These are the words that actually change an answer. Each one is defined the way this field uses it, and links to the page where the question gets worked through in full.
The terms that decide whether you can sell what you made, and who pays if someone objects. These are the questions most of this site exists to answer.
A promise in a vendor contract to cover your legal costs and damages if the output you generated triggers an intellectual-property claim. Coverage is usually capped, usually limited to paid tiers, and usually void if you edited the prompt to imitate a specific artist or brand. It is the single clearest signal of how much confidence a vendor has in its own training data.
Where we cover itThe grant in a tool's terms that permits you to use its output to make money. It is a permission, not a transfer of ownership, and it is routinely tied to an active subscription: on several platforms the right to keep selling work you already made ends when you stop paying.
Where we cover itYou pay once for the right to use the work and owe nothing per use afterwards. It does not mean free, it does not mean unrestricted, and it does not mean you own it. A royalty-free licence can still forbid resale, redistribution or use in a logo.
Where we cover itWork that carries no copyright protection, because it expired, was never eligible, or was deliberately released. Anyone may use it for anything. This matters for AI output because purely machine-generated work with no human authorship may not be protectable in the first place, which is a very different position from owning it.
Where we cover itA defence in United States copyright law that permits some unlicensed use of protected work, weighed across four factors including the purpose of the use and its effect on the market for the original. It is decided case by case after the fact, so it is a position you argue, never a permission you hold in advance.
Where we cover itThe requirement that a work originate with a person before it can be registered for copyright. The United States Copyright Office has refused registration for output generated wholly by a machine, while accepting the human-authored parts of a mixed work. In practice this makes your selection, arrangement and editing the part you can claim.
Where we cover itA right, governed by state law in the United States, to control commercial use of your name, image and voice. It sits outside copyright entirely, which is why a voice model trained on a performer can be a problem even when no recording was copied.
Where we cover itAny recognisable representation of a specific person, including a face, a voice, or a distinctive mannerism. Synthetic likeness is the fault line under most avatar and voice tooling, because a model can produce a recognisable person without ever reproducing a protected recording.
Where we cover itThe corpus a model was fitted on. It determines what the model can produce, and it is the thing nearly every active lawsuit is actually about. Vendors that license or own their corpus tend to offer indemnification; vendors that scraped tend not to, which makes the indemnity clause a readable proxy for a disclosure nobody publishes.
Where we cover itStating that content was generated or materially assisted by AI. It is required in different forms by platforms, advertising regulators and some public bodies, and the obligations do not agree with one another. The safe assumption is that the strictest platform you publish on sets your standard.
Where we cover itA verifiable record of where a file came from and what was done to it. Provenance is the constructive alternative to detection: rather than guessing whether something is synthetic, it lets a file carry a signed account of its own history.
Where we cover itThe Coalition for Content Provenance and Authenticity, and the open standard it publishes for attaching signed provenance metadata to media files. Consumer-facing implementations are usually branded Content Credentials. The signature survives only while the metadata does, so an ordinary re-upload can strip it.
Where we cover itThe mechanics worth understanding before you compare two products, because most of the spec sheet is meaningless without them.
Writing and refining the instruction given to a model to steer what it produces. In practice most of the gain comes from being specific about the things the model has no way to guess, and from iterating rather than from any single phrasing trick.
Where we cover itThe act of running a trained model to produce an output, as opposed to training it. Inference is what you are buying: it costs the vendor compute per generation, which is why almost every creative tool meters usage instead of selling flat access.
Where we cover itIncreasing the resolution of an image or video, with a model inventing the detail that was never captured. Because the detail is generated rather than recovered, a good upscaler is judged on whether it stays faithful to the original rather than on how sharp the result looks.
Where we cover itGenerating moving footage directly from a written description. The limiting factors in practice are clip length, temporal consistency from frame to frame, and whether you can hold a character or a setting steady across separate generations.
Where we cover itBuilding a synthetic voice that reproduces a specific speaker from a sample of their recorded speech. The technical bar is now low enough that consent and provenance, not audio quality, are the parts that decide whether a given use is workable.
