AI writing glossary

Using natural-language conversation and revision to guide AI from a rough idea toward finished writing. Vibe Writing Hub adds safeguards: a person must judge the output, check it, and take responsibility for the final work. The term is still developing and has no single accepted definition.

Keep separate: More specific than AI-assisted writing. Common public use may not include the Hub's safeguards.

A term Andrej Karpathy popularized in 2025 for directing AI-generated software in everyday language while paying little attention to the code. The term is now used more broadly, but carefully reviewed AI-assisted development is different from that original, deliberately loose practice.

Keep separate: Writing can borrow iteration and version control from coding, but code tests cannot fully test truth, meaning, or voice.

Writing made with help from AI at any stage, including questions, outlines, drafts, translation, editing, or review. The label only says AI helped. It does not decide who the author is, whether the use was allowed, or whether it must be disclosed.

Keep separate: A broad category, vibe writing is one iterative, human-directed approach within it.

AI designed to create content such as text, images, audio, video, or code from an input. The output is newly generated, but that does not prove its ideas, facts, or wording are original, accurate, or free of legal restrictions.

A model trained on patterns in language so it can generate or assess text. An LLM is a language model trained at a large scale. Its size and fluent wording do not make its output evidence or prove that it understands the subject or is telling the truth.

Input used to guide a model's response. It can include instructions, questions, sources, examples, limits, or a requested format. A prompt is only one part of the context available to the model, and it cannot guarantee that the model will follow every direction.

Guiding a task through several turns. After reviewing each response, the person makes the next request more specific. Good iteration keeps settled decisions clear and gives each turn a purpose. It is more than repeatedly asking the model to 'make it better.'

All information available to a model when it creates a response. Depending on the system, this may include current and earlier messages, system instructions, retrieved documents, tool results, examples, and service instructions hidden from the user. The model may not use all available information equally or correctly.

The limited amount of input and generated content a model can process for one response. It is often measured in tokens, which are pieces of words or text. Even material inside this limit can be missed or used unevenly. A large window gives more capacity, not dependable memory or attention.

The Hub's term for when an AI conversation stops following the writer's current goal, facts, limits, or settled choices. Old instructions or AI-made assumptions start shaping later revisions. This is a practical workflow diagnosis, not a standard scientific measure of a model.

An AI agent is usually a system set up to pursue a goal in several steps, often choosing actions or using tools with some independence. Agentic writing uses this approach for research, drafting, or revision. There is no settled boundary between an agent and an automated workflow, and the label does not remove the need for review or human accountability.

A short record kept outside the AI conversation and left unchanged. It states the reader, purpose, known facts, and assumptions the writing must not make. The writer can compare AI output with this independent starting point and spot unsupported additions.

A possible structure, passage, question, or edit produced by AI for a person to review. It is not approved writing or evidence. Calling it a proposal separates generation from the person's decision to accept, adapt, reject, check, or replace it.

A real examination by someone with enough skill and authority for the task. The person checks the output against its purpose, sources, effects, rules, audience, and risks. Simply reading smooth prose or clicking approve may not count as meaningful review.

A short, up-to-date record of confirmed facts, wording that must stay, settled choices, open questions, and the next step. It carries current decisions forward without treating the whole conversation, including rejected ideas, as trustworthy instructions.

A revision request that says what may change, what must stay fixed, and how the result will be checked. A smaller change makes unwanted side effects easier to find, but it still cannot guarantee that the model will follow the request.

The Hub's four-step model: the person sets the goal and limits, AI offers options, the person reviews and decides, and a responsible person approves release and handles corrections. It combines editorial practices and is not a standard scientific process.

The Hub's check for prompting over several turns: Target one decision, Hold accepted decisions, Request a bounded change, Examine the difference, Abandon or branch when needed, and Decide whether another turn is worthwhile. It is a practical framework, not a formally validated standard.

A system setup in which a person handles a defined step in an automated process, such as approval, correction, or deciding when to seek more help. The label does not tell us whether that person has enough time, information, skill, authority, or real power to intervene.

Keep separate: Having a person in the process does not guarantee effective oversight.

A statement, directly made or reasonably implied, that evidence can check. Common examples include names, dates, numbers, quotations, causes, titles, conditions, and current rules. Opinions can also rely on facts that need checking.

Breaking writing into claims that can be checked, tracing important claims to suitable evidence, and testing whether that evidence supports the exact wording. Unsupported claims must be revised, qualified, removed, or referred for further review. Fact-checking uses outside evidence, it is not a judgment that the writing sounds convincing.

