Data and AI
Data & AI
AI moves faster than the language around it, so half the battle is just knowing what people mean when they say "grounding" or "RLHF" in a meeting. This glossary keeps the definitions short and plain. Use the letters below to jump around, or read it start to finish if you're feeling ambitious.
The share of predictions a model gets right, plain and simple. It works fine when your categories are balanced. The moment one category dominates the data, accuracy starts lying to you, which is why it rarely stands alone.
Poking a model on purpose to find where it breaks. You feed it tricky, malformed, or hostile inputs and watch what falls out. Better you find the failure than a user does.
AI that doesn't just answer, it acts. It can plan a task, take steps, use tools, and adjust based on what happens, without a human nudging it along at every turn. Think less chatbot, more intern who actually finishes the assignment.
A system that pursues a goal on your behalf. It reads some environment, decides what to do, and does it, often chaining several actions together. A good one knows when the job is done.
A structured review of an AI system: what data it learned from, how it decides, where it might discriminate or leak, and if there is model drift. Auditors check the system against a standard or regulation and write up what they find. Evidence is key, not vibes.
A formal agreement that lets an AI company use someone else's content for training or retrieval, usually for a fee. The owner keeps the rights and sets the terms. This is the legal plumbing behind a lot of today's AI models.
A longer-term arrangement where a content owner supplies data to an AI developer, with established, shared standards for quality, formatting, and usage. Think less of one-off sale, more ongoing supply relationship. Both sides usually want structure and predictability.
The set of principles guiding how AI gets built and used, covering fairness, harm, consent, and accountability. It asks not just whether a system works, but who it works for and who it hurts. Ethics without follow-through is decoration, so the useful part lives in enforcement.
The rules, roles, and processes an organization uses to keep its AI in line. Governance decides who signs off, what gets logged, and what happens when something goes wrong. Think of it as the org chart and rulebook wrapped around the technology.
Laws and rules governments set for how AI can be built and deployed. The EU AI Act is the most cited example, sorting systems by risk and attaching obligations to each tier. Regulation always lags the technology, but it's catching up.
The practice of finding, ranking, and reducing the ways an AI system could cause harm or fail. That covers bias, security holes, bad outputs, legal exposure, the whole lot. You can't remove risk, so the work is deciding which ones you'll tolerate and which you won't.
The field focused on keeping AI systems from causing harm, whether by accident or misuse. It spans everything from stopping a chatbot giving dangerous advice to worrying about and planning safe guardrails around far more capable future systems. Safety is a design goal, not something you bolt on later.
The examples a model learns from. Text, images, audio, code, whatever the model is meant to handle. The old programmer saying holds here with a stubborn vengeance: garbage in, garbage out.
How openly an AI system's workings, data, and limits get disclosed. It ranges from publishing model cards to explaining a single decision to the person it affected. Users can't trust what they can't see.
Content that's been cleaned, structured, and tagged so a machine can actually use it. Consistent formatting, clear metadata, no broken markup. Most raw content isn't ready, which is why preparing it is a whole business.
The work of making an AI system's behavior match what people actually want, not just what they literally asked for. A misaligned model can follow instructions to the letter and still do the wrong thing. This gets harder as models get more capable.
Software that does things we'd normally call intelligent: understanding language, spotting patterns, making decisions, generating content. It's a broad umbrella stretched over everything from spam filters to modern LLMs that write whole dissertations. The definition keeps moving as the tech does.
Fun fact: The phrase is older than the technology by a long way! John McCarthy coined "artificial intelligence" in a 1955 proposal for a summer workshop at Dartmouth, which the Rockefeller Foundation funded with a $7,500 grant.
An AI agent that runs with minimal supervision, making its own calls about how to reach a goal. The more autonomy you grant, the more you're trusting its judgment. That trade sits at the center of most agent design.
A standard set of data everyone uses to test and compare models on the same footing. Without a shared yardstick, every vendor's "best in class" means nothing. The catch: once a benchmark goes public, models start gaming it.
Running a model against those standard datasets to measure how it performs. You get numbers you can compare across models and over time. Just remember a benchmark score is a proxy at best, not the real world.
