Terms, A to W
Every term below opens with its definition. Filter by name, alias or definition, or jump to a letter.
78 terms, A to W
A
- AB 2013 (California)#
- A California law, in force since January 1, 2026, that requires developers of generative AI systems to publish documentation of their training data. It must say where the datasets came from and whether they were purchased or licensed, so keep provenance records for every dataset you license. Frankfurt Kurnit Klein & Selz, (opens in a new tab)
- Data licensing
- Anonymization#
- Processing personal data so that the people in it can no longer be identified. Effectively anonymized data falls outside UK data protection law; data that can still be linked back to people is only pseudonymized. ICO, (opens in a new tab)
- Trust
- Appraisal#
- An indicative price range for your data, worked out from public prices and your answers before anyone sees the data. It is not an offer: a firm price follows an NDA, reports such as a Repo Check share.json or a system inventory, and licensee review.
- Free appraisal
- Assignment for the benefit of creditors (ABC)#
- A US alternative to bankruptcy in which a company transfers its assets to an assignee, who sells them and pays the creditors. Code and workspace data can be among those assets, so the assignee may be the one who licenses them.
- Shut-down startups
B
- Benchmark#
- A fixed set of tasks with known answers used to compare AI models. Once its items leak into training data the scores stop meaning much, which is why AI developers pay for test material that has never been public.
- Who licenses it, and for what
- Biometric data#
- Data about a person's body or voice that can identify them, such as voiceprints and face geometry. Laws such as Illinois' Biometric Information Privacy Act require consent, so voice and video licenses usually need releases that name AI training. Biometric Update, (opens in a new tab)
- Call and meeting recordings
- Bona vacantia#
- Property left in a UK company when it is dissolved, which passes to the Crown under section 1012 of the Companies Act 2006. Data and code still held by the company go with it, so license or transfer them before the company is struck off. legislation.gov.uk, (opens in a new tab)
- After the export: what to keep, and who can sign
C
- Chain of title#
- The documents that show the licensor owns the data or has the right to license it: employee and contractor IP assignments, client consents, author or talent agreements and platform terms. Gather it before you license: it is what lets a licensee rely on your right to grant the license.
- Trust
- Client consent letter#
- Written permission from a client to license code or data you created for them. Agencies need one for client work unless the contract already gives them ownership, and an unpaid invoice does not by itself give you the rights.
- Agency project archives
- Company archive#
- An export of a company's operational data across its tools and years, licensed as one dataset. Headcount, years of records and the number of linked tools set most of its price.
- Company archives
- Confidence level#
- How well a price in the Rightmark Price Index is known: High for published rates, Medium for reported and program ranges, Low for inferred estimates. Read Low rows as rough.
- Price Index
- Copyleft license#
- An open-source license, such as those in the GPL family, whose terms follow the code into anything built from it. Copied copyleft components must be removed or disclosed before code is licensed, because undisclosed copyleft can unwind a deal.
- Repo Check
D
- Data bounty#
- A published request for a specific type of dataset at a stated price range, listed on the Wanted board. Bounties show what is in demand right now.
- Wanted
- Data card#
- A short document that describes a dataset: what it contains, where it came from, dates, languages, formats, labels and known gaps. Keep one with every dataset you deliver.
- Buyers
- Data inventory#
- A list of every system in a company archive with its admin owner, years covered, users or mailboxes and rough volumes. Have someone technical fill one in before you discuss terms: it is how an archive gets scoped.
- Operational data
- Data licensing#
- Granting someone the right to use your data under set terms while you keep ownership. An AI data license names the permitted uses, the term, whether the licensee may sublicense and whether it is exclusive.
- Data licensing
- Data monetization#
- Earning revenue from data you already hold, by selling insights, selling access or licensing copies to AI developers.
- Data monetization in the AI era
- Data processing agreement (DPA)#
- A contract that sets how a processor handles personal data on your behalf: purpose, security, subprocessors, retention and deletion. Ask for one whenever a licensee or vendor will handle personal data before it is de-identified.
- Trust
- Data provenance#
- The documented history of where data came from, who created it, what rights attach to it and how it was handled. It matters because laws such as California's AB 2013 make developers disclose their training-data sources.
- Trust
- Decontamination check#
- Evidence that data was never public and was not written by an AI model. It matters because public or model-written material adds little to training and can spoil evaluations.
