AI training data glossary: the market's words, defined

This glossary defines the words used when a company licenses or sells its data to AI labs: who the parties are, what kinds of data change hands, how prices are stated, what a licence says, which privacy rules apply and how a sale runs. Each entry gives a plain definition, one example from our records and a link to the page that goes deeper. Three distinctions trip up readers more than any single word. First, a price always sits on one side of the market: what a worker is paid to make data, what a buyer offers for existing data, what a licence actually closed at, what a catalogue asks and what a vendor charges to collect to order are different numbers for the same unit, and for an hour of speech they run from $24 per hour of audio paid to the speaker to $122 per hour of audio charged by a vendor. Second, a licence is not a sale: after a licence the holder still owns the data, while a sale or assignment moves ownership, and under US copyright law an exclusive licence also transfers the rights it covers. Third, exclusive is not non-exclusive: an exclusive licence stops the holder licensing the same data to anyone else and fetches 2.5 to 5 times a non-exclusive one, while a non-exclusive licence lets the holder license the same data elsewhere.

The glossary is written for a founder, CFO, counsel or journalist meeting these words for the first time, and it uses the words the way buyers, laws and our own agreements use them. Where a buyer's term differs from ours, the entry says so. Every figure links to the record or computation behind it, so you can check the number before you rely on it; the method is in how we source every number, and the whole body of evidence sits under the research map.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. Where an entry describes how we work, it links the clause of our mandate that sets it.

This glossary is research from public sources, not legal advice.

Contents

1. The market and its parties
2. Kinds of data
3. Units, prices and the sides of the market
4. Licence terms
5. Rights and privacy
6. How a sale runs
7. Questions people ask
8. What this means for you

1. The market and its parties

The market has more parties than a buyer and a holder. Many companies that pay for data license it onward, some only list it, and a few act for one side alone. Knowing which party you are talking to tells you whose interest it serves.

AI lab. A company that trains large models and is the end user of most training data: OpenAI, Anthropic, Google, Meta and a handful of others. Labs sign some licences themselves and buy much of their data through vendors. OpenAI alone has signed 29 of the disclosed licensing deals on record, out of 111 in all, struck with 44 different licensees. Deeper: disclosed deals and how AI companies get training data.

Buyer. Any company that acquires, or may acquire, rights in a holder's data, with its group companies. An AI lab is one kind of buyer; the others are data vendors, resellers, exchanges and research groups. Our board tracks 41 buyers in 5 kinds, and 25 of them take a company's records. Deeper: the buyer board and buyer programs compared.

Holder. The company or person that holds the data and decides whether to sell or license it. It is our word for the party a buyer's pages may call "the seller", "the licensor" or "the data owner". A holder can be a living business, a closed company acting through its directors or a liquidator, or an individual such as a founder without a company. We describe 8 kinds of holder on their own pages. Deeper: what labs buy.

Data vendor (and annotation vendor). A company that sells data, or the work of making it, to AI labs: human data firms, environment builders, dataset publishers and annotation vendors. An annotation vendor is paid to label data a lab already has (boxes on images, transcripts of audio, segments of video, rankings of model answers), usually per item or per second of media, as in one vendor's rate of $0.08 to $0.25 per second for labelling first-person footage. Some vendors build from data licensed from the companies that hold it; most make it to order. Our register lists 150 vendors, of which 30 say they pay for data. Deeper: the vendor register and AI training data companies.

Human data company. A vendor whose product is the work of people, often credentialed experts, who write, judge or demonstrate for a lab: Mercor, Scale AI, Surge, Handshake AI, micro1. These are among the fastest growing firms in the market; Mercor reported over $2 billion in gross revenue run rate in June 2026. Our register counts 83 vendors selling expert and human data. Deeper: expert work below and the vendor register.

Reseller. A buyer that licenses a holder's data onward to labs rather than keeping it for its own models, and pays the holder out of what the labs pay. Mercor's catalogue, for one, offers every dataset to labs for licensing. A holder licensing to a reseller carries the end buyer's credit risk and should ask who the onward recipients are. Deeper: when AI data buyers pay.

Data exchange. A site where anyone can list a dataset at an asking price and buyers browse and buy, with the exchange keeping a share of each sale; Datarade is the best known, and it lets the holder set the price and keeps a share. Prices on an exchange are asking prices, the "listing" side of the market below. Deeper: data on offer.