Systematic skew in a model's outputs that favors or harms certain groups. It usually rides in on the training data and gets baked into the predictions. The uncomfortable part is that a biased model can be highly accurate and still unfair. Tricky.
Book Interchange Tag Suite, an XML vocabulary for marking up the structure of scholarly and technical books. It extends JATS, the journal standard, so publishers can handle books and journals with one toolkit. Worth noting it's an NLM project, not a formal NISO standard the way JATS is.
See also: JATS XML
Drawing a box around an object in an image and labeling it, so a computer-vision model learns where things are. Every self-driving demo you've seen was fed millions of these boxes. Tedious work, and the model is only as good as the boxes.
Prompting a model to reason step by step instead of blurting out an answer. Showing the work tends to improve accuracy on anything that needs logic or math. It also lets you see where the reasoning went off the rails, though the newer reasoning models keep much of that step-by-step work internal, where you never see it.
The field teaching machines to interpret images and video. Detecting objects, reading text, recognizing faces, sorting defects on a line. If a task involves a camera and a decision, computer vision is usually behind it.
The systems and processes for collecting, recording, and honoring people's permission to use their data. Regulations like GDPR made this non-optional. Done right it's an audit trail; done wrong and it's a lawsuit.
Making content usable by people with disabilities, including those on screen readers or other assistive tech. Standards like WCAG spell out what that requires. For AI, accessible source content also happens to be cleaner, better-structured content.
Tracing and citing a piece of content back to its source or creator. It matters for credit, for licensing, and increasingly for figuring out what an AI model was trained on. Attribution is easy to lose and hard to reconstruct after the fact.
Converting physical or analog material into digital form. Scanning books, transcribing tapes, photographing archives. It's step one for anyone who wants to search, license, or train on old material.
Adding structure and metadata to raw content so it becomes more useful: tags, links, entities, classifications. It's the difference between a pile of documents and a labeled, searchable library. Enriched content is what makes downstream AI work. Builds upon the cleanup remediation provides.
A platform where content owners and buyers meet to trade usage rights, often for AI training. It standardizes the messy parts: discovery, terms, payment. The AI boom turned this from a niche into a growth industry.
Forcing messy content into one consistent format. Same date style, same encoding, same structure across the whole pile. Boring and invisible, but completely necessary before anything downstream can trust the data.
Organizing content into a defined shape, with headings, fields, and hierarchy a machine can parse. Unstructured text is hard to query; structured content answers questions. Most enrichment pipelines start here.
Converting content from one format or structure to another, say a PDF into tagged XML. The goal is usually to make the content reusable, searchable, or machine-readable. It's the workhorse behind digital publishing.
The specific permissions attached to a piece of content: who can use it, how, for how long, in what context. Rights are narrow by default. Assuming you can do something the license never granted is how people end up in court.
The amount of text a model can hold in mind at once, measured in tokens. Anything outside the window may as well not exist to the model. Bigger windows help, but they don't make a model remember across sessions.
Staying within the law when you use someone else's creative work. For AI, that covers both what goes into training and what comes out. The rules are still being fought over in court, which makes compliance a moving target. Tricky.
The legal rights an author holds over their original work, controlling copying and reuse. It's automatic the moment the work is fixed in a tangible form. AI has thrown a decade's worth of unsettled questions at a system that predates it.
Material covered by copyright, which is most of what you'll find online whether it says so or not. Absence of a notice doesn't mean it's free to take. Assume protected until proven otherwise. Careful.
Labeling raw data so a model can learn from it. Manually tagging images, marking sentiment, flagging entities. It's the unglamorous labor most machine learning quietly depends on.
Fixing or removing bad data: duplicates, errors, gaps, nonsense. It's the least loved and most important step in any data project. Skip it and every model downstream inherits your mess.
Adding external or derived information to a dataset to make it more valuable. Appending a company's industry, a location's coordinates, a customer's segment. Richer data, sharper models.
Attaching the correct answer to each example so a supervised model has something to learn from. Same idea as annotation, and the two terms get used interchangeably. Quality here sets the ceiling for everything after.
