Why it’s not just about being the product

Surely one of the most mentioned quotes in the wake of the Cambridge Analytica/Facebook story is “If you are not paying for it, you’re not the customer; you’re the product being sold.” While it’s been said about everything from TVs to automotive black boxes, in Facebook’s case it applies to the extensive personal data it collects from and about their users and how it analyzes and packages it in precise demographic groups so that advertisers can pitch users anything from shoes to candidates. 

Facebook trying to show me that the world is better with it in it.

Users have long been accustomed to getting services for free on the Internet, without really understanding their true cost. Google’s entire business model is to build products that delight users and are extremely useful while using data collected from users using those services to serve ads that are “more relevant” to them. 

Yet what started as simple contextual advertising based on what users searched for and basic demographic targeting has evolved into Facebook having 29,000 data points for every user, tracking users wherever they go on the web and on their phone, both virtually and geographically, and aggregating those points into very specific groups such as video game players who are likely to spend on in-app purchases or even users interested in health issues such as “diagnosis with HIV or AIDS… erectile dysfunction, and binge-eating disorder awareness.” Cambridge Analytica made its decisions on who to target for the 2016 elections based by building a behavioral model based on the analysis of seemingly unrelated data points: “researchers had figured out how to tie your interest in Kim Kardashian West to certain personality traits, such as how extroverted you are (very), how conscientious (more than most) and how open-minded (only somewhat). And when your fondness for Ms. Kardashian West is combined with other interests you’ve indicated on Facebook, researchers believe their algorithms can predict the nuances of your political views with better accuracy than your loved ones.”

That detailed and actionable information is a boon for advertisers and gets results. Which is why when asked at yesterday’s hearing “are you willing to change your business model in the interest of protecting individual privacy?” Mr Zuckerberg declined to answer.

It’s in some ways discouraging to think, as a user, just how pervasive personal data collection is, even when the product itself isn’t free. To use a Mac, a PC, or a mobile phone, products far from free, users are still required to agree to data collection. 23andMe sell genetic testing kits and sells that data to universities and pharmaceutical companies for research, Sonos sells high-end speakers and also collects listening and activity data, and Fitbit sells fitness trackers and also collects exercise data.  Tesla, and other car manufacturers, sell expensive cars and also collect driving data, including video. So the entire “if you’re not paying for it” qualifier is irrelevant. In the age of big data, the user is always the product, even when they pay.  

It’s wrong to blame users for this and insist that they entered this agreement freely and willingly and gave their consent when they started using the product. It’s not that clear cut.

  • There is almost no flexibility for users, it’s either accept all the terms and conditions or not use the product. Though some companies, including Facebook, give users some privacy and data options, users cannot opt out of most data collection.
  • Users don’t understand the data being collected, its breadth and its depth. It’s not their fault, it’s hidden in 30-40 page legal documents for most sites, through Google and Facebook are trying to use a friendlier language. It’s also intentionally vague. It “may include” some things and often provides a simplistic example of data categories with “such as..” which are generally the more benign applications.
  • It’s not really a choice for many people. For Facebook, there are professional, academic, NGOs, and community groups that use the platform is their only means of communication, where members have no choice but to sign up. For other products, such as fitness trackers – there isn’t an alternative that doesn’t collect and make use of data. (Maybe that can become a selling point in the future: be the product that doesn’t collect data, the user owns it all.) At some point users have bought the product and want to start using it. They may not have an option to return it at that point if they don’t like the terms and conditions.

With all that, is it even possible to say that users have given their informed consent? I think the answer is clearly no, which is why, understandably, Americans think that they have lost control of their personal data. It’s not that they don’t care, it’s that they don’t feel they can change it. The anger right now around Facebook seems like a major turning point in terms of awareness, but it also seems that change won’t come from within. Personal data is just too valuable.