- Buyers
- De-identification#
- Removing or masking information that identifies people so that they cannot reasonably be identified. US health data is de-identified by HIPAA Safe Harbor or Expert Determination; company archives by removing names, contact details and other personal data.
- Trust
- Deletion certificate#
- A licensee's signed confirmation that licensed data was deleted at the end of the term or on request. Ask for the right to one in every license.
- Trust
- Derived dataset#
- A dataset built from licensed data, such as labels, embeddings, training tasks or synthetic rewrites. A license should say whether the licensee may keep or sell derived datasets after the term, and whether exclusivity covers them.
- Data licensing
- DICOM#
- The standard file format for medical images, which stores each image with a header of metadata. Headers and text burned into the pixels can hold patient details that must be removed before licensing.
- Medical imaging with reports
E
- eDiscovery#
- Search and export tools built into office suites, such as Purview eDiscovery for Exchange, OneDrive and SharePoint, and Vault for Gmail and Drive. They search and export mailboxes, chats and files across a whole organization rather than one account after another.
- Outlook and OneDrive
- Egocentric video#
- First-person video of people doing real tasks, recorded from a head or chest camera. Robotics teams license it to teach robots how hands handle objects.
- First-person task video
- Evaluation set#
- Data held back to test models rather than train them. It loses its value once it leaks into training, so AI developers pay for material that has never been public.
- Who licenses it, and for what
- Exclusive license#
- A license that gives one licensee sole use of the data for its term and scope. It is priced at a premium agreed deal by deal. Company-archive lump sums usually require one for a term; code is licensed non-exclusively by default.
- Data licensing
- Expert Determination (HIPAA)#
- A HIPAA method for de-identifying health data, the alternative to Safe Harbor. A qualified expert analyzes the dataset and finds that the risk of identifying anyone is very small; unlike Safe Harbor, it can keep useful details such as dates. HHS, (opens in a new tab)
- Clinical notes and EHR data
F
- Fine-tuning#
- Further training of an existing model on a smaller, focused dataset to teach it a task, a domain or a style.
- Who licenses it, and for what
G
- Git history#
- The full record of commits, authors and changes in a repository. At standing rates it often carries more than half of a repository's price, so squashing or migrating it away lowers the price.
- Repo Check
- Grounding#
- Using licensed content when a model answers a question, such as news articles retrieved and cited, rather than to train the model. A license should say whether it covers training, grounding or both.
- News and magazine archives
H
- HIPAA#
- The US law whose Privacy Rule governs health information held by healthcare providers, health plans, clearinghouses and their business associates. Health data normally moves to AI developers only after de-identification, after which the Privacy Rule no longer applies to it. HHS, (opens in a new tab)
- Health and life sciences
I
- Indemnity#
- A promise to cover the other side's losses if a stated risk happens, such as a claim that you did not own the data. Some code agreements leave the owner's indemnity uncapped, so license only what you are sure you own.
- Data licensing
- Introduction fee#
- A one-off payment for an introduction. A Rightmark Scout's share is set by the Scout terms instead.
- Scouts
- IP assignment#
- A signed agreement that transfers the rights in someone's work, such as an employee's or a contractor's code, to the company. Collect assignments from every contributor, so your chain of title is complete.
- Private codebases
L
- Legal hold#
- A setting that preserves a person's mailbox, chats and files even if items are deleted. In Exchange Online, a hold placed before an account is deleted keeps the mailbox as an inactive mailbox; in Gmail and Drive, deleting the user removes the data and its holds, so keep leavers' accounts instead.
- Outlook and OneDrive
- License#
- Permission to use something on stated terms, such as data for training AI models; to license is to grant that permission. The owner keeps ownership, and the license sets what the licensee may do, for how long and where.
- Data licensing
- License types#
- The main forms of AI data license: a non-exclusive license paid per sale, an exclusive license for a term (usually a lump sum), a revenue share, an annual fee for ongoing access, and a bounty for data made to order.
- Data licensing
- Lump sum#
- A single payment for a license, usually made after delivery and approval. Company-archive lump sums usually require exclusivity for a term.
- Data licensing
M
- MBOX and PST#
- Standard file formats for exported email. MBOX is an open format that many tools read, and PST is Outlook's data file; Vault exports either for Gmail, and Purview exports PST or individual .msg files for Outlook.
