BIG thread on "Piercing the Algorithmic Veil." Let's Go.

First on what the Algorithmic Veil is.

It comes from the name of a legal decision hold a company's shareholders or directors responsible for a company's crimes. https://twitter.com/hyonschu/status/1266114928080912384

When a corporation commits a misdeed, its shareholders and employees are generally not responsible for the misdeed. The corporation as a "person," is.

When the corporate veil is pierced, its shareholders and directors may now be held responsible for the misdeed.

The algorithmic veil, then, is holding the shareholders and directors responsible for the "misdeeds" of its algorithms. This is a huge shift.

When Big Tech companies commit misdeeds, however construed, they have always had the power to blame the algorithm or a bug in the software. https://www.engadget.com/youtube-bug-censoring-comments-against-china-191547339.html

This could persist as long as the companies wanted it to persist, because they could always hide behind the algorithms to say that "it's just math" – to imply that there was no hidden bias and that its actions were fair.

But what people are noticing is that this isn't true. There appears to be a huge bias in these algorithms.

That's because these algorithms are /not/ just math. They involve dozens, if not hundreds, of people to identify what is classified as OK or NOT OK.

This presents a problem: at what point are we allowed to say something is human intervention and what is algorithmic output? It didn't matter, because there was no one being held responsible. We defaulted it to being "the algorithm's fault."

When frequently called out for their actions, Big Tech would simply ignore these calls, or more often than not, not give anything beyond a general response.

And sometimes, they would just respond as "it was flagged by the algorithm" and take days, or even months, to "fix" it.

This, of course, leaves people frustrated. Without any responsibility for what the algorithm does, these companies were free to hide behind the algorithms, even if the actions were completely done by human hands.

In any case, the algorithm, a non-human entity, took all of the culpability. Not the engineers who created them, the data "scientists" who patched them together, nor the managers that authorized them.

As long as there was some algorithm in the pipeline, the problem could always be blamed on it.

No one would question the algorithm because they were proprietary and need not be exposed to the public for scrutiny. But even if they did, sometimes the algorithms are so complex that the people who write them don't know what they do. This is called a "black box" algorithm.

Here is the problem. We have no idea how much of these decisions are black box, how many of them are simple Yes/No algorithms, key word searches, or just a human in another country just pressing OK/NOT OK buttons on a terminal.

Here is how I believe the algorithm to /generally/ be working. They take a small subset of the content from the entire database (sampling). Then they get humans to say OK/NOT OK and flag them and let the algorithms find ones that are similar in content and label as OK/NOT OK.

This is called /supervised classification/: you have a bunch of racist content, and if any other content happens to have some of those words or combinations of words (shared features), then the content that has never been seen by a human is labeled racist – EVEN IF IT ISN'T.

There are a few problems here. First, the people who initially classify the content have biases. They might see content by The someone not actually racist to be "racist" or "incorrect" because of their own biases. This can get built into the algorithm.

Secondly, they have a vested interest in erring on the side of caution. So more likely than not, borderline content can be seen as "dangerous" for the safety and friendliness of the site (for advertisers, not you.)

Third, because of how impossibly large the entire database is, it is physically impossible to test the algorithm to see if it is behaving right on the whole database. There will be a lot of missed cases that are classified incorrectly. A LOT.

People DO test the algorithms to see if it is working as it "should." This is called model validation. People take a subset of the output and manually review to see if it is working "properly." But again, it is entirely left up to human bias.

I've identified at least three points that are subject to human bias. But it gets worse.

Because it is a firehose of content, the heuristics a person might be used to flag something might be lazy, or might overreach, or just be flat-out biased. And because it runs through an algorithm, it gets multiplied by a factor into the millions, if not billions.

This, is fundamentally how Big Tech filters content. Any small bias that enters into the algorithm manually gets multiplied huge. That's the power of leveraging such algorithms.

Now, what seems to be happening is this:

Big Tech has gotten away with this for a very long time because they said they were a "communications platform." They did NOT mess with what was presented on the platform, and so they claimed no liability for what was there.

Now that people have seen and realized that this is not the case, things are being editorialized with what appears to be explicit intention, the platforms are now in jeopardy of now being unable to present themselves as not biased. They are no longer a communications platform!

Note: It does not matter in what direction they are biased.

Because now they have shown that they are willing to edit, moderate, and censor content, they are about to lose the immunity.

So now, Big Tech is at a weird crossroads. Do they take a completely hands-off approach, or do they continue to "fact-check"?

If they go completely hands-off, a lot of bad stuff can go down on the platforms.

If they continue to editorialize or censor, then they have to show that they are NOT stifling free speech.

Problem: They do not understand fully what the algos are doing.

If they want to continue to use algorithms, now someone has to be responsible to ensure the algorithms are "behaving" or that there was no censoring of free speech, otherwise the company can be in legal trouble.

In other words, Big Tech might no longer be able to hide behind the algorithms.

Now that companies are on the hook for it, they can no longer point fingers at the algorithms, they now must take a strong look at their processes and ensure that these bad classifications do not fall through the cracks or lose their immunity as communications platforms.

From the EO:

(5b) Complaints described in section 4(b) of this order will be shared with the working group, consistent with applicable law. The working group shall also collect publicly available information regarding the following:

(ii) algorithms to suppress content or users based on indications of political alignment or viewpoint;

Now the algorithms will be under intense scrutiny from all sides.

Do you think the heads of these organizations will hold their people responsible for the outputs of these algorithms now?

Piercing the Algorithmic Veil means holding people accountable for the actions of whatever algorithms they provide for public consumption. It means no longer placing the blame on "just math" or "software bugs" that seemingly appear to no more than just coincidence.

The details of how this will work out remain to be seen, but it's going to get very interesting on how this works out, not just in social media, but in other arenas of so-called "artificial intelligence" – search engines, "intelligence engines" and autonomous vehicles.

(i.e., who gets blamed if an AV goes off-road and slams into an outdoor farmer's market? Is it the algorithm's fault? The engineer? The engineering manager? The CEO?)

In any case, people will have to claim responsibility again. And that is always a public good.