Sell-side broker. An agent appointed by the holder to find buyers, run the sale and negotiate, paid by the holder and acting for it alone. It is not a data broker in the sense state registries use (see data broker registry): it does not buy the data, take title to it or license it in its own name ( clause 2.3). That is the role AIData.Broker takes: under our mandate we act for the holder alone and take no instructions from any buyer about the data ( clause 2.2). The alternative is a direct sale, in which the holder licenses to one buyer on that buyer's own program, which is quicker for a single buyer and leaves the holder to compare terms and negotiate alone. Deeper: how it works and direct or through a broker.

Wind-down service. A firm that helps a company close, and offers to sell its remaining assets, data included, as part of the service. SimpleClosure's AssetHub, for example, says unique assets fetch tens to hundreds of thousands of dollars. Deeper: closed and insolvent companies and SimpleClosure alternatives.

Bankruptcy estate. The assets of a company in bankruptcy, run by a debtor in possession or a trustee and sold under the court's supervision. The largest public offers for one company's records so far came from one: Google's $10M winning bid at auction for Spirit Airlines' de-identified business records, still awaiting the court's approval when we read it, and a late offer of $12.5M from micro1. Deeper: selling customer data in bankruptcy.

2. Kinds of data

Labs buy data for different stages of training, and the stage decides what the data must look like and what it is worth. We price 9 kinds of data and have catalogued 156 narrower kinds, 95 of which a buyer has named.

Pretraining data. The very large, mostly general body of text, code, images or video a model first learns from, measured in tokens or hours. Volumes are enormous: Alibaba's Qwen3 models were trained on about 36 trillion tokens. It is bought in bulk at low prices per unit, and most of the court cases over training concern data of this kind. Deeper: where data comes from and fair use and AI training.

Fine-tuning data (SFT). Examples of the exact behaviour a lab wants, usually a prompt paired with a model answer written or checked by a person, used after pretraining to teach a model a task or a style. SFT stands for supervised fine-tuning. One vendor prices a constructed entry at about $6 per entry. Deeper: AI training data prices.

Preference data (RLHF). Judgments of which of two or more model answers is better, used to train a reward model that then steers the main model; RLHF stands for reinforcement learning from human feedback. Price follows the judge: a crowdsourced ranking costs $0.50 to $2 and one by a domain expert $5 to $10. Deeper: AI training data prices.

Evaluation set. A held-back set of questions or tasks, with known answers, used to measure a model rather than train it; a public one is a benchmark. Labs pay most for private sets a model cannot have seen, because, as one practitioner put it, public benchmarks have leaked into pretraining data. Our register counts 122 vendors selling evaluations. Deeper: the vendor register.

RL environment. A simulated workplace (a code base, a spreadsheet, a browser, a robot simulator) with tasks and an automatic grader, in which a model practises by trial and error. It is the newest large category: 89 vendors on our register sell environments, and public prices for one run $1,200 to $2,500 per environment. Two parts make an environment trainable: the verifier, the check that decides automatically whether a model's attempt succeeded, and the gold solution, a known correct answer that proves the task can be solved. micro1's job posts for coding environments ask builders for a golden reference solution for each problem. Deeper: RL environments.

Company records. The working files of an operating business: email, chat, documents, tickets, CRM entries, code and the records of how work was done. Labs want them to teach models how real work happens. Public prices for one company's records run $174K to $693K per company, and our valuation draws on 109 records of what whole companies' records sold for or were offered at. Deeper: how to sell company data to AI labs and how much company data is worth.

Source code. A company's repositories with their history: commits, pull requests, reviews and linked issues, which show a model how code changes and why. Public prices run $934 to $24.2K per repository, and some buyers pay far more for old code bases, one by 10x its cap on a modern one. Deeper: sell your codebase to AI companies.

Expert work. Time bought from people with professional knowledge (doctors, lawyers, engineers, analysts) to write, grade or demonstrate for a model. It is priced per task or per hour of work, the unit for time rather than output; public prices run $55 to $101 per hour of work, the people who do the work are paid a median $80 per hour of work an hour, and a lab pays Mercor $160 to $187 an hour for one expert hour, of which Mercor keeps 30% to 40%. Deeper: market rates.