See also: Data Annotation
Protecting personal information from misuse and giving people control over their own data. Laws like GDPR and India's DPDP Act set the floor. For AI, privacy questions start with the training data and never really stop.
The documented history of where data came from and what's been done to it. Provenance answers the awkward question "can we actually use this?" without a shrug. It's becoming a legal necessity, not a nicety.
How fit a dataset is for its purpose: accurate, complete, consistent, current. Bad data quietly poisons everything built on it. You rarely notice until the model's already wrong.
The principle that data is subject to the laws of the country where it's collected or stored. It shapes where companies can put their servers and whose rules apply. Cross-border AI runs straight into this.
Checking that data meets the rules before you use it: right format, sensible values, nothing missing. It catches problems at the door instead of deep in a pipeline. Cheap insurance.
Deliberately selecting, organizing, and maintaining a dataset for a purpose. Curation is editorial judgment applied to data. What you leave out matters as much as what you keep.
A branch of machine learning built on neural networks with many layers. The depth lets it learn complicated patterns from raw data on its own. Nearly every recent AI leap runs on it.
See also: Neural Network
The end-to-end path content takes from raw draft to published output across formats. It covers authoring, editing, structuring, converting, and distribution. In modern shops a lot of those steps are automated.
A way of turning words, images, or other data into vectors, lists of numbers that capture meaning. Things that mean similar things land near each other in that number space. It's the trick that lets machines do semantic search instead of dumb keyword matching.
See also: Vector Database
Techniques that make a model's decisions understandable to humans. Instead of "the model said no," you get why it said no. This matters most where decisions carry weight: loans, hiring, healthcare.
A single number that balances precision and recall, useful when you care about both. It's the harmonic mean of the two, which punishes a model for being lopsided. Reach for it when your data is imbalanced and plain accuracy would flatter you.
See also: Precision, Recall
Taking a pre-trained model and training it a bit more on your own narrower data. You keep the general knowledge and add specialization. Far cheaper than building a model from scratch, which is the whole point.
See also: Foundation Model
A large model trained on broad data that serves as a base for many downstream uses. Adapt it once and reuse it everywhere: chat, search, code, summarization. GPT and its cousins are the obvious examples.
See also: Fine-Tuning
AI that creates new content rather than just sorting or scoring existing content. Text, images, audio, video, code, all generated from a prompt. It's the branch that put AI on every front page.
A reference dataset labeled so carefully it's treated as ground truth for evaluation. You measure other work against it. Building one is slow and expensive, which is exactly why it's trusted.
The verified correct answer you compare a model's output against. It's the anchor for both training and evaluation. If your ground truth is wrong, your model learns to be confidently wrong.
Tying a model's output to a reliable source instead of letting it free-associate. A grounded answer can point to where it came from. It's one of the main defenses against confident nonsense.
When a model produces something fluent, confident, and false. It isn't lying, since it has no notion of truth; it's filling gaps with plausible-sounding filler. This is the single biggest reason you check an AI's work before trusting it.
Labeling done by people rather than machines. Humans catch nuance, context, and edge cases automated labeling misses. Slower and pricier, and for hard tasks it's still the gold standard.
Signals from people about whether an AI's output was good or bad, used to improve it. Those judgments become training data of their own. It's how models learn preferences no rulebook could spell out.
A setup where a person reviews or approves what the AI does before it counts. It keeps a human accountable for high-stakes calls. The design question is always where to place the human, not whether to.
Labeling the contents of an image so a vision model can learn from it. Boxes, outlines, tags, and points, depending on the task. It's the raw fuel for anything that sees.
The stage where a trained model actually does its job on new input. Training is the studying; inference is sitting the exam. Most of what a deployed model costs to run happens here.
Journal Article Tag Suite, the ANSI/NISO standard for marking up scholarly journal articles in XML. It's what publishers and archives use to exchange articles in a format everyone can read. Practically universal in academic publishing, and the base that BITS extends to books.
See also: BITS XML
Pulling structured facts out of unstructured content: names, dates, relationships, events. It turns prose a human has to read into data a machine can query. This is how you build a knowledge graph from a pile of documents.
