt-hour (TWh): Units of energy used over time. A kilowatt-hour is 1,000 watts running for one hour, and a terawatt-hour is one billion kilowatt-hours. LBNL estimated U.S. data centers used 176 TWh in 2023, or 4.4% of U.S. electricity.¹¹ The IEA estimates data centers worldwide used about 415 TWh in 2024, about 1.5% of global electricity.⁷ The two figures cover different areas, so they should not be compared directly.
Knowledge cutoff: The date after which a model has no information from its training. Anthropic publishes a "reliable knowledge cutoff" for each Claude model.³
Large language model (LLM): The engine behind chatbots. A model trained on huge amounts of text to predict the next word, which lets it write and answer questions.¹
Latency: The delay between asking for something and getting a response, measured in time. Cloudflare notes that latency is a measure of time, not of how much data is moved.⁴¹
Liquid cooling: Cooling computer chips with liquid, which carries heat away better than air. LBNL calls it an emerging technology for cooling dense equipment such as AI servers, with forms that include rear-door heat exchangers, direct-to-chip and immersion.¹¹
Machine learning: A way of building AI in which software learns patterns from examples instead of following hand-written rules. IBM calls it the subset of AI focused on algorithms that learn patterns from training data and then make predictions about new data.⁴²
Machine translation: Software that translates text or speech from one language to another. Modern tools such as DeepL and Google Translate use AI models and handle whole documents, but they can still miss tone, idioms or legal meaning, so have important translations checked by a person.⁴³
Meta AI: Meta’s AI assistant. Meta says it is available on iOS and Android, on the web and desktop, on AI glasses and in its apps.⁴⁴ See Which AI?.
Model: The trained AI system itself. Companies give models names and version numbers (for example, a newer model may be faster or more capable).
Multimodal: AI that works with more than one kind of input or output, such as text, images, audio and video.
Narrow AI: AI built to do a specific job or set of jobs, such as translating or recognizing faces. IBM says the term describes nearly all current AI, in contrast with the hypothetical general AI.¹⁶
Neural network: Software made of many simple connected units, arranged in layers, that learns patterns from data. IBM describes it as a machine learning model that stacks “neurons” in layers and learns weights from data.⁴⁵
NIST AI Risk Management Framework: A U.S. government guide that helps organizations manage AI risks. NIST says it is intended for voluntary use and was released on January 26, 2023.⁴⁶
On-premises (on-prem): Software or servers run on a company’s own equipment in its own buildings, not rented from a cloud provider.
Open-source AI: The Open Source Initiative defines it as an AI system released so that anyone can use it for any purpose, study how it works, modify it, and share it. It says open source models and weights must also come with the data information and code used to produce them.⁴⁷
Open-weight model: A model whose trained settings (its weights) are published so that anyone can download and run it. That is not the same as open source: under the Open Source Initiative’s definition, the data information and code used to build the model must be shared as well.⁴⁷
Opt-out: Choosing to leave something you are included in by default. For AI, it often means turning off the setting that lets a company use your chats to improve its models. OpenAI, for example, says individual ChatGPT users can choose whether their conversations help improve its models.²³
Parameters: The learned numbers inside a model that determine how it turns an input into an output. Training adjusts them. IBM describes them as the learned values within a machine learning model.⁴⁸
Perplexity: An AI assistant built around answering questions with sources. Perplexity says its answers link to numbered citations.⁴⁹ See Which AI?.
Personally identifiable information (PII): Details that can identify you, such as your name, address, phone number or account numbers. A good rule is to keep PII out of chatbots unless you know how the service handles it.
Power purchase agreement (PPA): A long-term contract to buy electricity from a specific power plant at an agreed price. EPA says renewable physical PPAs usually run 10 to 20 years, and that a financial (or “virtual”) PPA involves no physical delivery of power to the buyer.⁵⁰
Power usage effectiveness (PUE): A measure of how much extra electricity a data center uses beyond its computers, such as for cooling. LBNL defines it as the facility’s total electricity use divided by the electricity used by the IT equipment.¹¹ A PUE of 1.0 would mean no overhead; lower is better.
Prompt: What you type or say to an AI. Clear, specific prompts get better answers.
Prompt engineering: Writing prompts carefully, with clear context, instructions and examples, to get better answers. See Prompting.
