What Is AI Training Data, and Why Do AI Companies Pay for It?

Last updated: October 27, 2026

AI training data is the text, images, audio, code and records an AI model learns from. A model reads millions of examples, spots the patterns in them, and uses those patterns to answer questions, write and make decisions.

AI companies now pay for private business data because the free public internet is close to used up, and it doesn’t show how real work gets done inside a company. That’s why buyers like Mercor and micro1 pay businesses $20,000 to over $2 million for their records.

If you’re thinking of selling your own records, start with our guide to how to sell your company’s data to AI companies.

Key takeaways

AI training data is any set of examples a model learns from. Most of it so far has come from the public web.
Epoch AI estimates models will use up the stock of public human-written text between 2026 and 2032.
AI companies now get data four other ways: licensing deals, paid experts, company records and synthetic data.
Business records are valuable because they show finished work with known outcomes, which public data rarely does.
OpenAI and Anthropic say they don’t train on business or API customers’ data by default.

What is AI training data, in plain terms?

AI training data is the collection of examples an AI model studies to learn how to do something. For a chatbot that’s mostly written text: books, websites, emails, manuals and conversations. The model doesn’t store a copy of each page. It learns the patterns, such as which words and ideas tend to follow which.

Think of it as the model’s reading list. A model trained on legal contracts gets better at contracts, and one trained on support tickets gets better at answering customers.

The quality of the reading list matters as much as its size. Clean, accurate, specialized examples teach a model more than a larger pile of random web pages.

Where do AI companies get training data today?

AI companies get training data from five places: the public web, licensing deals with publishers and platforms, paid experts who create examples, records licensed from businesses, and synthetic data generated by other models. The public web was the main source until about 2024. The other four are now growing fast.

Source	What it is	Example
Public web	Pages collected by web crawlers	Common Crawl, a free archive of the web
Licensing deals	Paying publishers and platforms for their archives	Google’s reported $60 million a year for Reddit data
Paid experts	Hiring lawyers, doctors and engineers to write and grade examples	Contractors on Mercor and micro1
Company records	Licensing a business’s own files, tickets and workflows	micro1’s $1 billion commitment to company data
Synthetic data	Examples generated by other AI models	Rewritten or simulated records

News publishers were early sellers. News Corp’s deal with OpenAI could be worth $250 million over five years, according to The Wall Street Journal. Our list of companies that buy business data for AI covers who buys from businesses.

Why do AI companies want private business data when the internet is free?

AI companies want private business data because public text is running out and doesn’t show how work gets done. Epoch AI estimates the usable stock of public human text at about 300 trillion tokens, and expects models to use it up between 2026 and 2032.

The bigger reason is what the public web is missing. It has plenty of articles about accounting, but very few real workpapers with the final answer attached.

Mercor puts it plainly on its company data page: AI models are being trained to work inside companies, and “public data and synthetic datasets cannot replicate that.” Labs want the emails, tickets, SOPs and project files that show each step and how it turned out.

That’s also why older records sell. A closed legal matter or a finished construction project has a known outcome, which makes it a better teaching example than work still in progress.

What’s the difference between training data, fine-tuning data and evaluation data?

Training data teaches a model general skills from a huge mix of examples. Fine-tuning data is a smaller, focused set that adapts a trained model to one job. Evaluation data is held back and used only to test how well the model performs. Business records can be used for all three.

Type	What it does	Typical size	Business example
Training (pre-training) data	Teaches broad language and reasoning	Huge, much of the public web	Rarely a single company’s files
Fine-tuning data	Specializes a model for a task	Much smaller and focused	Support tickets with the correct resolution
Evaluation data	Tests the model without teaching it	Small and held back	Real accounting tasks graded by experts

Most company data deals feed fine-tuning and evaluation, not the first round of training. micro1, for example, says it turns company data into training environments where AI agents practice real business workflows.

What do “human data” and “expert data” mean?

Human data is training data that people create or grade on purpose, rather than data scraped from the web. Expert data is human data made by specialists, such as lawyers, doctors or engineers, who write example answers, rank a model’s responses and check its work in their field.

This is the business Mercor and micro1 started in. Both hire contractors through recorded AI interviews, and micro1 advertises rates from about $20 to $200 an hour depending on the role.

Company data is the newer cousin of expert data. Instead of paying an accountant to write 500 example workpapers, a buyer licenses an accounting firm’s real ones.

What are examples of AI training data by industry?