Advertisements

Cooking with voice: misadventures in the kitchen with Google Home

Two weeks ago I attended an event at Yummly, a recipe app recently bought by Whirlpool that’s trying to make a smarter kitchen. There is a big push to make cooking easier while having appliances do more by tying their behaviour, whether it be finding ingredients in the fridge, identifying them on a supermarket shelf, or controlling the oven settings. While that future is a bit farther away, there is a future that’s already here: cooking via voice command. Today, only 18% of smart speaker owners are currently using it for cooking requests, but that’s bound to increase as kitchens become smarter, and doing this well will have real value in a smart kitchen.

Voice UI in the kitchen has so much potential to be better than an open recipe book that needs to be propped up and constantly referred to, each time while scanning the page for the right location to find the information, or a tablet that needs to be turned on with a password or touch and scrolled through to find the right information. That said, voice cannot just recite a recipe meant to be read. There need to be a few changes.

Intrigued, I wanted to see what the current state was. I decided to make Chai tea. (I’m paraphrasing the dialog I had with Home.)

Me: OK, Google, how do you make Chai tea?

Home: I found a recipe for Homemade Chai on Epicurious, would you like to make it?

Then Home said that there were 9 ingredients, that it would list them and pause between each one. It then proceeded to tell me ingredients like “2-inch piece fresh ginger, cut into thin rounds” Is that peeled? Not? I decided to peel it anyway. The second item was a 2 cinnamon sticks. Conveniently here I could ask Home if I could replace sticks with powder. Home read me an answer, with a ratio (I’ll get back to this later as it didn’t help me at all) but I ended up finding sticks in my cupboard. At this point I realized that I’d only be making half the recipe amount because I didn’t have enough sticks.

As it continued adding spices, I decided to just dump them all in a saucepan. Still listing ingredients, Home told me I needed water, tea bags, milk, and sugar in addition to the five spices. I just added all in the pot.

Home: Would you like to hear the instructions?

Me: Yes.

Home: Add all the spices to a saucepan and crush them slightly.

Me: [Too bad, they were all in the water/milk mixture but I tried crushing them a bit.] Next!

Home: Bring to a boil.

Me: [Brought to boil] Next!

Home: Lower heat and simmer for 10 minutes, would you like me to set a timer?

Me: Yes!

Home: Timer set for 10 minutes.

Home: [10 minutes pass.] Alarm!

Me: OK, next!

Home: Remove from heat. [Continues alarm.]

Me: OK Google, stop the timer! Next!

Home: Add tea bags and steep for 5 minutes.

Me: ! [Teabags were already added before, skipping this step.]

Home: Discard tea bags. Add milk and sugar.

My Chai tea with artistically placed spices

Me: [Discard, sure. Milk and sugar were already added.] Next!

Home: Bring tea just to simmer over high heat, whisking until sugar dissolves.

Me: [Nope, skipping this, too, as it was done already and incorrectly.] Next!

Home: Strain chai into teapot and serve hot. This was the last step.

Me: [Strain and sip.]

Even with all my mistakes, it actually turned out OK. Then again, Chai is a forgiving recipe.

Beyond the actual recipe, I do want to share a few thoughts on where cooking with voice is better than with a written recipe and where it still fails.

First, what I liked:

  • Home did a great job finding a good recipe. It found one fast, from a reputable site, and, above all, one that turned out tasty. Yes, Home is powered by Google, and, yes, I’ve heard that they do search quite well. Still, this was well done and it’s important to get this right.
  • When a step included doing something timed, Home asked me if I wanted to set a timer. This was a great help. That said, the interaction can be smoother and my asking for the next step (or any other voice command) should have turned the alarm off, instead of requiring me to ask it to stop specifically. Also, it seems that this only works for a certain minimum because when it said to steep the tea bags for five minutes, it didn’t offer a timer.
  • Home did a good job at parsing the recipe, pausing between each ingredient and instruction step, for no matter how long it took me to do that step. It never lost the thread of communication. That said, it wasn’t always waiting to hear my next question and often required “OK, Google” to continue.