Prompt injection: An attack that hides instructions in text an AI reads, such as a web page or document, so it follows them instead of yours. OWASP defines it as user prompts altering a model’s behavior in unintended ways, even when the input is invisible to humans.⁴⁰ NIST lists it among attacks on generative AI.¹
Rack: A tall metal frame that holds servers stacked on top of each other in a data center. Data center capacity is often counted in racks.
RAG (retrieval-augmented generation): A method where an AI looks up information in a trusted source first, then writes its answer using what it found. AWS describes it as having the model reference an authoritative knowledge base outside its training data before it responds.⁵¹
Reasoning model: A model that works through a problem in steps before it answers. OpenAI says reasoning models use internal reasoning tokens before producing a response, which can improve results on harder tasks but uses more tokens.⁵²
Red teaming: Deliberately trying to make an AI misbehave, so the problems can be found and fixed before release. NIST describes AI red-teaming as exercises that look for potential adverse behavior or outcomes of a model and how they could occur.¹
RLHF (reinforcement learning from human feedback): A training step in which people rate or rank an AI’s answers, and the model is adjusted toward the answers people prefer. OpenAI’s 2022 paper describes fine-tuning a model first on human demonstrations, then further with reinforcement learning from human feedback.⁵³
Server: A computer that stores data or runs programs for other computers over a network. A data center is, at its core, a building full of servers.
Super intelligence (SI): A hypothetical AI that would be smarter than humans at nearly every task. It does not exist today: IBM calls it a hypothetical software-based system, and AWS calls it a theoretical concept.^8,9 Today’s chatbots are narrow AI, not super intelligence. The term is in the news because the White House now uses it: on Sept. 29, 2026 President Trump signed an executive order directing the federal government to use “Super Intelligence” instead of “artificial intelligence,” according to CBS News.¹⁰ On this site, “AI” stays the name for the technology you can use today. See also the news item and the myth.
System prompt: Hidden instructions a company gives a chatbot before you type anything, telling it how to behave, what tone to use and what to avoid.
Temporary chat / incognito: A chat mode that stays out of your history and isn't used for training. Names and rules differ by app; see Privacy.
Token: A chunk of text, often part of a word, that a language model reads and writes one at a time.¹
TPU (tensor processing unit): A chip designed by Google specifically for AI workloads. Google describes TPUs as custom-developed chips used to accelerate machine learning workloads.⁵⁴
Training: The long, expensive process of building a model by showing it huge amounts of data and adjusting its parameters until it gets good at its task. It happens before you use the model; using it is called inference.
Training cluster: Many AI chips wired together so they can work on training one model at the same time. The wiring matters because the chips must constantly share results.
Training data: The text, images and other material a model learns from before you ever use it.
Transformer: The design behind most of today’s language models. It was introduced in a 2017 research paper, “Attention Is All You Need,” which described a network built solely on attention mechanisms.⁵⁵
Voice cloning: Using AI to copy someone's voice from a recording. The FTC warns a short clip can be enough.⁶
Water consumption and water withdrawal: USGS defines withdrawal as water removed from the ground or diverted from a surface source for use. Consumption is the part of that water that evaporates or is otherwise not available for immediate reuse locally.⁵⁶ A data center can withdraw much more water than it consumes, so check which one a number refers to.
Water usage effectiveness (WUE): A measure of how much water a data center uses per unit of computing. LBNL defines it as the facility’s total water consumption divided by its IT equipment’s electricity use, reported in liters per kilowatt-hour.¹¹
Watermarking: Hiding a marker inside AI-made images, audio, video or text so it can be detected later. Google says its SynthID tool embeds invisible watermarks and is designed to survive changes such as cropping or compression.⁵⁷ A missing watermark does not prove something is real.
Web search / grounding: When a chatbot looks things up online before answering and shows links to its sources.^4,5
Workflow automation: Setting up a chain of steps that runs by itself when something happens, for example: a new email arrives, AI summarizes it, and the summary is posted to your team chat. Tools such as n8n and Zapier do this, and many now add AI steps.⁵⁸ See AI Systems.
Zero data retention: An arrangement some AI providers offer approved business customers in which the provider does not keep customer prompts and answers beyond processing them (details vary by provider). OpenAI says its Zero Data Retention excludes customer content from its abuse-monitoring logs and is subject to prior approval.⁵⁹ It is not something a free consumer account normally has.

Sources59 sources for this pageSee this page’s sources · All sources & credits