Egocentric video. Footage filmed from the point of view of the person doing a task, with a head or chest camera or glasses, used to teach robots and world models how hands and objects move. Made to order, a finished hour from a simple GoPro rig costs $15 to $20, and our valuation prices first-person footage captured to a spec at $15 to $40 an hour. Deeper: sell videos to AI companies.

Teleoperation demonstration. A recording of a person steering a robot through a task by remote control, capturing the robot's own sensors and motions, so a model can imitate it. It is priced per episode, per demonstration or per hour; one collector charges $1 to $3 per episode for mid-complexity manipulation, and made to order a demonstration median is $13 per demonstration. Public prices for physical-world data run $4.48 to $25 per hour of video. Deeper: where data comes from.

Speech and audio. Recorded speech with transcripts, conversations, call recordings and music, priced per hour. Public prices run $65 to $266 per hour of audio, and custom recorded speech costs 2.7x catalogue speech at the top of each range. Call recordings raise consent questions; see all-party consent. Deeper: sell audio and speech data.

Text archive. A body of written work held by one party: a publisher's back catalogue, a forum's posts, a news archive. The archetype is Reddit, which licensed its content to Google for a reported $60 million a year. Public prices per book or title run $881 to $3,082 per book or title. Deeper: Reddit's data licensing deals and news publisher deals.

Image library. Photographs and illustrations with captions and metadata, priced per image or per library. Public prices run $0.27 to $5.27 per image, and a curated image has fetched 100x a bulk stock image. Shutterstock's early licences ran $25 million to $50 million each. Deeper: sell photos for AI training.

Synthetic data. Data generated by a model or a simulator instead of recorded from people or businesses. It has no holder to pay, so its price is a cost to produce, such as $0.0000024 per token for generated text. Labs use it to fill gaps, and still buy real data to check it against. Deeper: market rates.

Display licence. A licence to show content in a product's answers, with credit and links, rather than to train on it. Many disclosed deals are of this kind: 40 of the deals on record license content to be shown, not to train. A holder should know which it is signing, because the two are priced and limited differently. Deeper: disclosed deals.

3. Units, prices and the sides of the market

A price in this market means nothing without its unit and its side. Our market rates page states each price in a comparable unit and sorts it by side; we have converted 911 public prices to a comparable unit, and 59 markets have enough prices for a median.

Unit. What one price buys: an hour of audio, an image, a token, a task, a repository, a whole company's records. Media priced per minute is converted to the hour so prices compare. Two prices in different units cannot be compared until one is converted, and some cannot be converted at all. Deeper: AI training data prices.

Per hour of audio or video. The standard unit for speech, sound and footage, counted in hours of usable, finished recording, not hours of work. On the licence side, speech medians $76 per hour of audio an hour and video $42 per hour of video, while buyers offer holders of video $95 per hour of video. Deeper: sell audio and speech data and sell videos to AI companies.

Per image. The unit for photographs and illustrations. Licence prices median $0.029 per image an image, far below asking prices of $1.71 per image, because large licences buy in bulk. Deeper: sell photos for AI training.

Per token. A token is the piece of text a model reads at a time, roughly part of a word, and the natural unit for text and code at scale. Buyers offer holders of code $0.00004 to $0.0001 per token a token. Deeper: sell your codebase to AI companies.

Per book or title. The unit for published works. Licences have paid $2,750 per book or title a title, and buyers' public offers to authors and publishers run $355 per book or title. Deeper: news publisher deals.

Per repository. The unit for source code held as projects: one repository with its history. Licences closed at a median of $1,579 per repository, and buyers offer $1,478 per repository. Some buyers price a single pull request instead, at $1,000 per pull request. Deeper: sell your codebase to AI companies.

Per task and per environment. The units for environments, evaluations and expert work: a task is one problem with its grader or its answer, an environment a whole simulated workplace with its tasks. The sides differ widely: an engineer is paid $130 per task to write a coding task, while a lab pays a vendor $2,837 per task for a task made to order. Made to order, an environment median is $42.7K per environment, and our valuation prices a finished one at $20,000 to $300,000 an environment. Deeper: RL environments.

Per company. The unit for company records: everything one business holds, sold as one lot. Sales that happened median $28.3K per company a company, and buyers' public offers median $316K per company, which is why an offer figure alone overstates what closes. Deeper: how much company data is worth.

