Horrible Remedies for Broken Recommenders
This is not a happy story.
About a month ago, my then-pregnant partner and I learned that our baby’s heart had stopped beating. Our child, Minne, was to be stillborn.
It’s been an emotional rollercoaster, and I would kindly ask everyone not to bring it up if they meet with me in real life. There is, however, one anecdote in this story that deserves to be said out loud.
I want to talk about how freaking hard it is to get the internet to turn off ads for baby products after having a stillborn child or a miscarriage. I’m happy to report that I may have found a remedy to the problem, but it’s truly a sad state of affairs.
We did a lot of “how to be a good parent” searches during our pregnancy. It is, therefore, no surprise that online stores figured out that we may be in the market for baby products and that ads started popping up. That said, after what we’ve been through, the last thing we want is to be reminded of baby’s.
There are many culprits, but I’d like to zoom in on the ads from bol.com. You may not have heard of the website if you’re not Dutch, but it’s the “Amazon” of the Netherlands. I’m picking this store, not because they are the only culprit, but because their ads went beyond their own website. The day after it happened my partner and I were desperately looking for distraction on YouTube. We were eager to watch a new episode of a playlist we enjoy only to get we’d get ads for bol.com diapers in the middle of a video that was supposed to calm us down.
I was initially a bit surprised that I even got this ads in the first place. Although Google usually brings me to the book description on bol.com I rarely end up buying my book there in favor of the local book store. Either way, from glancing at my front page, it was pretty clear what bol.com thought of me.
I figured I’d reach out to their customer service. If these ads went away, that’d be one less risk of a trigger on our end. Eventually, I got a hold of two people, one via their website and one via their social media channel. Unfortunately, they told me that there was nothing that they could do. They told me that I could turn off their emails and notification, but they could not change the front page of their website or change the ads I received from them on YouTube.
I should add that the folks I spoke to were understanding and kind, and they wholeheartedly agreed that it was a sorry state of affairs. But in the end, they genuinely could do nothing.
There are many “this is broken” kinds of thoughts that I experienced while this was happening. Fun fact: I have trained many of the data scientists at bol.com a few years ago. So I knew a few people who still worked there. For a moment, I contemplated reaching out to some of them on LinkedIn to discuss this. But how much impact would that make? If I were to reach out to a senior data scientist at Bol.com, would they even be able to go against their team’s quarterly OKRs to fix this? How much work would it even be? Would it require a complete overhaul? Would I need to do this for every website out there?
For a few years now, I’ve been trying to understand under what circumstances machine learning systems fail in real life. This anecdote is a perfect example of a common theme; it’s rarely the machine learning technique that’s at fault whenever a system fails. Instead, it’s unintended side-effects combined with an inability to correct a system.
I’d love, love, to give bol.com more information on what I want. But this information should be interpreted as constraints, not mere click-data. It’s not just about baby products either. To mention another example: I don’t own a driver’s license, so I’d love to be able to explain to my browser that car commercials are pointless.
Part of me even thinks it’s a more straightforward “data science”-problem when you ask the user what they want instead of playing the guessing game with tensors. Instead of worrying about an optimal recommendation, maybe there’s something to be said to allow users to customize instead. But why would bol.com ever consider this? Allowing for customization might be something that users are interested in, but it may cause them to customize for products with lower margins.
This last bit thought got me thinking. Maybe, just maybe, there was a simple but incredibly awkward remedy to the problem.
Let’s make an assumption by pretending that bol.com only optimizes towards product margin. I can’t imagine this is actually true, but for arguments sake, let’s say it is. If it really is true it makes my situation very simple. If I want the baby ads to go away, I need to find a product or category with a higher margin and convince bol.com that I am interested in that.
I tried out a bunch of products, but the best solution seems to be … Chromebooks.
I spent a few minutes searching for Chromebooks on bol.com. I clicked many links and opened many tabs in an attempt to steer the recommender. After a while the front page no longer had the baby ads, and I even confirmed that YouTube had stopped displaying the diaper ads.
While I’m happy to report this trick seems to work, the fact that this had to be the solution leaves a bad feeling in my stomach. I reported back to customer service and requested they start advising this to people in a similar situation. The folks on the other end responded positively. My impression was that they genuinely didn’t consider this was an option, and I hope they will spread the word.
I’ve zoomed in on bol.com here, but I should stress that there are many offenders in this space. It’s effortless to sign up for a free “baby box” from Albert Heijn, but it’s disappointingly tough to get them to stop sending you pictures of healthy babies that remind you of a stillborn child. Not just via email, mind you, I’m talking physical letter mail here.
Most providers of baby boxes will gladly remove you from their system when you call them. You can tell that they deal with this request frequently and that there is a process in place. But in my experience they will also tell you there’s a 4-week waiting period for the changes to cause an effect. To make things worse, when you sign up for a baby box, your data typically gets shared with 3rd parties too. My phone call may have removed me from one database, but nobody dared to guarantee anything about what 3rd parties might do.
But here’s the thing: it’s not like my situation is rare. In 2020, 735 children (0.4% of all births) were stillborn in the Netherlands. Another 538 children died within the first 28 days after delivery. About 1/4 pregnancies in the Netherlands don’t make it past the 12th week due to a miscarriage.
This effects many thousands of people. And if any of these people want to stop seeing ads that reminds them of a traumatic experience … the best thing we can do is ask them to click on Chromebooks?!
What about websites like Zalando? Should my wife be required to search the service for high-margin alternatives to maternity clothing in order not to be reminded of what happened? What about instagram? You’re able to click a button that reads “this ad is not relevant for me”. That’s not a bad feature to have but it’s incredibly painful to have to do repeatedly. The button may provide a signal, but it’s by no means an off-switch. It took weeks before the pregnancy ads went away.
Whenever I walk in the park and see a newborn in a stroller, I’m confronted with a traumatic memory. I couldn’t be happier for the parents, but my heart does sink for a moment every time. The park is something that I am going to have to learn to live with. But am I really to accept that it’s required for me to get an email that reads “everything you need for your baby” with pictures after everything that happened? Just imagine you’d get such an email while you’re at work in the office just before an important meeting. Is that really something we don’t want to be able to turn off immediately, given how many people it affects?
I hope my “hack the recommender by clicking Chromebooks”-trick might serve people in a similar situation but I hope we can all agree it’s a horrible remedy. The bigger hope that I have is that this piece helps folks recognize that it’s past time we start thinking about building an “off switch” for these systems.
Without it, many of our emails, letters, recommenders, searchbars and frontpages remain beyond broken because they are unfixable for the end-user.
I also propose we consider this “off-switch” not just for pregnancy related products but also for other products that could cause trauma. A former alcoholic won’t enjoy that beer ad and somebody who is dealing with an eating disorder may not want to see slim-fit diet-ads either.
It deserves highlighting that everything that I’m describing here is a well-documented problem. I’m certainly not the first to talk about it. For more information, I might recommend checking:
For attribution, please cite this work as
Warmerdam (2021, Oct. 21). koaning.io: Beyond Broken. Retrieved from https://koaning.io/posts/beyond-broken/
BibTeX citation
@misc{warmerdam2021beyond, author = {Warmerdam, Vincent}, title = {koaning.io: Beyond Broken}, url = {https://koaning.io/posts/beyond-broken/}, year = {2021} }