YouTube has me figured out pretty well. If you sampled my current recommendations, you’d get such videos as How to Make the Perfect Omelet (so simple, yet so hard), Bacon Fat Tortillas from Fresa’s Chicken al Carbon in Austin (I want to eat that, now), The Black Hole Bomb and Black Hole Civilizations (did you know that you could use a Black Hole to power entire civilizations?), and Engine failure after Takeoff - Briefing (hey, you never know when you might need that).

Granted, if you examined my viewing history, it wouldn’t be hard to figure out that I’m into cooking, planes, astronomy and astrophysics, and video games. But it’s unlikely that any human was ever directly involved in creating YouTube’s recommendations to me. Instead, an algorithm using machine learning figured out what to recommend from what I’ve watched in the past.

But this algorithm has another purpose: To keep me on YouTube, watching as many videos as possible. YouTube (and by extent, its parent company Google) wants to make money by showing me ads. The more videos I watch (aka. “engagement”), the more ads they can show me. So instead of being trying to find videos that I might want to watch next, it is tuned to recommend me videos that keep me watching. Which might not sound like a big difference, but is a big distinction in what you’ll see.

Machine Learning sounds strange and exotic, but in reality, it’s anything but. It is not a new field, and has been explored since the 80s. Only recently has it gained more recognition and its software ecosystem has now grown to the point where you can create a machine learning algorithm by only stringing together a few boxes, and providing learning examples.

Machine learning consists of two parts: A very, very big equation with lots of variables, and a feedback system that tunes those variables via training so that if you give it a certain input, it will produce a certain output. Depending on what you use to tune those variables, it can then do things like detect faces, figure out what your camera is looking at, or recommending videos on YouTube.

Here’s the thing: these equations don’t care about you, or anyone else. They do exactly what they were tuned to do, and what they were trained on. If you train this algorithm to recommend videos that keep me watching, that’s exactly what it will do.

The trouble starts when this collides with human behavior: We generally lose interest pretty quickly, so we tend to move on and do something else after a short while. However, we are also emotional beings, and respond strongly to stimuli that induce things like fear or outrage. So, to keep us engaged, you need to be constantly stimulated with something that creates that exact emotional response.

Put this together with the above algorithm, and the next thing you know, YouTube recommendations get more and more extreme. Because the algorithm only cares about keeping you watching, no matter the reason. Which is why I eventually get recommended videos featuring plane crashes and other accidents. If YouTube thinks you like watching political videos, you’ll get recommendations for more extreme viewpoints. It gets even worse when children’s videos are involved. People have figured out how to manipulate that algorithm, and show algorithmically created content to children that frightens, traumatises, or even abuses children. Because children don’t know any better, they can’t determine wether what they’re seeing is normal and skip it if it’s not. That’s how these producers get their cut of YouTube’s ad money.

Machine Learning certainly has a wide array of applications where it can be used for great good. There are trials running with machine learning algorithms assisting doctors in recognizing tumors, or recommending treatment options to patients. Self-driving cars are poised to reduce accidents by a huge factor. But like any other technology, it is not a cure-all, and it will still fall to us humans to actually ensure these algorithms are created and used ethically.

Other interesting links from around the web:

• The bitter truth about taste buds, genes and flavour — Ever wondered why we can taste bitter so well?
• How We Found 7 Earth-like Planets Around 1 Star (YouTube) — Ever wondered how we can know so much about a star system 40 light years away, even though we can’t actually see any of its planets?
• Creation and consumption — Ever wondered what all the computers around the world are actually used for?
• Why sports sound better in your living room — Ever wondered why sports games sound so much better when you’re watching from home?

📖 Weekly Longread 📚

“Cut off from their unit, the tiny band of American soldiers was outnumbered and outgunned in the deserts of Niger, fighting to stay alive under a barrage of gunfire from fighters loyal to the Islamic State.” — ‘An Endless War’: Why 4 U.S. Soldiers Died in a Remote African Desert

🦄 Unicorn Chaser 🦄

The Philharmonic Turntable Orchestra