Second, what I’d like to improve:

The bigger challenge is that the entire paradigm needs to change. Today a recipe is recited as it is written: first ingredients, then instructions. The big difference between reading a written recipe and following spoken instructions is that in the former the chef can go back when reading instructions and see the amount and preparation instructions for each ingredient. In the latter, reading instructions that refer to the ingredients is problematic. For example, the instruction to put the first five ingredients in the saucepan. But what were those five? When using a written recipe, a quick glance upwards gives the answer. When listening to spoken instructions, they need to list those ingredients in that step, even though the written recipe doesn’t include them.

Another example is ingredients list. It’s read aloud, one ingredient at a time, with prep and amounts. Some people cook like this and measure and prepare all ingredients before cooking. Others, like me, just like to make sure they have enough of each ingredient, then measure and prepare it at the step where it’s added to the recipe. When reading a recipe, cooks can work according to their preferred mode. With voice instructions, that’s currently impossible. What’s needed is an initial read-aloud of the ingredients with quantities so that cooks can verify they have all of them, and then a repeat of the ingredients for actual assembly and prep. 

Beyond the paradigm change, there are a few smaller things that can be improved:

  • Can I replace X with Y? Here’s where voice can shine. My recipe called for cinnamon sticks and I wanted to replace with powder. I asked Home if I can do that and it read me the relevant passage with a usable ratio of powder per stick. But the recipe called for two sticks, could Home just have given me the correct amount instead of the ratio? Yes, in this case it was to multiply by two, but this is something computers, in general, do better than humans. Do the math for us! Tell me “you’ll need 1 teaspoon ground cinnamon instead of 2 sticks.”
  • How much are we making? Add Home skipped the part about how many servings a recipe yields, which would have been very helpful to me. I only realized it when it read the amount of required water and I realized it was too much. It would be helpful to mention this bit of info ahead of time.
  • Can I make half? Another place where voice can shine is quantities. My recipe said it serves 6 and I only wanted to make it for 2. Home could ask after saying the current quantities of a recipe how much I’d like to make. Then it could automatically adjust the ingredient amounts, and just read what is necessary for the quantity I want. If there’s an undividable ingredient in the recipe, say an egg, Home could calculate the corresponding ingredient amount and say something like “There’s an egg in this recipe. You can make either 2 servings or 4 instead of the receipe’s 8. Which would you like?”
  • How much do I need again? This can allow users to reference the information they need when they need it.
  • How many calories/protein/carbs/vitamin D does this recipe have? Helpful for those on specific diets and this is just another calculation based on an understanding of the ingredients. It could also be compared to the RDA for each.
  • What size pot/pan should I use? I switched to a larger saucepan as I was adding water. Home can do that calculation for me and recommend a 4 quart saucepan from the beginning.

These are just a few things that can help make cooking easier with voice. If there’s a linked screen, Home could also show what the finished dish should look like, videos of prep techniques, or what an ingredient looks like. Voice could really make cooking more accessible to many more people, and that may be the biggest win of all.

Superbowl Ads 2018: social media has left the building

My favorite Super Bowl moment: Leslie Odom Jr singing America the Beautiful
Source: NFL on YouTube

Every year I recruit our Super Bowl party viewers to help me tally the number of hashtags, URLs, and social media handles and pages mentioned in the ads. This year we tallied mentions from just before America the Beautiful and to right after the final Hail Mary attempt by the Patriots. All in all we counted 80 ads as we didn’t include NBC show promos or local ads. Of those 80, a scant few had social media handles: 3 for Twitter, 1 each for Snapchat and Instagram, and none for Facebook. This is more or less consistent with previous years. Also consistent with the last few years is the percentage of ads including URLs, still surprisingly high at about a third of all commercials. Are advertisers afraid that viewers won’t find the site on their own, and if so, will a second on the screen really motivate going to the site?

