I installed Red Hat Linux in June 1998 and became an open source believer overnight. What I did not know at the time was that “open source”, as the name of a movement, was barely four months old. The words themselves were older, but the people who had just chosen them as a banner had spent the previous fifteen years arguing about a different word, “free”, and that argument never really ended. It just moved. Today it is being replayed, almost line for line, over AI models, and you cannot follow the AI version without knowing the original.
This is chapter 1 of the Open Source AI Hub. It is the theory the rest of the hub is built on: where the ideas came from, what the licenses actually demand, and why the split between “free software” and “open source” still decides how we should judge a model that calls itself open.
Before the name: the GNU project and the four freedoms
Free software is software that respects four freedoms of its users: to run it for any purpose, to study and change it, to redistribute copies, and to distribute modified versions. Richard Stallman announced the GNU project on 27 September 1983 with the aim of building a complete Unix-compatible operating system that anyone could use, share and improve, and he founded the Free Software Foundation in October 1985 to fund the work and defend the idea.
The four freedoms were written down and numbered in the 1990s as the Free Software Definition, and the numbering starts at zero: freedom 0 is the freedom to run the program as you wish, freedom 1 to study how it works and change it, freedom 2 to redistribute copies, freedom 3 to distribute your modified versions. The famous clarification came with them: free as in free speech, not as in free beer. Price was never the point. Control was.
One detail matters more than the rest for what follows. Freedom 1 requires access to the source code, because you cannot study or change what you cannot read. Source availability is therefore a means to an end, not the end itself. Keep that in mind when we get to AI models, where “the source” is not a single thing.
Copyleft and the GPL
Copyleft is the practice of using copyright law to guarantee freedom rather than restrict it: a copyleft license lets you copy, change and redistribute a work on the condition that every redistributed version, modified or not, carries the same freedoms. Stallman’s instrument for this was the GNU General Public License. Version 1 arrived in February 1989, version 2 in June 1991, and version 3 on 29 June 2007, adding explicit patent terms and provisions against hardware that refuses to run modified software.
The GPL’s proof of scale was Linux. Linus Torvalds announced his kernel in August 1991 under his own terms and moved it to the GPLv2 in 1992, and it has stayed there since. The kernel is the clearest demonstration that copyleft does not stop commercial adoption; it shapes it. Companies contribute because they must share, and because everyone else must too.
Two relatives are worth one sentence each. The Lesser GPL relaxes copyleft for libraries so that proprietary programs can link to them. The Affero GPL, published in November 2007, extends copyleft to software offered over a network, closing the gap that lets a company run modified GPL code as a service without ever distributing it. That gap, as we will see, is one of the reasons several AI tools chose the AGPL.
Permissive licenses: MIT, BSD and Apache-2.0
A permissive license grants the four freedoms without requiring that downstream versions remain free. The MIT and BSD licenses, both from the late 1980s, ask for little more than that the copyright notice travel with the code. You may take permissive code into a closed product, and many do.
The Apache License 2.0, published in January 2004, is the permissive license that matters most for AI. It adds two things MIT and BSD lack: an explicit grant of patent rights from every contributor, and a retaliation clause that ends those rights for anyone who sues over patents in the covered work. That patent clarity is why Apache-2.0 is the default choice for model weights released by companies with large patent portfolios, and why it appears on Qwen, Mistral’s open line, gpt-oss and OLMo. One compatibility note that still trips people up: the Free Software Foundation considers Apache-2.0 compatible with GPLv3 but not with GPLv2.
Debian draws the line: the DFSG (1997)
The Debian Free Software Guidelines are the first written test for whether a license counts as free, ratified in July 1997 alongside the Debian Social Contract and drafted mainly by Bruce Perens. A distribution that packages thousands of programs needed a checklist, not a philosophy, to decide what could go in. The guidelines demanded free redistribution, source code, permission for derived works, and, crucially, no discrimination against persons or groups and no discrimination against fields of endeavor. A license that said “not for commercial use” or “not for military use” failed the test, however noble the reason.
