In the mid-1990s I worked the Unix support desks at the University of Virginia, answering a phone hotline. Most calls were about commercial Unix systems. Then Linux questions started coming in, from ordinary users calling a support line about an operating system that not long before had been a hobbyist curiosity.1 I installed Linux early myself.
Later I built systems for TheStreet.com. We started on a commercial Netscape web server and moved to Apache, the open-source HTTP server.2 The move felt like freedom, in a practical sense: we could build, adapt, and run the system without living inside one vendor’s decisions about what we were allowed to do with it. I’ve thought about that feeling a lot while writing this book, because I think it’s what people are reaching for when they call a model open.
Over the years I contributed mostly to Java projects, including Ant and Maven.3 That’s where I learned the difference between using open source and taking part in a project’s community. I’ve also shipped open-source libraries and then left them alone for twelve or thirteen years, which is a good way to find out what a license does once the author stops paying attention.
Writing for O’Reilly and working as a part-time journalist for several years gave me access to many of the people in this book. It gave me a close view of where they disagree. That experience is why I’m asking the question on the cover: But is it open?
My bias
I’ve spent roughly thirty years working with and around Apache- and MIT-style licenses. These are permissive licenses: they grant broad reuse rights with relatively few conditions, including requirements to preserve notices.4 That shapes what I notice and what I distrust.
When I hear the word open, I check things in roughly this order. Is the source code available? Can I actually get to it? Can I modify it? Can I redistribute it without being required to open up the unrelated code around it? Can I use it commercially? That ranking is mine, and I don’t expect everyone to share it.
A copyleft advocate would rank those differently, and for good reasons. Copyleft uses license conditions to preserve recipients’ freedoms when covered software is passed on. The ordinary GPL allows private modifications and is generally triggered by distributing copies, so a cloud provider can modify a project, run it as a hosted service, and give very little back to the people who built it. The AGPL addresses that gap: users interacting with a modified version over a network must be offered its corresponding source.5 Neither license requires submitting patches to the original maintainers.
Companies weigh those obligations against compliance costs, adoption, and investment. Some use an open-core model, where the core is open source and some modules or services stay proprietary to pay for the work. Those are choices about how to fund development and where to place obligations. They deserve examination on their terms.
The same preference for practical control shows up with models. I’m comfortable with a developer withholding training data it cannot legally share, and I can see why a company holds back valuable data too. That leaves a real limit: without sufficient information about the data and training process, users cannot fully audit or reconstruct what they were handed. The distinction between publishing a corpus and documenting it will matter throughout this book.
The argument
Open source has a meaningful boundary. A license that forbids a field of use does not become open source because its publisher says so. The Open Source Initiative maintains the software definition and has published a separate Open Source AI Definition. The latter requires parameters, code, and detailed data information; it does not require every training datum to be shareable.6
Practical control has several dimensions, though. A release can let you run a model independently while withholding the means to reconstruct its training. It can publish substantial research while restricting commercial use. Those differences matter even when the release does not meet an open-source definition. The book’s argument is that we should name the freedoms and limits instead of asking one word to carry all of them.
I will compare what companies, standards bodies, researchers, and users mean by open with what was released, what rights accompany it, and what people can do afterward. That includes a question I don’t see asked often enough. If the training data isn’t shared, what should be shared about the later training that taught the model which answers to prefer: the reinforcement learning, the human and AI feedback, the reward models, and the evaluations that shaped its behavior?
I’m not a lawyer. Where legal questions remain unsettled, I’ll distinguish the text of a license, the claims people make about it, and what a court has decided.
Who this is for
If you’ve been in open source for years, I hope you’ll recognize the argument you’ve been having with your colleagues. If you’re new, this book will give you the history and vocabulary to join it. You don’t need twenty years of experience to ask what a release actually permits.
The organizations in these pages do different work. The Open Source Initiative and the Free Software Foundation defend definitions and freedoms. Foundations steward projects. Standards bodies coordinate shared specifications. Researchers publish evidence that other people can examine. Their disagreements are part of the subject, and the reader need not choose a camp before looking at them.
Chapter 1 sets out five questions to put to any release, starting with what happens when you press a download button.
Notes
Linus Torvalds, “What would you like to see most in minix?,” comp.os.minix, August 26, 1991, https://groups.google.com/g/comp.os.minix/c/dlNtH7RRrGA/m/SwRavCzVE7gJ (announcing “a (free) operating system (just a hobby, won’t be big and professional like gnu)”). ↩
Apache Software Foundation, “About the Apache HTTP Server Project,” https://httpd.apache.org/ABOUT_APACHE.html (members of the Apache Group formed the ASF in 1999). ↩
Apache Software Foundation, “Apache Ant,” https://ant.apache.org/; Apache Software Foundation, “Introduction,” Apache Maven, https://maven.apache.org/what-is-maven.html. ↩
Apache Software Foundation, “Apache License, Version 2.0,” January 2004, https://www.apache.org/licenses/LICENSE-2.0; Open Source Initiative, “The MIT License,” https://opensource.org/license/mit. ↩
Free Software Foundation, “Why the Affero GPL,” GNU Project, https://www.gnu.org/licenses/why-affero-gpl.html; Free Software Foundation, “GNU Affero General Public License, Version 3,” November 19, 2007, section 13, “Remote Network Interaction,” https://www.gnu.org/licenses/agpl-3.0.html. ↩
Open Source Initiative, “The Open Source AI Definition, 1.0,” October 2024, https://opensource.org/ai/open-source-ai-definition; Open Source Initiative, “Open Weights: not quite what you’ve been told,” https://opensource.org/ai/open-weights. ↩
From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.