A network of entities and the relationships between them, stored so machines can reason over it. People, places, things, and how they connect. It's what lets a search engine answer "who directed the movie she starred in" without breaking a sweat.
A model trained on enormous amounts of text that can understand and generate language. It predicts the next token over and over, and out of that comes fluent writing, code, and conversation. Big, capable, and prone to sounding sure of itself when it shouldn't be.
Data an AI developer has explicit legal permission to train on. It's the clean alternative to scraping and hoping. As lawsuits pile up, more builders are paying for the clean version.
The contract spelling out how one party may use another's content or data. It sets scope, duration, payment, and limits. Read the limits twice; that's where the surprises live.
Software that learns patterns from data instead of being told every rule by hand. Show it enough examples and it figures out the mapping itself. It's the engine under most of what gets called AI today.
Data about data. A photo's date, location, and camera; a document's author, format, and tags. Small, easy to ignore, and the difference between a searchable archive and a black hole.
Keeping metadata consistent, accurate, and usable across a whole content collection. Good management means you can actually find and trust your assets. Neglect it and your archive slowly turns to sludge.
When a model's performance decays over time because the world moved and the model didn't. The data it sees in production drifts away from the data it trained on. Left unwatched, a model quietly gets worse while everyone assumes it's fine.
Measuring how well a model performs before and after you ship it. You test it on held-out data with metrics that match the task. Skipping this is how bad models reach production.
AI that handles more than one kind of input or output: text, images, audio, video, together. A multimodal model can look at a photo and describe it, or read a chart and answer questions. It's closer to how people actually take in the world.
Several AI agents working together, or against each other, to get something done. They divide labor, negotiate, or specialize. Useful for problems too big or varied for one agent to handle alone.
Spotting and classifying the proper nouns in text: people, organizations, places, dates. It's a workhorse of language processing. Half of knowledge extraction starts with getting NER right.
The field that gets computers to work with human language. Translation, summarization, sentiment, search, and the chatbots you argue with. It's the bridge between how people write and how machines compute.
A model loosely inspired by the brain, built from layers of connected nodes that adjust as they learn. Each connection carries a weight that gets tuned during training. Stack enough layers and you get deep learning.
Optical Character Recognition, turning an image of text into text a computer can edit and search. It's what lets you search a scanned contract or digitize an old book. Modern OCR is good, though it still trips over bad scans and odd fonts.
A formal map of the concepts in a domain and how they relate. It defines the categories and rules so machines and people share the same vocabulary. Think of it as the grammar underneath a knowledge graph.
Of everything the model flagged as positive, how much it got right. High precision means few false alarms. You optimize for it when a wrong "yes" is costly, like accusing someone of fraud.
Records of which output people preferred when shown options. It's the raw material for teaching models human taste. RLHF runs on exactly this.
The input you give a model to steer its output. A question, an instruction, an example, or all three. The quality of what you get back tracks closely with the quality of what you put in.
The craft of writing prompts that get a model to do what you want, reliably. Part wording, part structure, part trial and error. It sits somewhere between programming and persuasion.
Agreements where publishers grant rights to their catalog, increasingly for AI training and retrieval. For many publishers it's a new revenue line built from content they already own. The terms are still being worked out across the industry.
The process of checking that outputs meet a standard before they go out. In data and AI work, that means testing labels, outputs, and pipelines for errors. QA is cheaper than the cleanup after a bad release.
Of all the actual positives out there, how many the model caught. High recall means few misses. You optimize for it when a missed "yes" is dangerous, like a disease screen that can't afford to overlook a case.
Training a model by reward and penalty as it acts in an environment. It learns a strategy through trial and error, chasing the reward. This is how AI learns to play games, steer robots, and more.
An umbrella term for building and using AI in ways that are fair, transparent, and accountable. It pulls ethics, safety, governance, and privacy into practice rather than talk. The word doing the heavy lifting is "responsible," and it's only real when someone's on the hook.
A setup where a model looks up relevant documents first, then generates an answer grounded in them. It lets a model use fresh or private information it was never trained on. Done well, it cuts hallucination and shows its sources.
