Off the Beat: Bruce Byfield's Blog

The conventional wisdom is that free software licenses are rapidly evolving. The copyleft licenses are supposed to be in decline, and the permissive licenses gaining popularity, according to  two widely-quoted studies from Red Monk by Stephen O'Grady  and Donnie Berkholz, In fact, writing in 2012, Berkholz declares that new project licenses are more likely to use a permissive license than anything else. However, on closer examination, whether these conclusions are accurate is open to question.

For one thing, both the Red Monk studies and their main source, Black Duck Software and its Open Hub site (formerly Ohloh) are business-oriented. Because permissive licenses are more flexible and less likely to generate compliance problems, the possibility is strong that these sources could have a conscious or unconscious bias against copyleft licenses.

Unfortunately, no details have been published about how the data is collected or organized, so judging the conclusions is next to impossible. However, Berkholz's study does show a series of choices that could affect results. For example, he divides licenses into permissive, limited, and copyleft licenses, while other studies group what he calls limited as weak examples of copyleft licenses. He then ignores the limited group altogether.

Berkholz also "hypothesized" - - for no obvious reason -- that programming languages might somehow be involved in the evolution of license selection, then build his study around that assumption. No doubt Berkholz had his reasons for these decisions, but different organization and assumptions might  produce different results.

Sources of bias
In fact, such problems are common in any studies of free licenses. Some project repository sites, such as Canonical's Launchpad, have no license search built into their interfaces, so they tend to be ignored. By contrast, SourceForge returns pages of results, and, since the results can number in the hundreds, only offer an approximation of individual results unless you are willing to do some tedious counting. Other potentially useful sources of data, such as the repository contents of major distributions, have no provision for searching by license. Debian, for example, notes that its license "include" a short list but makes no guarantee that the list is complete, and goes no further than to note that a half dozen licenses are "common."

Where detailed licensing information is available, comparisons can still be difficult because of how licenses are categorized. Few sites, for example, bother to make a distinction between different releases of a license aside from the second and third versions of the GNU General Public License (GPL). In particular, distinctions between versions of the Affero General Public License and of the Lesser GNU General Public License and of various BSD licenses are only made by some sources. These categorizations many effect results in particular for copyleft licenses with their long histories. As a result, analysts often cannot simply make comparisons. Instead, they have to study how their sources have organized data, and make their own categorizations to compare with any accuracy.

Yet what really complicates any analysis of license use is biases in the data that are not immediately obvious. For example, the selection of licenses seems to depend heavily on how long the repository has been in existence. At one extreme, SourceForge, one of the longest established project repositories, returns 95,175 results for GPLv2, but only 14,550 for GPLv3 -- totals that are roughly five times larger than all the major BSD licenses combined. At the opposite extreme, the Android site  specifically recommends the Apache License while disparaging the Lesser GPL and ignoring altogether strong copyleft licenses, so, although the site lacks any license search, it is not hard to guess what license most contributors choose. The picture changes drastically depending on which source you use -- nor can you assume, in the absence of any definitive source, that you can simply lump all sources together and they will all balance out. The numbers are simply not consistent, no matter how you massage them.

The questions that matter
At any rate, the question of which types of licenses are gaining or losing popularity is perhaps misframed. If core projects for the Linux operating system continue to use copyleft licenses, it changes very little if dozens of small or specialized projects favor permissive licenses. 

Yet analyses of license use continue to act as if each piece of software is an equally important data point. A better indication of trends might be which licenses are chosen by core pieces of software. Even more interestingly, what license are new replacements for core software using? (Systemd, for what it's worth, uses the Limited GNU General Public License, while Wayland uses the Historical Permission Notice and Disclaimer).

Nor, in the end, may the relative status of copyleft and permissive licenses be the most important point to analyze. As Steven J. Vaughan-Nichols pointed out in the article that inspired this one, 73-85% of the projects on Github have no license whatsoever. This number is so large that even allowing for large margins of error and the fact that it is based on a single source of information, it is likely a cause for concern.

Still, the relative popularity of difference licenses remains a reoccurring interest. The lack of reliable data is not likely to stop speculation, any more than the same problem is likely to stop discussion of the relative merits of different distributions. Just remember that, unless an analysis is more thorough than any that have been done so far, the results are like cartoon characters, hanging in mid-air after walking off a cliff for a split second before plummeting downwards, thoroughly discredited once you start to examine them.