- Gmail and Drive
- Mirror clone#
- A copy of a git repository made with
git clone --mirror, which brings every branch, tag and ref with the full history. It is how to preserve code before you close a hosting account. GitHub Docs, (opens in a new tab) - Export guides
N
- Non-exclusive license#
- A license that leaves the owner free to license the same data to others. It is the norm for code, which can be licensed, and paid for, again with each new licensee.
- Data licensing
O
- Operational data#
- The everyday record of how a company works: email, chat, calendars, files, CRM records, tickets, wikis and meeting recordings. Licensed together across the same years, it becomes a company archive.
- Operational data
- Ownership pack#
- The documents that show a dataset can be licensed. What goes in depends on the route: for code, the contributor, license and secrets checks from Repo Check plus the owner's IP documents; for company archives, the signer's authority, a scope note and a system inventory.
- Buyers
- Own-use license#
- A right for a licensee or an intermediary platform to use the data for its own models or products, not only to pass it on. Whether one applies differs by route, so check it for yours, and whether it pays you.
- Data licensing
P
- Permitted use#
- The clause that names what a licensee may do with licensed data, such as training and evaluating AI models. A vague or any-purpose use is a warning sign.
- Data licensing
- Per-unit pricing#
- Pricing data by the unit AI developers use, such as per repository, per company, per hour, per image or per title, rather than per gigabyte. Each data type has its own unit.
- Data valuation
- PHI (protected health information)#
- Individually identifiable health information held by a HIPAA covered entity or one of its business associates. Once it has been properly de-identified, it is no longer PHI. HHS, (opens in a new tab)
- Health and life sciences
- Pretraining#
- The first and largest phase of training, in which a model learns general patterns from a very large corpus such as web text, books and code.
- Who licenses it, and for what
- Pseudonymization#
- Replacing identifiers such as names with consistent codes and keeping the key separately. Under the UK GDPR, pseudonymized data is still personal data for anyone who holds the key. ICO, (opens in a new tab)
- Trust
R
- Read-only scan#
- A scan that reads your data or code without changing anything, to find secrets, personal data and third-party rights. It runs on your side, as Repo Check does on your own computer, or inside the licensee's pipeline.
- Trust
- Repo Check#
- Rightmark's free script that measures a git repository on your own computer and writes a numbers-only share.json for a quote. It never uploads your source code.
- Repo Check
- Retention policy#
- A setting that deletes messages or files after a set period, such as the retention rules in Slack or Exchange Online. History a policy has already deleted cannot be exported or licensed, so check retention before you rely on an archive.
- Export guides
- Rightmark Price Index#
- Rightmark's quarterly table of what AI developers pay for private data, by data type and unit, with a source and a confidence level on every row. The current edition is Q3 2026.
- Price Index
- Rightmark Scout#
- A person who introduces data owners to Rightmark and earns a share of Rightmark's income from those owners, under the Scout terms. Scouts are paid from Rightmark's own income, so owners pay nothing for the introduction.
- Scouts
- RL environment#
- A simulated workplace built from real material, such as a codebase with its tickets and chat, where an AI agent practices tasks and is scored automatically. Reinforcement learning, training by trial and reward, is what these environments are for.
- Who licenses it, and for what
S
- Safe Harbor (HIPAA)#
- The HIPAA de-identification method that removes a fixed list of identifier types, such as names, addresses and record numbers, and keeps only the year of dates linked to a person. The holder must also have no actual knowledge that what remains could identify someone. HHS, (opens in a new tab)
- Clinical notes and EHR data
- Sample review#
- The owner's chance to inspect representative samples of processed data before it is used. Company-data licensees usually offer one, though details vary.
- Trust
- Sanctions screening#
- Checking owners, licensees and their ultimate owners against sanctions lists, such as OFAC's and the EU's, including the ownership rules that extend sanctions to companies owned by sanctioned persons. How screening applies to owners and licensees is being set with counsel; see Trust.
- Trust
- Secrets scanning#
- Searching code and its whole history for keys, passwords and tokens before anything is shared. Live secrets must be rotated, not only deleted, because old commits keep them.
- Repo Check
- Squashed history#
- Git history collapsed into a few commits, often by a migration or an import. It removes much of what AI developers pay for, so find the original repository if it still exists.
- Repo Check
- Standing-rate license#
- A non-exclusive license for code, priced per sale by the Rightmark standing-rate model for the Q3 2026 edition: per token of original code and, at a lower rate, per token of git history. The same repository can be licensed again to each new licensee.