Where we cover itAligning a speaker's mouth movements to an audio track, whether the audio was generated, dubbed or recorded separately. It is the difference between a usable avatar video and an obviously synthetic one, and it degrades fastest on languages the model saw least.
Where we cover itSplitting a finished mix back into its component parts, typically vocals, drums, bass and everything else. It is what makes an AI-generated track editable after the fact, and the quality of the separation sets a hard ceiling on how much you can fix.
Where we cover itA format that stores musical events, which note, how hard, how long, rather than recorded sound. MIDI export is the feature that decides whether a generated piece is a finished asset or a starting point you can rearrange in your own software.
Where we cover itProducing the final output file from a generation, at full quality and full length. Render time and render queue position are where free tiers hurt most: the output is often identical to the paid tier, and it arrives much later.
Where we cover itThe terms that appear on a pricing page or an export dialog, where the difference between two plans usually hides.
The unit most creative-AI tools bill in, standing for one metered generation or some fraction of one. Credits are not comparable between vendors: a credit can mean a second of video on one platform and a whole image on another, so a price per credit tells you nothing until you know what one credit buys.
Where we cover itA visible mark burned into output, usually to distinguish a free tier from a paid one. Distinct from an invisible watermark, which encodes provenance data and is meant to survive editing. The first is a paywall, the second is a disclosure mechanism, and vendors use the same word for both.
Where we cover itThe neutral term for any content generated or substantially altered by a model. It covers everything from an upscaled photo to a fully generated performance, and it is the wording most platform policies and disclosure rules are actually written in.
Where we cover itSynthetic media that depicts a real, identifiable person doing or saying something they did not. The distinction from synthetic media generally is the identifiable person, and that is also what moves the question from platform policy into right of publicity and, in several jurisdictions, criminal law.
Where we cover itPairs of terms that get used interchangeably and are not interchangeable. Each one is a decision that changes what you are allowed to do.
Royalty-free is a LICENCE: someone still owns the work and has given you permission on terms, usually pay once and owe nothing per use. Public domain is an OWNERSHIP status: nobody holds copyright and no permission is needed. A royalty-free licence can be revoked, restricted or tied to a subscription; the public domain cannot.
Where this bitesA commercial licence is permission to USE output for money. It is not a transfer of copyright, and on several platforms it lapses when the subscription does, which can end your right to keep selling work you already made. Ownership would survive cancellation. Almost nothing in this field grants ownership.
Where this bitesSynthetic media is the neutral category: anything generated or substantially altered by a model, including an upscaled photo. A deepfake is the subset that depicts a real, identifiable person doing or saying something they did not. The identifiable person is what moves the question from platform policy into right of publicity and, in several jurisdictions, criminal law.
Where this bitesNearly every tool grants a commercial licence, which says you may sell the output. Very few offer indemnification, which says they will pay your legal costs if someone objects. The first is permission and costs the vendor nothing; the second is a financial commitment and is the clearest signal of confidence in their training data.
Where this bitesPublic domain means the work is unprotected and you need no permission. Fair use is a defence you raise about a work that IS protected, weighed after the fact across four factors. One is a status you can verify in advance; the other is an argument a court accepts or rejects once you are already being sued.
Where this bitesHuman authorship decides whether a work is REGISTRABLE for copyright at all. Getting over that bar is not the same as owning the whole output: the Copyright Office has accepted the human-authored parts of a mixed work while refusing the machine-generated parts. What you can claim is your selection, arrangement and editing.
Where this bitesVendors use one word for two opposite jobs. A visible watermark is a PAYWALL, burned in to distinguish a free tier from a paid one and removed when you pay. Provenance metadata is a DISCLOSURE, meant to survive editing and travel with the file. Paying removes the first and should never remove the second.
Where this bitesCopyright protects a fixed work; right of publicity protects a person. A voice model trained on a performer can infringe the second while touching neither a recording nor a composition, which is why "we never copied any audio" is not the defence it sounds like. They are different laws with different owners and different remedies.
Where this bitesThe rights entries describe how these terms are used in the creative-AI industry. They are not legal advice, and licence terms change without notice. Check the vendor terms that apply to your account before you rely on any of it.