Every industry has records AI companies want. Law firms have closed matter files, accountants have workpapers, medical practices have de-identified physician notes, and manufacturers have run records that tie a process to a measured result. Mercor’s listings describe what it buys in 24 industries.

Industry	Training data AI buyers ask for
Law firm	Complete files from closed matters and completed deals
Accounting and tax firm	Tax returns, workpapers and bookkeeping records
Investment firm	Closed-deal records, credit files and portfolio reporting
Payments and fintech	Years of transaction, order, inventory or shipment tables
Cybersecurity	Security alerts, investigations and case write-ups
Semiconductor and materials	Run records that tie each recipe to a measured result
Medical practice	De-identified patient records with years of physician notes
Any office-based business	SOPs, support tickets, CRM records, email and Slack

Descriptions from Mercor’s listings, checked October 27, 2026. For what each industry pays, see how much company data is worth to AI.

Does an AI model memorize the documents it’s trained on? Could it repeat my files word for word?

Sometimes, yes. Models mostly learn patterns, but they can memorize and repeat some training text word for word. A 2023 study by researchers from Google DeepMind and several universities extracted gigabytes of training data from open models and from ChatGPT.

Text that appears many times, or is unusual, is the most likely to be memorized. A single company’s records make up a tiny share of a model’s data, and nobody has published a test of whether one firm’s licensed files can be pulled back out.

The safe assumption is that anything you license could resurface. Leave out anything you couldn’t stand to see quoted. Our guide to whether selling company data to AI is safe covers how to cut that risk.

Will synthetic data replace the need for real business data?

Not on its own. Synthetic data, made by AI models, is cheap and useful for filling gaps, but models trained mostly on other models’ output get worse. A July 2024 study in Nature found that training on recursively generated data causes “irreversible defects.”

The study called this model collapse: rare cases drop out first, and after a few generations the output drifts toward repetitive nonsense.

Synthetic data works best when it’s built from real examples. micro1 says some company datasets get synthetic rewrites that keep the workflow but remove direct ties to the original records. Real data still has to come first.

Do ChatGPT and Claude train on what businesses type into them?

Not by default on business plans. OpenAI says it doesn’t use business data from ChatGPT Business, Enterprise, Edu or its API for training unless you opt in. Anthropic says it won’t use inputs or outputs from its commercial products, including the API, unless you choose to share them.

Plan	Used for training by default?	Source, checked
ChatGPT Business, Enterprise, Edu and API	No, unless you opt in	OpenAI, updated January 8, 2026
Claude for Work, Team, Enterprise and API	No, unless you share data or send feedback	Anthropic, updated August 18, 2026
Claude Free, Pro and Max	Only if you turn on Model Improvement, or a chat is flagged for safety review	Anthropic, updated March 16, 2026

Feedback is the catch on both. Clicking thumbs up or down can send that conversation for training, and Anthropic keeps feedback for up to five years. Check your admin settings, and turn off feedback for the workspace if your staff handle client data.

Using an AI tool isn’t the same as selling data to an AI company. On a business plan, your prompts stay yours. A data license is a deliberate deal where you’re paid for records.

Common mistakes about AI training data

The most common mistake is assuming a business’s data has no value because the internet is free. These are the others:

Thinking AI models store a searchable copy of everything they read. They mostly learn patterns, though some text can be repeated.
Assuming synthetic data has made real data unnecessary.
Confusing your staff using ChatGPT at work with your company selling its data.
Thinking only big companies’ data is worth buying. Mercor’s listings start at five employees.
Assuming old records are worthless. Buyers often prefer finished work with known outcomes.

Frequently asked questions

What is AI training data?

AI training data is the collection of examples, such as text, images, audio, code and records, that an AI model learns from. The model learns patterns from the examples rather than storing copies of them.

Where does AI get its training data?

From the public web, licensing deals with publishers and platforms, paid experts who create examples, records licensed from businesses, and synthetic data generated by other models.

Why do AI companies pay for business data?

Public text is running out, and it rarely shows how real work gets done. Business records like tickets, workpapers and project files show each step and its outcome.

Is AI running out of training data?

Public text is. Epoch AI estimates models will use up the stock of public human-written text, about 300 trillion tokens, between 2026 and 2032.

Does ChatGPT train on business data?

Not by default. OpenAI says it doesn’t train on data from ChatGPT Business, Enterprise, Edu or its API unless the customer opts in.

What is the difference between training data and evaluation data?

Training data teaches a model. Evaluation data is held back and used only to test how well the model performs.

What Is AI Training Data, and Why Do AI Companies Pay for It? (2026)