§ Built on Stolen Data

§ 2021-07-02

A few days ago, GitHub released a new tool for developers called CoPilot. It's a software plugin for text editors as VSCode that allows developers to "Just Hit Enter" to autocomplete the code they've written so far. I recommend checking out CoPilot's website. Their showcases are pretty incredible.

In fact, they're so awe-inspiring that I had a few chats with people about it in recent days. "Man! We're the next to get replaced by AI!" and so on, where the initial reactions. I'm sympathizing with those emotions - but I was also fairly skeptical. I immediately had doubts about that bot helping me write helpful unit tests - which I find the most tedious and analytically-laborious task in writing software. I don't think it'll understand different testing contexts.

But some people on the Internet managed to get ahold of this program quickly and started exploring its outputs. Suddenly it became conceivable that the "public" data GitHub had used to train CoPilot may include restrictively licensed code like those residing under ,e.g., the GPL.

Someone pointed out on Twitter:

github copilot has, by their own admission, been trained on mountains of gpl code, so i'm unclear on how it's not a form of laundering open source code into commercial works. the handwave of "it usually doesn't reproduce exact chunks" is not very satisfying pic.twitter.com/IzqtK2kGGo

— eevee (@eevee) June 30, 2021

Shortly after, another person followed up by having CoPilot generate the infamous fast square root algorithm from Quake III (GPL licensed).

I don't want to say anything but that's not the right license Mr Copilot. pic.twitter.com/hs8JRVQ7xJ

— Armin Ronacher (@mitsuhiko) July 2, 2021

Indeed, the later tweet author even tricked the auto-completion algorithm into suggesting a license for the code as well.

Now, I think it shouldn't come as too big of a surprise that the inputs of a machine learning algorithm can reappear as fragments in output later. I guess that for ML-generated images, our brains are just not good enough to compute matches. Hence, we credit most contemporary algorithms as "pretty original". In contrast, it seems relatively easy for code-copy-cats to be debunked. Particularily, when this one suspicious-looking random-ass hex value "0x5f3759df" is auto-suggested.

For me, what I found much more scandalous is, however, this other dimension where it appears we've caught Microsoft stealing from all GitHub users. Could it be that it simply had dumped all sorts of restrictively licensed open source code (including MINE!) into their model? Could it be true that they're now profiting by stealing from my work?

That'd be outrageous! And so, as I was informing myself some more, this subtle anger - triggered by my sense of justice feeling hurt - started growing.

It only got worse when I read more Hacker News comments. Specifically, one which claimed that many image recognition algorithms were likely just trained with copyrighted image data too. I mean, I don't even know what to say about that.

Indeed, now when I'm writing these lines - it's so absurd - I feel like doubting myself. Is it really that bad? But see, for me, it didn't even occur until now that training a model on copyright-protected images is an option. My most prolonged understanding of how these models were built was: (1) Launch a successful social network. (2) Setup shitty T&Cs. (3) Harvest data from users legally (but maybe immorally).

However, to now learn that a giant like Microsoft dares to just release a product built on looting is genuinely a shock. Particularly when considering that I could be part of the wronged ones. Or; that I could be part of the damaging party by having accidentially used such an illicit algorithm .

Right now, it feels absurd. I'm classifiying the usage of copyrighted content in machine learning as deeply immoral and illicit. Involuntarily, an image of a blood diamond comes to mind.

I think it should be illegal and pursued to distribute such a piece of code.

Edit on July 8, 2021:

oh my gods. they literally have no shame about this.

GitHub Support just straight up confirmed in an email that yes, they used all public GitHub code, for Codex/Copilot regardless of license. pic.twitter.com/pFTqbvnTEK

— ✨ Nora Tindall 🪐 (@NoraDotCodes) July 7, 2021