- Private codebases
- Sublicensing#
- A licensee's right to grant rights in your data to others. If a licensee or platform may sublicense, the license should say to whom, for what, and whether you are paid for every grant.
- Data licensing
- Synthetic data#
- Data generated by a model or a simulation rather than recorded from the real world. It scales common patterns cheaply but inherits the blind spots of whatever generated it.
- Synthetic data and its limits
T
- Task-grade repository#
- A repository that builds, runs its tests and carries real bug-fix history, so AI developers can turn it into checkable training tasks. Such repositories are indexed at $10,000 to $100,000+ (Low confidence, estimated) against $150 to $1,500 (Medium confidence) for a standing-rate license. Indicative estimate, not an offer. Task value is priced only after review.
- Private codebases
- Teleoperation data#
- Recordings of a person controlling a robot remotely, with the robot's cameras, joint positions and often force sensors. Robotics teams use it to teach robots by demonstration.
- Teleoperation and robot logs
- Term#
- How long a license lasts and what happens when it ends. Exclusive company-archive licenses run for a fixed number of years agreed deal by deal; on code routes, licenses already granted survive even if you withdraw unlicensed code.
- Data licensing
- Token#
- The unit a language model reads: a short piece of text, often part of a word. Code licenses are often priced per token of original code and of git history.
- Data valuation
- Training data#
- The text, code, images, audio, video and logs a model learns from. The term also covers the tests and tasks used to evaluate and improve models after their first training.
- AI training data
- Training task#
- A problem built from real work with a checkable answer, used to train or test an AI model. AI developers pay $200 to $2,000 per task built from real software work (Medium confidence). Epoch AI, (opens in a new tab)
- Rightmark Price Index: Training task built from real code
- Translation memory#
- A database of sentences paired with their human translations, usually stored as TMX files. Specialist domains and rare language pairs keep their value; common pairs have lost most of theirs.
- Translation memories
U
- US bulk data rule#
- The US Department of Justice rule, in force since April 8, 2025, that restricts transactions giving countries of concern access to Americans' bulk genomic, biometric, health, financial, geolocation and other sensitive personal data. US Department of Justice, (opens in a new tab)
- Trust
W
- Warranty#
- A statement in a contract that something is true, such as that you own the data and have the right to license it. Owners warrant that they own what they license; if a warranty proves false, the indemnity decides who pays.
- Data licensing
- Whole-slide image#
- A high-resolution digital scan of an entire pathology glass slide. Slides linked to diagnoses and outcomes are what oncology AI teams license.
- Pathology slides
No term matches that. Try a shorter word, or ask us about your data.
Questions
What is de-identified data?
Data from which information that identifies people has been removed or masked, so individuals cannot reasonably be identified. Under HIPAA, health data is de-identified by Safe Harbor (removing a fixed list of identifier types) or Expert Determination (a qualified expert finds a very small risk of re-identification).
Source: HHS, (opens in a new tab)
What is data provenance?
The documented history of where data came from, who created it, what rights attach to it and how it was handled. It matters because laws such as California's AB 2013 make developers disclose their training-data sources.
What is HIPAA Expert Determination?
A HIPAA method for de-identifying health data, the alternative to Safe Harbor. A qualified expert analyzes the dataset and finds that the risk of identifying anyone is very small. Unlike Safe Harbor, it can keep useful details such as dates.
Source: HHS, (opens in a new tab)
What is a non-exclusive license?
A license that leaves the owner free to license the same data to others. It is the norm for code. An exclusive license gives one licensee sole use for its term and scope, is usual for company-archive lump sums, and is priced at a premium agreed deal by deal.
What is chain of title?
The documents that show the licensor owns the data or has the right to license it: employee and contractor IP assignments, client consents, author or talent agreements and platform terms. Gather it before you license: it is what lets a licensee rely on your right to grant the license.
Sources
- HHS, (opens in a new tab)
- Frankfurt Kurnit Klein & Selz, (opens in a new tab)
- ICO, (opens in a new tab)
- ICO, (opens in a new tab)
- Biometric Update, (opens in a new tab)
- legislation.gov.uk, (opens in a new tab)
- GitHub Docs, (opens in a new tab)
- Epoch AI, (opens in a new tab)
- US Department of Justice, (opens in a new tab)