The record closest to the information being checked, such as a full study, original data, an official rule, a direct transcript, or a firsthand account. 'Primary' means close to the subject. It does not mean neutral or sufficient, so the source's authority, purpose, completeness, and date still matter.

How closely a source supports the exact claim attached to it, including the claim's strength, scope, date, and tense. A source can be genuine and respected but still not support that particular statement.

Using more evidence from genuinely independent sources to confirm, question, or clarify a claim. Two links are not independent support if both repeat the same press release, dataset, anonymous account, or AI output. Claims with greater consequences or uncertainty need stronger confirmation.

Clearly identifying where information, wording, ideas, or contributions came from. Attribution names the source, provenance is the broader record of an item's origin and history. Both can help readers judge credibility. But naming or citing a source does not by itself prove a claim, grant permission to use material, or excuse too much copying.

NIST's preferred risk term for AI-generated content that confidently presents false or mistaken information, including made-up citations or misleading reasoning. The term describes a failure in the output without suggesting that the model has human-like perceptions.

Keep separate: Often called hallucination, neither term supplies an error rate for every model or task.

A common but human-like label for AI-generated content that is false, made up, or unsupported. Institutions use different terms. This glossary prefers NIST's term 'confabulation' when precision matters. Neither word means that every model error has the same cause.

Asking a model to criticize, check, or revise output made by itself or another model. This can point out possible problems and suggest what to investigate, but the review is still another model response. It is not independent proof that an important claim is correct.

The link between a person and the ideas and expression presented as that person's work. The rules differ by field and law. For responsible AI-assisted writing, a byline alone is not enough: the person should control important choices and be able to explain and support the result.

The overall impression created by what a writer chooses, values, claims, qualifies, leaves out, and takes responsibility for, expressed through language. Voice is not a verbal fingerprint or just a set of habits that a model can fully recover from writing samples.

The social or emotional attitude a passage shows toward its subject and reader, such as warm, formal, skeptical, restrained, or urgent. Tone can change within one work while its broader style and authorial voice remain the same.

Repeated choices in words, sentence structure, pace, organization, imagery, and conventions. A model can describe and copy surface style, but a style match does not prove the same experience, judgment, identity, or authorship.

The Hub's term for a revision that moves away from the writer's intended position, detail, emphasis, relationship with the reader, or way of expressing ideas. It is not a biometric measure based on traits used to identify a person, and a style score alone cannot prove it.

Presenting another person's words or ideas without the credit expected in the relevant school, profession, or publication. It is an integrity issue, and its rules vary by context. It is not the same as textual similarity or copyright infringement.

A measured or noticed resemblance between a text and material in a set used for comparison. A match may be a quotation, common wording, independent overlap, or improper copying. A tool may also miss paraphrases or sources it cannot access. Similarity is a clue, not a verdict.

Law that protects qualifying original expression and sets rules for rights, ownership, licenses, exceptions, and infringement. The law differs by jurisdiction. In the United States, AI help does not prevent protection for human-authored expression, but material created purely by AI is not protected. Other jurisdictions differ. This is not legal advice.

A person's informed and freely given agreement to a specific use. For AI-assisted writing, meaningful consent may require explaining the tool, purpose, uploaded material, who will receive it or provide the service, and relevant risks. Permission to read or edit a document does not automatically permit sending it to an outside AI system.

Information that must, or is reasonably expected to, have limited access and disclosure. That duty may come from law, a contract, professional rules, employment, or a relationship. Whether an AI system is suitable depends on the actual duty and the system's real controls. Get qualified advice when needed.

Telling the relevant people how AI contributed, with the detail and placement required by policy, law, method, or reasonable audience expectations. A useful disclosure names the important contribution. It does not make forbidden, false, non-consensual, or infringing work acceptable.

The duty and real ability to answer for a result: approve what it means, support its claims, respond to people it affects, and correct harm or errors. AI cannot take on this responsibility. It must remain assigned to people or organizations that can hold legal duties.

A setup that allows qualified people to understand, monitor, question, stop, or correct an AI-supported process based on its possible effects. Real oversight requires information, authority, skill, time, and a working way to intervene, not just a person somewhere in the process.

How important information would be to a reader's, evaluator's, client's, or collaborator's understanding or decision. The Hub uses this idea when considering disclosure: would a reasonable affected person care that AI significantly shaped the claims, argument, reporting, evaluation, personal voice, or presented words?

Keep separate: This editorial use does not replace specialized legal or accounting tests.