Don’t touch that dial!
“Once is happenstance. Twice is coincidence. Three times is enemy action.” — Ian Fleming, Goldfinger
“Follow the money.” — William Goldman, All the President’s Men
Last September, I had my right hip replaced. In February, my surgeon swapped out the left one. I am a proud cyborg.
Hip replacements have come a long way since 1990, when my mom had her first replacement (she’s had four hip replacements in total — I call her “the centaur”). She spent a week in a hospital bed and two more weeks as an in-patient in a rehab hospital. I walked the same day as my surgeries (with a walker) and was downhill skiing within three months of the first replacement.
I’m four months into my second replacement and I’ve recovered nearly all of my strength (I’m swimming a mile nearly every day at our community pool), but my flexibility is still terrible.
I’ve found a YouTube yoga instructor with a good collection of “gentle 20 minute yoga” videos and on alternating days, I dig the mat and yoga blocks out from under the bed and try to gently but persistently work my robotic joints into something a little more human.
I stopped doing yoga some years ago, the arthritis made it too painful. There was a time when I could just listen to a yoga instructor as they named the pose and find my way into it, but not these days. It’s often the case that I have to grab my phone off the floor next to the mat and squint at it to figure out what move I’m supposed to be making.
When that happens, there’s about a one-in-three chance that I will accidentally tap a part of the screen that instantaneously swaps the video I’m watching now for a recommended video. What’s more, there’s no graceful way to go back to that previous video. A couple times, I’ve been completely unable to find it.
This isn’t an isolated incident, nor, I think, is it a coincidence.
During the lockdown, I binge-watched the entirety of MASH — all eleven seasons, plus AfterMASH and the pilot for RADA*R — on HBO Max. Years ago, I watched all those episodes in syndication, many times over, so often that I can picture the action if I can hear the dialog. Often, I’d just put my phone in my pocket and listen to the episodes while I puttered around the house or went for a walk, treating it like a podcast.
It was great. Pure comfort watching/listening. Except… While you watch a video, the entire HBO Max screen is taken over by dozens of thumbnails leading to recommended videos. When the video ends, even more thumbnails appear onscreen.
If any part of your hand so much as grazes your screen while you’re using HBO Max, you’ll be whisked off to another video, and not the next episode of the show you’re watching! As with YouTube, the video will be some recommended program, and, as with YouTube, getting back to your chosen video requires far more effort than simply going with the flow and watching the recommendation.
This isn’t just a couple of the streaming apps — it’s common to all the over-the-top streamers, as well as social media apps like TikTok and Reddit and Twitter (Reddit and Twitter’s web design have virtually no screen real-estate that you can just click on to switch your browser’s focus to that part of the window, say, so you can use your spacebar to page down).
What’s going on here?
I can’t say for certain, but here’s my theory: the companies behind these apps are obsessed with how “sticky” they are — how long your use-sessions are. Their product managers, designers and programmers are all assessed based on whether they can tweak their services to increase your “engagement time.”
These firms take their gospel from Steve Jobs: “People don’t know what they want until you show it to them.” They invest heavily in recommendation systems, hoping that they can interest you in a new thing before you get tired of the current thing and go somewhere else.
In other words, all this tech is designed by people who get paid more when you follow a recommendation, and paid still more when you stick with it. It’s their “Key Performance Indicator” — the dread KPI.
It’s not that they’re (necessarily) trying to trick you into clicking something that you’re not interested in. Rather, they’re using all kinds of automatically generated UI and layout experiments to determine the “optimal placement” for a recommendation to “maximize clickthrough.”
Some of the most delightfully perverse tales of machine learning systems going awry relate to “reward hacking,” when a machine learning model figures out how to achieve the goal its designers set for it that satisfies the formal requirement without actually achieving the goal itself.
For example, you might create an autonomous Roomba and try to get it to stop scuffing your baseboards and furniture by training it to avoid collisions with its front bumper. But if you do that, you might find — as one hacker did — that your Roomba starts driving backwards, crashing into everything in the house, but never generating a single front-bumper collision.
When the CEO decides that the shareholders want to see longer session times on the app, and tasks the product manager with making session times longer, who then makes programmers’ and designers’ performance evaluations dependent on the number of times you click or tap on a recommendation, it’s inevitable that your screen will become a minefield.
This is the Fatfinger Economy, where a design choice that increases accidental taps can send millions of dollars cascading through the system.
The Fatfinger Economy has many accelerants, but none so profound as the bezel-free, wrapround mobile screen. Now, the edges of your phone are tappable surfaces, plastered with hot-zones that teleport you from the thing you chose to watch to something an algorithm wants you to watch, even if you pick up your phone gingerly by its edges, like it was a film photograph you didn’t want to get fingerprint smudges on.
It’s a one-way journey. Getting to the recommended video is instantaneous; getting back to the video you accidentally switched away from is so cumbersome that it’s impossible to believe it’s not deliberate.
You treasure what you measure. Or, more formally, “When a measure becomes a target, it ceases to be a good measure.”
Recommendation systems are not, in and of themselves, bad. There’s a big old infosphere out there and you’ll never find your way to all the parts of it that would please you best on your own. Serendipity is a glory, and automating serendipity is a noble cause.
But “did someone click on a recommendation?” is a poor proxy for “was this a good recommendation?” It’s too easy to dress up a bad recommendation as a good one, or to otherwise trick someone into a click.
Even “did someone click on the recommendation and stick with it?” is easily gamed: all you need to do is take away the “back” button.
The Fatfinger Economy isn’t an explicit conspiracy — and more’s the pity, because then we could find the people behind it and make them cut it out. Rather, it’s an emergent property of perverse incentives.
Like a Roomba playing demolition-derby with your furniture legs as it runs full tilt in reverse, the Fatfinger Economy is a systemwide case of reward-hacking.
I solved my problem by downloading all my favorite 20-minute yoga sessions using youtube-dl, and now I watch them on my phone with VLC. I freely stipulate that this might not be a solution for everyone, but I am, after all, a cyborg.