Getting explicit permission to use a piece of content before you use it. It's the legwork of confirming who owns what and securing the yes. Tedious, and far cheaper than skipping it.
Tracking and enforcing who's allowed to do what with a body of content. It covers the licenses, the terms, and the systems that keep them straight. At scale this becomes its own discipline.
Training a model using feedback generated by another AI instead of humans. Faster and cheaper than paying people to rate everything. The catch is you're trusting one model to judge another, so the quality of the judge is everything.
Training a model to prefer outputs that humans rate highly. People rank responses, that becomes a reward signal, and the model learns to chase it. It's a big reason modern chatbots feel helpful instead of unhinged.
Testing a model specifically for harmful behavior before release: toxic output, dangerous instructions, easy jailbreaks. A different lens than accuracy, since a model can be sharp and still unsafe. You want to find the failure in the lab, not the wild.
The structured descriptive data attached to academic works: authors, affiliations, DOIs, citations, funding. It's what makes research discoverable and citable across the system. Standards keep it consistent so machines and databases can use it.
Labeling every pixel in an image by what it belongs to, so the model understands the scene at a fine grain. Not just "there's a road" but exactly which pixels are road. Heavier than bounding boxes and far more precise.
Attaching machine-readable meaning to content, marking what a term is, not just that it appears. Tagging "Apple" as a company versus a fruit, for instance. It's what lets software understand content instead of merely storing it.
A compact language model built to run cheaply, fast, and often on-device. It trades some breadth for efficiency and privacy. For a narrow, well-defined job, small often beats big.
Converting spoken audio into text. It's behind voice assistants, captions, and transcription. Good now in clear conditions, still fallible with accents, crosstalk, and noise.
Content organized into defined fields and hierarchy a machine can parse without guessing. The opposite of a wall of freeform text. Structure is what makes content reusable across formats and feedable to AI.
Training on labeled examples where each input comes with its correct answer. The model learns to map one to the other. It's the most common flavor of machine learning, and it lives or dies on label quality.
Artificially generated data used to train or test models when real data is scarce, private, or skewed. It can fill gaps and protect privacy. The risk is a model that learns a simulation and then stumbles on reality.
A hierarchical classification scheme, a way of sorting things into nested categories. It gives content a consistent backbone for organizing and finding it. Simpler than an ontology, and often the first structure a content team builds.
Marking up text with labels: entities, sentiment, categories, relationships. It's the fuel for training language models to understand written content. Slow, detailed work, and the quality shows in the result.
The unit a language model actually reads, usually a word or a chunk of one. Models measure input, output, and cost in tokens, not words. "Unbelievable" might be three tokens; the model doesn't see letters the way you do.
The full body of text a model learns from. Its size, mix, and quality shape what the model knows and how it's biased. The corpus is the model's whole education, for better and worse.
Reusing what a model learned on one task to jump-start another. You inherit the general knowledge instead of starting cold. It's why you can fine-tune a giant model on a small dataset and still get good results.
Learning from data with no labels, letting the model find structure on its own. It clusters, groups, and spots patterns nobody pointed out. Useful when labeling everything would be impossible or absurd.
A database built to store embeddings and find the nearest ones fast. Ask it for "things similar to this" and it answers in milliseconds across millions of items. It's the retrieval engine behind most RAG systems.
Labeling objects and actions across the frames of a video, often tracking them as they move. It's image annotation with time added, which makes it far more work. Anything that understands motion was trained on it.
Using XML's structured, tagged format to prepare content so machines can reliably parse it. Clean tags mean a model or pipeline knows what each piece of content is. It's a big reason structured publishing formats matter for AI training.
A human-readable format for configuration and structured data, common across AI and software workflows. It's designed to be easy to write and easy to read, using indentation instead of brackets. You'll meet it the first time you configure almost any modern tool.
When a model handles a task it was never explicitly trained on, generalizing from what it does know. No examples, just the instruction. It's a headline capability of large models, though less reliable than giving a few examples.
Asking a model to do something with only an instruction and no examples in the prompt. You're betting on what it already learned. Fast and often good enough; when it isn't, you add examples and move to few-shot.