Per contract year. The unit for large, ongoing licences, where a buyer pays a fixed sum for each year of access. Text licences median $16M per contract year a year. Deeper: AI training data market size.

Pay (side). What a vendor or a platform pays the people who make data: a speaker, a filmer, an engineer, an annotator. For speech, $24 per hour of audio an hour. It is the bottom of the chain and not what a holder of existing data should expect. Deeper: market rates.

Offer (side). What a buyer publicly offers a holder for data that already exists, on its own pages or in its program. For video, $95 per hour of video an hour. An offer is a starting point and often the top of a range. Deeper: market rates.

Licence (side). What a sale actually closed at: a disclosed licence, a winning bid or a settlement. For speech, $76 per hour of audio an hour. It is the side that counts most in a valuation, because money changed hands. Deeper: disclosed deals.

Listing (side). What a catalogue or a data exchange asks for a dataset off the shelf. For speech, $118 per hour of audio an hour. An asking price is not a sale. Deeper: data on offer.

Collection (side). What a buyer pays a vendor to collect or make data to its own specification, also called made to order. For speech, $122 per hour of audio an hour. It sets a ceiling for existing data of the same kind, because a buyer can always have it made instead. Deeper: market rates.

Cost (side). What it costs to produce data that has no holder, such as synthetic data, or the vendor's own cost of production. For speech, $244 to $333 per hour of audio an hour. Deeper: market rates.

Median, range and spread. The median is the middle price of three or more comparable prices, and the range the span from low to high; we publish a median only where at least three prices compare, so one anecdote cannot set the number. The spread is how much more buyers pay than the makers of data are paid, for the same kind and unit: the vendor's margin plus its work. For speech, buyers pay 5.1x what speakers receive; for tasks, 9.5x. Deeper: figures and how we source every number.

Fixed fee and revenue share. The two basic ways a licence pays. A fixed fee is a single price, or a schedule of payments, after which nothing more is owed: micro1's terms make its licence royalty-free and paid in full once, and some buyers split the price into milestones, for example half on signature and half on delivery. A revenue share pays the holder a part of what the buyer earns each time it licenses the data onward, as Magma does with a revenue share, paid as your data licenses; the holder then depends on the buyer's later sales. Deeper: when AI data buyers pay and what is a data licensing agreement.

4. Licence terms

A licence is the agreement in which a holder grants a buyer rights in its data. Its terms decide much of the price, often more than the data itself. We have read 186 clauses from buyers' own terms and keep a register of the 15 terms a data licence turns on.

Licence. The agreement between the holder and a buyer granting rights in the data while the holder keeps ownership. AIData.Broker is not a party to it ( mandate clause 1.9). Deeper: what is a data licensing agreement.

Sale or assignment. A transfer of ownership of the data, or of the rights in it, to the buyer, after which the holder no longer controls it. A wind-down service may take assets outright: if SimpleClosure or an affiliate acquires them, it owns them. Deeper: what is a data licensing agreement.

Exclusive licence. A licence under which the holder may not license the same data to anyone else, for the field and period stated. It fetches 2.5 to 5 times a non-exclusive licence in our valuation. Some buyers take it by default: Kled's terms make the licence exclusive to Kled, and every licence on Polyshares' track record is a perpetual exclusive licence. Exclusivity can be narrowed to a window, limited in time or to one field, after which the licence becomes non-exclusive: Troveo says exclusivity is time-limited and has to be earned, and our register advises a defined field and a defined window. Deeper: exclusive vs non-exclusive data licence.

Non-exclusive licence. A licence that leaves the holder free to license the same data to others. It is the default in most buyers' terms, including micro1's nonexclusive, irrevocable, perpetual licence. Deeper: exclusive vs non-exclusive data licence.

Field of use (licensed purpose). The uses a licence permits, such as pretraining, fine-tuning, evaluation or retrieval, and those it excludes. Handshake AI says it is contractually restricted to AI model training and evaluation. A narrow field lets the holder license the rest elsewhere, and our register advises naming the permitted uses and excluding what is not granted. The narrowest common field is an evaluation-only licence, to test models but not train them, which fetches 0.2 to 0.5 times a full commercial licence in our valuation. Deeper: the licence terms register.