Hold on to that last clause. It is the one that most “open” AI licenses fail.
The Cathedral and the Bazaar, and the Netscape moment
Eric S. Raymond presented “The Cathedral and the Bazaar” at the Linux Kongress in May 1997, and O’Reilly published the book in October 1999. Its argument was about method, not morality: that a loosely coordinated crowd releasing early and often could out-develop a closed team of experts. Given enough eyeballs, all bugs are shallow.
The essay found its audience fast. On 22 January 1998 Netscape announced that it would release the source code of its Communicator browser, citing Raymond’s essay as an influence, and the code shipped on 31 March 1998 as the Mozilla project. A large company had decided, in public, that the bazaar was a business strategy. That decision came before the label “open source” had been adopted.
3 February 1998: the word “open source”
“Open source” was proposed as the name for the movement on 3 February 1998, at a strategy meeting in Palo Alto called in response to the Netscape announcement, by Christine Peterson of the Foresight Institute; she published her own account of it in 2018. The two words were not new. Caldera had used “open source code” in a September 1996 press release about opening up DR-DOS, and the phrase appears in scattered earlier writing. What was new was choosing it deliberately, as a label. The people in the room, Raymond among them, wanted a term that would sell the development model to businesses without the ambiguity of “free”, which in English means both liberty and zero price, and without the moral framing that made executives nervous.
Later that month Raymond and Perens founded the Open Source Initiative to steward the term, and Perens adapted the Debian guidelines into the Open Source Definition, with its ten criteria. In April 1998 an O’Reilly summit of free software leaders endorsed the new label. Within a year the phrase was everywhere, and I was one of the converts.
The split: ethics versus pragmatism
Free software and open source describe almost the same set of licenses for different reasons. The free software movement holds that users deserve the four freedoms as a matter of ethics, and that proprietary software is a social problem. The open source movement holds that openness produces better software, faster, and that this is reason enough. Same code, same licenses, different arguments.
The split was visible immediately. Perens left the OSI in February 1999 with a public note titled “It’s Time to Talk About Free Software Again”, worried that the new term was letting people forget the freedoms. Stallman made the case at length in 2007 in “Why Open Source Misses the Point of Free Software”, and has revised it many times since. The neutral compounds FOSS and FLOSS appeared in the early 2000s so that people could refer to the whole field without picking a side.
My own position, after living with open source since 1998, as a hobby first and a profession since 2000: I use “open source” for the method, because that is the word the industry understands, and I keep the four freedoms as the test, because they are the only part that protects you when a vendor changes its mind. The pragmatists won the market. The ethicists won the argument. Every license fight since, including the ones over AI, has been the two sides rediscovering that they need each other.
Why this history matters for AI
Three things from this history transfer directly to AI models, and the rest of this hub depends on them.
Fields of endeavor. The Open Source Definition’s sixth criterion forbids licenses that restrict use in a field of endeavor. A model license that prohibits certain uses, or that stops applying above a certain number of users, fails that criterion exactly as a “non-commercial” software license would have failed the Debian test in 1997. This is not a technicality; it is the whole reason the definitions exist. Chapter 3 reads those licenses clause by clause.
The preferred form for modification. Freedom 1 needs the source, and copyleft licenses define source as the preferred form for making changes. For a program that is a text file. For a model it is weights plus training code plus enough information about the data to reproduce the thing, and the argument over how much of that must be released is the argument over the Open Source AI Definition. Chapter 2 takes it up.
Marketing versus meaning. In 1998 the community chose a friendlier word to win over business, and accepted the risk that the word would drift from the freedoms. In the 2020s “open” has drifted further still: it is applied to models you may download but may not use freely, and to weights without the recipe that made them. The remedy is the same one Perens reached for in 1999. Say what the license actually grants, and judge the artifact by that, not by the adjective on the launch post.
Next steps
Continue to chapter 2, What Is Open Source AI?, which applies these ideas to models. For why any of this matters beyond licensing, read Who Cannot Use AI, and Why?. And if you want to see what I was saying about open source in 2006, that post is still here.