Why upload filters don’t work (really simple math!)

“I can’t figure out how upload filters should work, but I’m not a technical person — surely someone who is can sort it out!”

That is a misconception. I’ll be happy to explain, requiring — I promise! — no technical understanding of what an upload filter is, or how it works.

The current draft of the EU Directive “on copyright and related rights in the Digital Single Market”, available here, (PDF in English), also known as the EU Copyright Directive, requires in its Article 17 (formerly Article 13), clause 4, that the service provider undergoes,

in accordance with high industry standards of professional diligence, best efforts to ensure the unavailability of specific works and other subject matter for which the rightholders have provided the service providers with the relevant and necessary information.

It is obvious that the only way any platform hosting user-generated content would thus have to intercept any such content on upload, failing which it would immediately become potentially liable for a copyright violation. This would require a technical facility commonly called an upload filter.

We don’t need to talk about how upload filters work

Now, an upload filter is immensely complex and there are tons of technical difficulties — the only time this has been attempted on a large scale is YouTube’s Content ID, and it is exceedingly unreliable and prone to overblocking. But for the purposes of this discussion, it doesn’t matter whether implementing an upload filter is difficult to do.¹

Assume for a moment someone has built a magnificent upload filter. Something that operates on magic pixie dust that catches all copyright violations.

Interactions

Now, let’s call every instance of someone uploading content to the internet an “interaction”. Every tweet, every Facebook post and comment, every comment on your favorite news site, every blog post you write, every picture that you take and post to a WhatsApp group of 50 people or more, every YouTube video and comment — let’s call all of those “interactions.”

And let’s make an outrageously overblown assumption: suppose that on the internet today, 1% of such interactions infringe someone’s copyright. Again, let me reiterate that this is ludicrously high. The vast majority of internet interactions today are either completely trivial and thus irrelevant to copyright, or works of your own, or a perfectly legal fair-use way of using someone else’s work, such as when you quote a passage of a book. But purely for the sake of this discussion, let’s say it’s 1%.

So then let’s look at 10,000 interactions that completely random people make on the internet.

Those would then break down like so:

	Total
Perfectly legal	9,900
Infringing copyright	100

Catching copyright violations. Or non-violations.

OK. Now, suppose we built a perfect upload filter, i.e. one that catches all copyright infringements. Remember, the Directive calls for “best effort to ensure the unavailability” (emphasis mine) of potentially infringing content. It does not allow providers to balance for freedom of expression or the like, so to err on the side of caution, they must strive to over- rather than underblock. So a perfect filter is one that has no false negatives — meaning if content infringes, it is always caught.

Now, suppose further that the filter mis-identifies content (meaning, flags content as infringing when it is not) with a rate of only 2%. That means it has 2% false positives. That, now, is ridiculously low for any automated screening procedure.²

So that means that out of our 10,000 interactions tracked by our “perfect” content filter, the numbers break down like this:

	Total	Flagged as legal	Flagged as infringing
Perfectly legal	9,900	9,802	198
Infringing copyright	100	0	100
Overall	10,000	9,802	298

Congratulations, a coin toss beats your upload filter.

That leads us to the question: if you upload something and it gets flagged, how likely is it that it is actually infringing any copyright? Answer: 100 in 298. Roughly one in three. Yes, that is worse than a coin toss. And remember, this is assuming an implausibly high rate of infringements overall, and a ludicrously low false-positive and false-negative rate on your filter.

Go ahead and play with the numbers, tweak the false-negatives and false-positives, whatever. As long as what you’re looking for is exceedingly rare, automated filters detect it with poor accuracy.

And if you leave all parameters the same, but consider a probably much more realistic infringement rate of 1 in 1000, that is, 0.1%, then things look like this:

	Total	Flagged as legal	Flagged as infringing
Perfectly legal	9,990	9,800	200
Infringing copyright	10	0	10
Overall	10,000	9,800	300

Now there’s a one-in-thirty chance that an upload block is legitimate. Assuming there is an appeals process, and all false positives get appealed, then that means the human going through the appeals will have to undo a block 29 times out of 30.

A cheap optimization

I’d like to propose an optimization here: any website seeking to implement a content filter should consider to just use a random number generator to reject your upload, comment, tweet, or post with a certain probability that is demonstrably larger than that of an upload filter. I’d posit that that would be by far the safest, cheapest way to comply with the directive — if it becomes law.

Of course, everyone who is now in favor of this directive (including its Article 17) will hate that.

---

Footnotes