Term and perpetual licence. The term is how long the licence runs. Disclosed licences are short: Reddit's January 2024 arrangements ran two to three years. A perpetual licence has no end date, and most buyers' standard terms ask for one, often irrevocable as well, so the holder cannot take it back even if the buyer breaches. A perpetual licence is worth more than a term licence and should be priced closer to a sale. Deeper: the licence terms register and what is a data licensing agreement.

Model survival. Whether models trained on the data may be kept and used after the licence ends or the data is deleted. Buyers ask for it as a matter of course: deleting the data, under micro1's terms, does not require it to delete or retrain any model. Our register treats it as a priced term: perpetual model rights should cost more. Deeper: the licence terms register.

Sublicensing and resale. Sublicensing is the buyer's right to grant others rights in the data. Kled may sublicense through several tiers; Defined.ai's licence to its buyers is non-sublicensable and non-transferable. A licence with unlimited sublicensing is, in effect, a resale right. A resale restriction is the opposite promise, not to sell the data on as received: micro1 may not resell the data as a stand-alone data product, which still leaves it free to use the data in what it builds. Deeper: what is a data licensing agreement.

Prior-licence warranty. The holder's promise that it has not already licensed the same data on the same or similar terms to anyone else. micro1's terms require it in so many words: the holder has not previously granted a third party the same rights. A holder that has licensed before must disclose it or narrow the new grant. Deeper: buyer programs compared.

Warranties and indemnity. Warranties are the holder's promises about the data: title, authority to license, lawful collection, non-infringement and consent. Holders commonly give title and authority without qualification and qualify the rest to knowledge, tied to a disclosed rights inventory. An indemnity is a promise to compensate the other side for losses from defined claims, such as a third party's claim that the data infringed its rights; our register advises limiting it to precisely defined claim types, with a cap. Deeper: what is a data licensing agreement.

Acceptance window. The period after delivery in which the buyer may inspect the data and reject it. micro1 gives itself 30 business days after delivery to reject in writing, and payment often waits until the window closes. What counts as a proper delivery is set beside it: our register advises specifying format, schedule and acceptance, whether new data follows later, and avoiding service levels the holder cannot meet. Deeper: when AI data buyers pay and how to prepare data for sale.

Deemed acceptance. The rule that data counts as accepted if the buyer does not reject it within the window. In micro1's terms, if no notice is given in time, the data is deemed accepted. Without such a rule, a buyer's silence can delay payment indefinitely. Deeper: when AI data buyers pay.

Clawback. A buyer's right to take back money already paid. micro1 may demand a refund of everything paid if the data manifest proves materially inaccurate, even after acceptance; pre.dev may deduct a refunded buyer payment from later payouts; Kled requires a refund if consent is withdrawn. Our fee follows the money back ( mandate clause 8.6). Deeper: when AI data buyers pay.

Deletion and audit. Deletion is the buyer's duty to delete or return its copy of the data when the licence ends or on request, usually with written certification: Grepped deletes source data and working copies within 30 days of final delivery, confirmed in writing. Deletion of the copy rarely reaches trained models; see model survival. Audit is the holder's right to see records showing the buyer kept to the licence, for which our register suggests an annual certificate of compliance rather than open access to the buyer's systems. Deeper: the licence terms register.

Flow-down and output limits. A flow-down clause requires the buyer to pass the licence's limits on to anyone it shares the data with, so the limits travel with the data. In every licence we take part in, we propose that the buyer will not attempt re-identification or pass the data on without these terms. An output limit restricts what a trained model may reproduce, such as verbatim passages of confidential text or code, and our register recommends one where regurgitation would leak confidential material. Deeper: the de-identification standard and the licence terms register.

De-identification control. Who chooses the method and the provider that removes personal data, and who answers for the result. Buyers split: Mercor extracts and anonymises with its own pipeline, while under pre.dev's agreement the clean-up is the holder's. Our register favours the holder choosing or co-choosing the provider. Deeper: the de-identification standard.

First look. A buyer's right to see the data, or make an offer, before other buyers. It shortens a sale and weakens the holder's price. Our mandate rules it out: no buyer gets a first look. Deeper: how prices are negotiated.

Last look. A buyer's right to see the other offers and then improve its own after the deadline. It discourages other buyers from bidding their best. Our mandate rules it out alongside the first look ( clause 4.3). Competition moves price: in the Spirit Airlines sale the last offer was 2.5x the first. Deeper: how prices are negotiated.

