Skip to content
Outrich Data

Data type

License your private source code for AI training

A private codebase can often be licensed for AI training if your company owns the code outright and no client or open-source license restricts it. Outrich Data reviews ownership, strips secrets and personal data, and pays for a non-exclusive licensed copy. You keep the code, the IP and the right to keep selling your product.

What counts as source code

Source code here means the private repositories your engineers have built and maintained: application code, internal tools, infrastructure definitions, tests and build scripts. The surrounding history matters as much as the files. Commit messages, pull request discussions, code review comments and linked issues record why each change was made.

It includes code for retired products and internal systems that no longer earn revenue. It does not include code you wrote for a client who owns the result, third-party code you only have a license to use, or anything you hold in escrow or under a customer's confidentiality terms.

Why AI labs want it

Coding models have already been trained on most public repositories. What they have seen little of is private, production code: large systems with real business logic, long histories, legacy modules and the compromises that come with paying customers. AI labs seek that material for training and for evaluation sets that a model cannot have memorized.

History is the scarce part. A bug report, the failing test, the pull request that fixed it and the reviewer's comments form a complete worked example of software engineering. Labs use these chains to train and test models on tasks such as fixing a defect, reviewing a change or migrating a module.

Removed before licensing

What comes out first.

  • API keys, passwords, tokens, certificates and private keys, including those buried in commit history
  • Internal hostnames, IP addresses and connection strings
  • Customer names and customer data in fixtures, seed files, logs and comments
  • Author names and email addresses in commit metadata
  • Vendored third-party code and files under licenses that do not permit this use
  • Code written for a client under a contract that assigns ownership to the client
  • Details of unpatched security weaknesses in systems still in production

Data title search

What gets checked.

  • Chain of title. Confirm that employees and contractors who wrote the code assigned their rights to the company in writing. Contractor work without an assignment is a common gap.
  • Open-source obligations. Code that includes or derives from GPL, AGPL or other copyleft components carries license terms. An inventory of dependencies and copied snippets is needed, and the result should be reviewed with counsel.
  • Client work. Work-for-hire and custom development agreements often give the client ownership or an exclusive license. That code is excluded.
  • Third-party SDKs and commercial libraries. Their license terms may prohibit redistribution of their code or headers.
  • Customer agreements and escrow. Source code escrow terms, exclusivity clauses and confidentiality commitments to customers can restrict disclosure.
  • Export controls. Some encryption and dual-use code is subject to export rules. Check before licensing to a buyer in another country.

Who signsThe CTO or head of engineering sponsors it and the general counsel approves, because the license touches core IP. Board or investor consent may also be needed, depending on your financing documents.

Value

What moves the price of source code.

FactorEffect on value
History and review depthRepositories with years of commits, pull request reviews and linked issues are generally worth more than a snapshot of files.
Test coverageA working test suite lets a lab verify model output automatically, which makes the code far more useful for evaluation.
Language and domainLess common languages, legacy stacks and specialized domains such as billing, logistics or embedded systems are thinner in public data.
Never publicCode that has never been published or open-sourced is more useful for evaluation, since models have not seen it.
Clean ownershipDocumented IP assignments and a dependency inventory widen what can be included and shorten review.

Questions

Licensing source code: what companies ask.

Will licensing our code let a competitor copy our product?
A license covers defined uses such as training and evaluation, not resale of the code or use of it to run a competing product. You approve each buyer category and license scope in writing, and you can exclude repositories you consider too sensitive. Discuss the residual risk with counsel before deciding.
Our code uses open-source libraries. Is that a problem?
Using libraries as dependencies is normal and usually means those libraries are simply left out of the dataset. The harder cases are copied snippets and copyleft code mixed into your own files. A dependency and license scan identifies those, and affected files are excluded or reviewed.
How are secrets in old commits handled?
History is scanned, not only the current files. Keys, tokens and passwords found anywhere in the commit log are redacted before a package is built. Any credential that was ever committed should also be rotated on your side, which is good practice regardless.
Can we license code from a product we shut down?
Often yes, and it can be a good fit because commercial sensitivity is lower. The same ownership checks apply. If the product was built for a client, acquired from another company or covered by an escrow agreement, those terms need review first.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc