Should the government require that Google search, Facebook and YouTube recommendation algorithms be published? And should changes to those algorithms need to meet certain criteria at by regulators, in order to prevent ill effects on society?
I'm thinking there must come a time when our machine algorithms are so impactful in society that we can't afford to leave them in the stewardship of a single private company with only profit motives. I think we probably reached this point a while ago.
Many folks often wish that ranking/recommendation algorithms like PageRank search engine, or Youtube "watch next", or email spam heuristics, etc, be revealed to the public. E.g. the exact algorithm is required by law to be published on Github.
That would hurt society more than help because once the algorithms are public, the code's inner workings can then be gamed and abused.
That's the paradox or contradiction: for the algorithms to be effective, they have to be kept a secret to prevent gaming.
Recommendation/ranking algorithms are not to be confused with "security by obscurity" memes. Yes, you can publish the exact algorithm to derive a SHA-256 but that's not the same domain as a ranking algorithm. For Google to transition the algorithms from basic PageRank to Panda to Penguin to Hummingbird, etc ... the inner workings must be kept secret to be effective.
I know it drives people crazy that the algorithms are secret but you can't make them public in a way that stops abusers from using that knowledge against us.
What level of algorithmic transparency can one achieve without sacrificing gaming?
I ask this question because this is a common discussion for many aspects in our society. For example, we have designed laws over 100s of years despite the knowledge that they can be prone to gaming. The existence of gaming itself is not a good enough reason to have ZERO algorithmic transparency.
That's an important consideration, but I don't think it is known to be inevitable.
It seems like a worthy research goal, to develop algorithms which are not gameable (in defined ways) despite knowing the algorithms.
Consider how publishing good encryption algorithms does not render the encryption breakable.
And the research being done now on Differential Privacy techniques, to allow large scale data collection from individuals without compromising individual privacy.
I suspect there is an analogous class of algorithms that resist being gamed in ways that matter. Possibly requiring that the deployed algorithm depend on a secret key or random seed, but that would be fine as long as the secret does not encode another algorithm.
So I don't buy that argument off hand.
Plus people are already gaming the system, it's just the knowledge to do so is imperfect and unfairly distributed. Leveling the playing field there puts more focus on making the algorithm less vulnerable in the first place instead of effectively relying on security through obscurity.
Companies are the LEAST dependable ones to create truly uncorrupt algorithms for social media. They adjust everything to drive sales, to drive profit, to drive ad revenue, because year over year shareholders expect profit growth.
Creating social media presence used to be fair, but now as a content creator, the ONLY way to get noticed is to game the system. Rich people and companies prevail against algorithms because they can hire companies to boost their views, and to create followers, and to create massive campaigns.
This makes everything unfair, and invalidates the equality that the Internet was meant to create.
There is already a universal basis for social media timelines and algorithms, it's the one they started out with when they were trying to become important; A time line based on when content was posted, equal visibility for all users, conservative ad placement without pop-ups, User Interfaces based on functionality first, thorough testing of releases prior to updates, reading feedback from users before addressing stakeholder concerns... And most importantly "Not Becoming Evil" in the process.
This happens without it being public knowledge and is commonly referred to as SEO. Keeping the black box a block box hurts more than helps because the privileged few that do have it figured out completely control the narrative. In addition a publicly published algo helps re-introduce competition. I see it as no different than food products being required to list ingredients.
I'm not claiming that absolutely zero gaming is possible. For all the successes that the SEO industry achieves, it's still a version of the proverbial blind men (SEO experts) feeling different parts of a giant elephant (the opaque Google algorithm). That's why SEO practitioners still get blindsided by new search engine changes by Google.
But the algorithms are already being "gamed and abused" - that's why Google keeps changing them - so in the worst case revealing the algorithms is no different than our current situation.
The advantage of opening the algorithms is that everybody can see and understand how the algorithms might be gamed, so they can look out for sites and other actors gaming them.
Keeping the algorithms secret is a way for them to play stupid and shirk responsibility, and lets them do their own shady things without oversight.
As the game of cat and mouse goes on it would result in us refining our algorithms to the point where they actually measure what we want rather than an approximation of what we want. Eventually "gaming the algorithms" would be the same as producing high quality content. Yes refining algorithms to that point would be hard work but it would get figured out.
As far as I know, nobody in computer science has come up with a ranking algorithm that can't be gamed to render the algorithm useless. (Unless the so-called algorithm is a trivial case such as the runner with the fastest time across the finish line is the "winner" of the race.) For a multi-criteria algorithm that approximates a hazy idea like "video quality" that uses various signals as proxies for quality, the algorithm's exact source code must be kept a secret. Yes, you can publicize broad outlines of what the algorithm does but you can't reveal the actual code (or the actual machine learning weights, etc).
Yes, an algorithm like a SHA256 hash source code can be made transparently public and still can't be "abused" such that a bad actor can make the string "DEADBEEF" appear anywhere they want in the output hash. This class of algorithm behaves differently than ranking/recommendation algorithms.
https://static.googleusercontent.com/media/guidelines.raterh...
Disclaimer: this is just my understanding of how these newsfeed type algorithms work based on conversations with friends who work on these teams at FB/GOOG. Please do correct me if I got something wrong here.
There's not "an algorithm" that can be published and inspected, like good ol' Pagerank. It's probably more accurate to think of it like thousands of slightly different algorithms knotted together, tied to an enormous database that selects with algorithm gets used for which person at which point in time. The knot and the database are constantly changing.
You could perhaps publish the machine learning algorithms that created the recommendation engine, but those won't tell you anything useful without the training data (which, for these systems, is absolutely huge and complex).
I'm convinced that a big reason these social platforms haven't fixed their recommendation engines is that they actually have no idea why they recommend the things they do.
The systems are trained on "engagement", without regard to whether it's good or bad, happy or sad, constructive or destructive. It's just: which thing collects the most clicks from which people.
In a way, this makes these recommendation engines mirrors into human nature--which sounds cool, but is actually boring and banal. We already knew that it is easier to make someone unhappy than to make them happy. And we typically want people to build things that deliver happiness--even though we know it's harder.
We have a bunch of companies that run A/B tests all the time, and of course either A or B result in more time spent watching or less subscriber cancellations. So then they keep that change. Hence you get a product that over time is literally designed to be as addicting as possible...that can't be a net positive for society in my mind.
An interesting idea I have (not realistic of course) is that companies would have to "send their customers away." Like the mountain lion in that Trailer Park Boys episode where "if you love something you have to set it free," a positive way to service your customers is to help them get done with what they want to do as fast as possible.
"But my customers want to be entertained for X hours." Okay, provide a UI for them to find stuff to watch but do not autoplay videos (for example).
But using ML for controlling content? Please. I very much doubt anything like that is going to happen.