Tail. A period after a broker's mandate ends in which a sale to a buyer the broker introduced still carries the broker's fee. In our mandate, the tail runs from the day the mandate ends and covers only buyers on the written list we send ( clause 10.3). Deeper: fees.

5. Rights and privacy

Whether data can be licensed at all depends on who owns it, who appears in it and what they agreed to. Buyers' counsel ask the same questions each time; we keep 27 of them, of which 16 block a sale until answered.

Personal data and sensitive personal information. Personal data is information about an identifiable person. US state laws call it personal information, the GDPR personal data; both reach far beyond names, to email addresses, identifiers and combinations of details that point to one person, and licensing it for money is, in California, a sale under the CCPA. Sensitive personal information is a narrower class with stricter rules: health, sex life, beliefs, precise location, government identifiers and, in California, the contents of messages unless the business was the intended recipient. Most buyers exclude it or require it removed. Deeper: data privacy laws by country and employee emails and chats in AI training.

De-identification. Removing or altering personal details so the data can no longer reasonably be linked to a person, and keeping it that way. Under the CCPA data counts as deidentified only with technical measures, a public commitment and contractual bans on re-identification. Names are the easy part. A quasi-identifier is a detail that identifies no one alone but can in combination (a postcode, a birth date, a job title in a small firm), so de-identification must coarsen these too; our engine cuts postcodes to their first part, and our standard requires every mention of a person to be replaced. Deeper: the de-identification standard and de-identify on your computer.

Safe Harbor. One of two routes to de-identify health data under HIPAA: remove eighteen listed kinds of identifier, and have no actual knowledge that what remains could identify someone. It is mechanical and easy to audit, but strips dates and geography a buyer may want. Deeper: HIPAA de-identification for AI training.

Expert Determination. The other HIPAA route: a qualified expert judges that the risk of identifying anyone is very small for the intended recipient, and documents why. It can keep more detail than Safe Harbor, and determinations often carry an expiry. Deeper: HIPAA de-identification for AI training.

Anonymous data. In EU and UK law, data that no longer relates to an identifiable person by any means reasonably likely to be used. Truly anonymous data falls outside the GDPR, which is a much higher bar than removing names. Deeper: GDPR and licensing data for AI training.

Pseudonymisation. Replacing names and identifiers with stand-ins while keeping a key that could reverse it. The EDPB holds that pseudonymised data stays personal data, though the EU's top court has since judged it per recipient. Our tool makes stand-ins from a secret key the holder keeps. Deeper: GDPR and licensing data for AI training.

All-party consent. A rule that every party to a private conversation must agree before it is recorded. California requires it for confidential calls under Penal Code 632, and several other states have similar rules. It matters for call recordings and meeting audio in a lot. Deeper: sell audio and speech data.

Data broker registry. A state register that every business meeting the state's definition of a data broker must join, usually each year. In US state law a data broker is a business that collects and sells personal information about people it has no direct relationship with, so a company selling data about its own customers or staff is generally not a data broker under California's definition, though its other privacy duties still apply. California requires registration with its privacy agency and asks about sales to generative AI developers; Vermont, Oregon and Texas keep their own registers ( Vermont, Oregon, Texas). Buyers and resellers that gather data about strangers are the ones these registers catch. Deeper: is it legal to sell customer data.

Bulk sensitive data rule. A US Justice Department rule that bars licensing bulk US sensitive personal data or government-related data to countries of concern and covered persons, and counts it whether or not de-identified. We do not take such data to market ( mandate clause 6.3). Deeper: selling from your country.

Court-approved sale in bankruptcy. A sale of a bankrupt company's assets outside the ordinary course of business, approved by the court after notice and a hearing; lawyers name it after the section of the Bankruptcy Code that governs it, a section 363 sale. The genetic testing company's assets were sold this way: the 23andMe sale closed as a section 363 sale. Where the privacy policy promised not to sell personal data, the sale must fit that policy or pass the review of a consumer privacy ombudsman, a disinterested person the court appoints under the Bankruptcy Code to report on it. In the Spirit Airlines case the ombudsman backed the sale of de-identified records once passenger systems were left out. Deeper: selling customer data in bankruptcy and closed and insolvent companies.

ABC (assignment for the benefit of creditors). A state-law alternative to bankruptcy in which the company hands its assets to an assignee who sells them and pays creditors, faster and more privately than a court case. Florida defines it as a state-law liquidation run by an assignee, and the assignee sells the assets, customer data included. Deeper: selling customer data in bankruptcy.

