The free AI software most businesses build on is a set of open-source libraries, and free means no license fee, not no obligations. PyTorch and scikit-learn use BSD-style licenses that require keeping copyright notices, and TensorFlow uses the Apache License 2.0, which adds an explicit patent license that ends if the user sues over patents in the software. Every one of them is offered as is, with no warranty.
Key takeaways
- BSD 3-Clause permits commercial use if copyright notices are kept in source and in the documentation that accompanies binaries, and if contributor names are not used to endorse a product without permission.
- The Apache License 2.0 grants a perpetual, worldwide, royalty-free copyright license and a patent license, and the patent license ends on the date the user files patent litigation over the software.
- PyTorch moved under the Linux Foundation as the PyTorch Foundation on September 12, 2022, with business decisions handled by a governing board and technical decisions left to maintainers.
- scikit-learn made its first public release on February 1, 2010 and lists a BSD license, with a 2007-2026 copyright line held by the scikit-learn developers.
- No license guarantees support, security fixes, or fitness for a purpose. Those remain the user's own job.
What "free" covers in an open-source license
An open-source license gives permissions without charging for them. The user may run, modify, and in most cases ship the software inside a commercial product. In return, the license sets conditions, and these differ between license families. Breaking a condition can end the permission, which is the usual legal risk when a business copies a library into a product without reading the file named LICENSE or COPYING.
The licenses below are the ones attached to widely used machine learning libraries. They are separate from the licenses on downloadable models, which can include use restrictions and are a different question.
BSD 3-Clause: PyTorch and scikit-learn
The 3-Clause BSD License, approved by the Open Source Initiative, has three conditions. Source redistributions keep the copyright notice and the list of conditions. Binary redistributions reproduce the same notice in documentation or other accompanying materials. The names of the copyright holder and contributors may not be used to endorse or promote derived products without prior written permission. The license file in the scikit-learn repository carries this text with the line "Copyright (c) 2007-2026 The scikit-learn developers." PyTorch's license file follows the same three-part pattern, with notices from Facebook, DeepMind Technologies, NYU, NEC Laboratories America, and other holders.
Both files end with a disclaimer that the software is provided "as is" by the copyright holders and contributors, without express or implied warranties, including merchantability and fitness for a particular purpose.
Apache License 2.0: TensorFlow
TensorFlow's repository uses the Apache License, Version 2.0. Section 2 grants each contributor's "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable" copyright license. Section 3 adds a patent license limited to claims that a contribution necessarily infringes, and it ends for any user who files patent litigation alleging that the work infringes a patent, on the date the suit is filed. Section 4 requires giving recipients a copy of the license, marking changed files with prominent notices, keeping existing notices, and passing along a NOTICE file if one exists.
Who maintains the code
PyTorch was announced as moving to the Linux Foundation on September 12, 2022. The announcement named a governing board with leaders from AMD, Amazon Web Services, Google Cloud, Meta, Microsoft Azure, and NVIDIA, and said technical decisions remain in control of individual maintainers. The project's site states it has more than 2,400 contributors, a figure from 2022.
scikit-learn began in 2007 as a Google Summer of Code project by David Cournapeau. Researchers at INRIA took over leadership in 2010 and made the first public release on February 1, 2010, and INRIA holds the copyright on work by people it employed at the time. Its site says the project is financially supported by Probabl and other institutions, and describes it as community-driven.
Comparison of the three licenses for a business user
| Library | License family | Must keep or pass along | Patent terms |
|---|---|---|---|
| PyTorch | BSD-style, three conditions | Copyright notices in source and in documentation with binaries | No express patent grant in the text |
| scikit-learn | BSD 3-Clause | Same notices; no endorsement use of names | No express patent grant in the text |
| TensorFlow | Apache License 2.0 | License copy, change notices, NOTICE file | Express patent license, ends on patent suit |
Checklist before shipping free AI software
- List every library and its license file, including libraries pulled in as dependencies.
- Confirm that notices will appear in the product documentation or an about screen.
- Check whether the business holds patents that could conflict with the Apache patent termination clause.
- Plan for security updates and support, since the licenses disclaim all warranties.
- Keep a dated record of license versions, because projects can change terms between releases.
This is general information, not legal advice. A lawyer reviews license terms for a specific product.
FAQ
Can a company sell a product built on a BSD-licensed library?
The BSD text permits redistribution in source and binary form, with the notice conditions above. It does not require the product itself to be open source.
Is free software always free to use for any purpose?
Not always. The three licenses above have no field-of-use limits in their text, but other AI software, especially downloadable models, can carry custom restrictions.
Who fixes bugs in free libraries?
Maintainers and contributors do, on their own schedule. The licenses give no service commitment, so a business that depends on a library often budgets staff time for upgrades.
Does a library license cover the data a business trains on?
No. The library license covers the code. Rights in training data are a separate legal question.
Related reading
- Costs and savings reported from AI in business
- AI in software development
- Methods that make AI software faster and leaner
More in AI Tools and Industry.
Last verified: October 2026
Sources
- BSD 3-Clause License, Open Source Initiative
- Apache License, Version 2.0, Apache Software Foundation, 2004
- LICENSE file, PyTorch repository, GitHub
- COPYING file, scikit-learn repository, GitHub, copyright line 2007-2026
- LICENSE file, TensorFlow repository, GitHub
- PyTorch strengthens its governance by joining the Linux Foundation, PyTorch blog, September 12, 2022
- About us, scikit-learn documentation