Fair use. The US doctrine that permits some unlicensed use of copyrighted work, judged case by case on four factors. Courts have held training on lawfully acquired books to be fair use while sending pirated copies to trial. A working licensing market, the Copyright Office says, weighs against unlicensed training. Deeper: fair use and AI training and lawsuits and rulings.

Work made for hire. Work an employee creates in the course of the job, which the employer owns. Code written by a contractor is not automatically the company's: under US law it needs a signed agreement in a listed category, or an assignment. This decides whether a company can license its own code base. Deeper: sell your codebase to AI companies.

Provenance and training data summaries. Provenance is the documented history of a dataset: where each part came from, the rights it was collected under and what was done to it. Buyers ask for it because laws now require model developers to publish a training data summary: California's law requires developers to post where their datasets came from and whether they were licensed, and the EU AI Act requires a sufficiently detailed summary. A licensed holder may therefore be named, in broad terms, unless the licence limits it. Deeper: training data provenance.

Rights reservation (opt-out). A machine-readable statement that content may not be used for training, such as a robots.txt rule against AI crawlers. EU law requires model providers to respect such reservations. In our latest census, 260 of 2,141 sites that answered shut at least one AI crawler out. Deeper: the AI crawler census and is it legal to scrape data for AI.

6. How a sale runs

These are the words of a sale run by a sell-side broker, defined as our agreements define them. Our mandate is published in full at /legal/agreements/mandate, and the steps are on how it works.

Valuation and counts. A valuation is our estimate of a holder's data, made at /value from its answers and its counts, with a reference; counts are the figures it measures from the holder's exports on the holder's own computer (how many people, messages, documents, years and repositories), which are released to us while the records themselves are not. For company records the valuation draws on 109 records of what whole companies' records sold for, were offered or were estimated at, weighing a sale, an offer and a reseller's figure at 1, 0.5 and 0.25. It binds nobody. Deeper: value your data.

Lot. The data a holder offers, as one valuation describes it, sold under one mandate. It is our word for what buyers' pages may call a dataset. Our data on offer index tracks the lots buyers publish. Deeper: how to prepare data for sale.

Mandate. The agreement appointing a broker as the holder's agent for one lot. Ours appoints us to market the data, run a competitive process and help negotiate, and the holder alone decides whether to sell, to whom and on what terms ( clause 5.1). Deeper: how it works.

Exclusive period. The time from acceptance in which a broker is the holder's only agent for the lot. During ours, the holder appoints no other agent and lists the lot on no data exchange ( clause 3.1); afterwards the mandate continues on a non-exclusive basis ( clause 3.4). Deeper: fees.

Rights file. The holder's answers on who owns each part of the lot, which contracts limit it, who appears in it and whether they agreed ( clause 1.14). A typical question is whether customer content is covered by customer contracts. Deeper: the Rights check tool.

Datasheet. The one document describing the lot: its source, contents, measured size, rights, personal data, format, the licence offered and its known limits, approved by the holder before any buyer sees it ( clause 1.3). Deeper: the datasheet builder.

Blind profile. A first description of a lot to buyers that identifies nothing about the holder: the kind of data, its size and years in ranges, the industry and region in broad terms ( NDA clause 4.1). Labs choose from such descriptions; one lab buyer picked lots like a software startup's email, drive and code archive. Deeper: how it works.

Round and buyer undertaking. The round is how we take a lot to buyers: the blind profile to every buyer whose published rules fit and that we can screen, less any the holder left out; the holder's name and the datasheet to each buyer that gives its buyer undertaking, a written promise of confidentiality ( clause 4.2); then one deadline for offers ( clause 4.1). Competition among buyers moved the offers for Spirit Airlines' records by $7.5M from first to last. Deeper: how prices are negotiated and how it works.

Screening. Checking a buyer's name against sanctions lists, and its answers about who controls it, before it learns the holder's identity ( clause 6.1). Where the buyer is itself a reseller, we ask it to name its onward recipients ( clause 6.4). Deeper: how we verify holders.

Offer. A buyer's written proposal for the lot: price, rights, exclusivity, payment and who removes personal data. We pass every offer to the holder in writing with what it means ( clause 4.4), and the holder may reject any, including the highest ( clause 5.2). Deeper: the offer check and an offer in your inbox.

Pre-existing buyer. A buyer the holder lists when it accepts the mandate because it was already talking to it. While the holder deals with that buyer itself, no fee is due on that sale ( clause 9.1). Deeper: fees.

Introduced buyer. A buyer to which we disclosed the holder's identity, sent the datasheet or held a discussion about the lot during the mandate. A buyer that received only the blind profile is not one ( clause 1.8). Deeper: fees.

Introducer and referral program. An introducer is a person or firm that brings a holder to us and is paid a share of our fee, out of our own fee, on that holder's sales within a set period ( clause 7.5). A referral program is a buyer's own published offer to pay whoever introduces a company whose data it then licenses: of the buyers of company records on our board, 11 publish what they pay, and 6 of the ten buyers our selling guide compares do. When we run a sale we decline such payments and tell the holder what the buyer offered ( clause 7.2). Deeper: introducers.

Fee and gross proceeds. The fee is what the holder pays its broker. Ours is paid only out of money a buyer has actually paid the holder ( clause 8.4), with no retainer and no minimum ( clause 8.9), at rates that fall as the sale grows, from 15% through 10% to 6%. It is calculated on gross proceeds: everything a buyer pays the holder under a licence (upfront payments, instalments, milestones, royalties, renewals), before the holder's own costs and taxes ( clause 1.7). Deeper: fees.

Questions people ask

What is AI training data? Any data used to build or test an AI model: the text, code, images, audio and video a model first learns from, the examples and judgments that tune it, the tasks it practises on and the sets it is measured against. Buyers pay for each stage differently, from fractions of a cent per token to $174K to $693K per company for one company's records.

What is the difference between licensing and selling data? After a licence the holder still owns the data and, unless the licence is exclusive, can license it again. A sale or assignment moves ownership; under US copyright law an exclusive licence also transfers the rights it covers, but a non-exclusive one does not.

Why do prices for the same kind of data differ so much? Mostly because they sit on different sides of the market. For an hour of speech, the speaker is paid $24 per hour of audio, licences have closed at a median of $76 per hour of audio and a vendor charges $122 per hour of audio to record it to order. Compare prices only within one side and one unit, as market rates does.

What does exclusive mean in a data licence? The holder may not license the same data to anyone else, for the field and period the licence states. Exclusivity fetches 2.5 to 5 times a non-exclusive licence in our valuation, and should be limited to a field and a window.

What is RLHF data? Human judgments of which model answer is better, used to train the reward model that steers a large model. A crowdsourced ranking costs $0.50 to $2; one by a domain expert costs several times more.

What is an RL environment? A simulated workplace with tasks and an automatic grader, where a model learns by trial and error. Public prices for one run $1,200 to $2,500 per environment. Our RL environments guide covers who buys them.

Is de-identified data still personal data? It depends on the law and the method. Data meeting the CCPA's standard falls outside its personal information rules; under the GDPR only truly anonymous data does, and pseudonymised data usually stays personal. The DOJ's bulk data rule counts de-identified data regardless.

What is a sell-side broker for data? An agent the holder appoints to find buyers, run a competitive sale and negotiate, paid by the holder and acting for it alone. AIData.Broker works this way: the fee is paid only out of money a buyer has paid, at rates from 15% down to 6%.

What this means for you

When you meet a number, ask which side of the market it sits on and in what unit, before comparing it with anything: an offer on a buyer's page is not what licences close at, and a worker's pay is not what existing data fetches. When you meet a licence, read the grant before the price: whether it is a licence or a sale, exclusive or not, for which field, for how long, and whether models survive it. Then read the clauses that decide when money arrives and whether it can come back: the acceptance window, deemed acceptance and clawback. Our licence terms register and the 186 buyer clauses behind it show how widely buyers' terms differ.

Before any buyer sees anything, settle the rights: who owns each part of the data, who appears in it, what they agreed to and how personal details will be removed. The rights check, the de-identification standard and our guides to HIPAA and GDPR cover the ground, and the valuation gives a first estimate from counts alone, with no data uploaded.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. We put the lot before every buyer that fits at the same time, with no first look and no last look, and only the holder pays us, a fee fixed in the mandate and due only out of money a buyer has paid. The full process is on how it works, and every guide sits in the guide index.

This glossary is research from public sources, not legal advice.