My unofficial social media mentions tally

The real trend this year seems to be the continued decline of the hashtag. In previous years, more than 50% of ads had a hashtag, supposedly to encourage discussion of the brand on any (and every) social network. This year hashtag mentions continued their steady decline with only 10 ads, 13%, mentioning one.

The last second of the Tide ad, with hashtag
Source: Tide on YouTube

While one of the most talked ads of the Super Bowl, Tide, featured the hashtag #TideAd, it didn’t really need to as it was talked about everywhere anyway. Says Adweek: “But as of Monday morning, the top two brand stories trending on Twitter were Tide’s successful attempt to blanket the Super Bowl with ads referencing other ads and the backlash over Dodge Ram’s decision to use a Martin Luther King, Jr. speech to sell trucks.” Which may mean that advertisers have realized that in order to create a memorable and talked-about ad, it is no longer necessary to include social media promtps, it’s only necessary to create a good ad.  

I’m struck by this weird contrast between the steady mention of URLs in commercials vs. the decline of hashtags. If it’s no longer necessary to direct the conversation with hashtags, why the continued insistence of including URLs? What are the odds that a viewer will remember a URL, even a short one, flashed on screen for a second or two? If there’s a belief that viewers will even read the final frame, would it be more effective to just reinforce the brand name?

Finally, it’s interesting to note that even with the constant and ongoing upheaval in viewing habits, the Super Bowl has remained a reliable ratings magnet. This year it garnered a 43.1 rating, which really hasn’t changed that much in the last 40 years. In fact “the past ten Super Bowls rank as the ten most-watched all-time, and the past eight rank as the eight most-watched programs in U.S. television history.” The Super Bowl is probably the last show that Americans watch in such overwhelming numbers and consistently across the years. It’s no wonder advertisers, and agencies, love it.

 

A few thoughts about the ongoing misunderstanding around the use of personal data

Strava’s Heat Map can be beautiful. This was captured in the Nevada at the site of Burning Man.
Source: Strava

Over the weekend, an international security snafu erupted when it was “discovered” that anonymous Strava user data, actually released by the company last November, can be used to identify the location of up-till-now secret US Army bases, possibly putting soldiers in danger. 

Strava has defended the release of this information saying that  its map “represents an aggregated and anonymized view of over a billion activities uploaded to our platform, but also thatthe information was already made public by the users who uploaded it.” Strava also said that it’s “committed to helping people better understand our privacy settings.” It does sound a bit like Strava is blaming the users for not understanding what data is they’re sharing, what the default sharing setting are, and how to control those settings.

It comes back to this: “The controversy around Strava demonstrates a common issue with the relationship between tech companies and their users: People casually using an app often don’t understand what companies do with their data or how to properly protect it.”

This pervasive disconnect between what a company does with personal user data and what the users think it does with it plays out across so many different apps and services. As in most cases of “free,” if you’re not paying for the product, you are the product. Strava, a free app, has to make money somewhere, so along with premium membership it also sells aggregated, anonymized data such as cycling data: “We also have a metro business, which is aggregating and anonymizing commute data to sell that back to departments of transportation so they can better plan pedestrian bicycle routes in cities.”

Spotify: be humble.
Source: Adweek

This isn’t new. It just pops up every time a surprised media reports on some company’s data usage or another, sometimes with the backing of the company itself. In this case, Strava itself released the heat maps. Users were also outraged when Uber released its “walk of shame” data and Netflix and Spotify started using peculiar viewing/listening data in ads. Last year fury erupted when it turned out that Unroll.me, a service meant to unsubscribe users from unwanted emails, sold anonymized emailed Lyft receipts to Uber. CEO Jojo Hedaya’s response sounded familiar when he said it was “heartbreaking to learn that people were upset after discovering that Unroll.me sells data to make its service free. He believes the company wasn’t ‘explicit enough’ in telling users what it does, and that there will be clearer messaging in apps and the web.”

Facebook is another a free-to-use product that deeply profiles users and users often get upset when the depth of data is revealed. This week it reported that even with a decline in the number of daily users in the US and Canada and in the time spent on the site, an increase ad prices boosted revenue. Says Erin Griffith: “unfortunately for users, mining our personal data to better sell us stuff is the future of Facebook’s business. Over the years, Facebook has managed to rise above criticism over privacy and transparency by pointing to settings that purport to give users control over what they share and with whom. But those settings are complex and not well understood by users. Further, Facebook does not give users full control over their data.”

I’m curious if the disconnect between user’s expectation of privacy and what companies actually share comes from users not reading the terms of use or the intentionally vague phrasing of those terms. Perhaps the surprise comes from fact that the default settings for sharing are usually set at 11 and the controls for those settings are hidden deeply in the service to the point where users don’t realize they have a choice. Also, in many cases users don’t have a choice – either they consent to the collection of their personal data or they cannot use the app. In any case, I don’t think either data collection, the controls, or their default settings is going to change. Yet it would behoove companies to either communicate better about what data they collect and how it’s shared with their users or stop trying to use that data in their advertising or public relation attempts. Either get ahead of the outrage or don’t start it at all.

This week in Facebook: two steps forward, one step back

It’s been an interesting week (or was it a month?) of Facebook related news and as I’ve been critical about Facebook in the past year since the election, it’s nice to see that change is coming. Even more so, it’s commendable to see Facebook move from itsthe idea that fake news on Facebook influenced the election in any way is a pretty crazy idea” denial and to an awareness and admission of the monster they have created. To be fair, I haven’t seen other businesses, tech or otherwise, even come close to admitting something that detrimental to their business. However, it’s not all rainbows. Facebook seems to have taken two steps forward, but also one step back. Let’s start with the positive:

Step forward number one: admitting that Facebook may be bad for users mental health.

Can social media create truly meaningful social interactions?

Back in December, Facebook admitted something that critics and academics have been saying for a long time “Spending Time on Social Media Bad for Us.” In a separate post, Mark Zuckerberg said he would try to refocus the newsfeed on “One of our big focus areas for 2018 is making sure the time we all spend on Facebook is time well spent” and that he’s “changing the goal I give our product teams from focusing on helping you find relevant content to helping you have more meaningful social interactions.” Now, meaningful social interactions is a fine goal but it’s one that seems to contradict Facebook’s business goal, where the metric is engagement and the goal is to keep users on the site for as long as possible so that they can see more finely-targeted ads.

Spending less time on social media may be good for humanity but not so good for Facebook’s bottom line. This is why I admire this post. It’s extremely difficult as a successful business to admit that it harms humanity. Cigarette companies were forced to do so, big sugar hasn’t done so, and big soda is fighting it every step of the way. This isn’t and has never been a business norm so it’s refreshing that Facebook is going past denial and trying to do better.

Joe Edelman wrote two very interesting posts addressing this change. The first on what meaningful social interaction can mean online and a second that outlined possible ways to improve the experience and to actually promote significant interactions, and why it will mean completely remaking Facebook to reflect those different values. It’s fascinating to read someone who has both the deep understanding of what meaningful social interactions really mean and how to apply those principles to a virtual space. I highly recommend both.

Step forward number two: admitting that Facebook is bad for democracy.

In another candid post, Samidh Chakrabarti, Product Manager of Civic Engagement, went through all the talking points that Facebook critics (including me) have been making about how Facebook encourages divisiveness and doesn’t do enough to either block fake news stories or promote legitimate ones. Side note: Sarah Frier has mentioned that the same team at Facebook is also responsible for international work with “world leaders, some of whom use it against their citizens.”

In this article, the Washington Post takes an in-depth look into what happened a bit before and since the 2016 election to change Facebook’s mind. It was a long process and one mainly driven by Facebook’s employees.

The step back: using “the community” to judge the trustworthiness of news sources.

Facebook said last week that it would change its newsfeed algorithm (again) to prioritize “trustworthy” news sources and that it will let “the community” decide what are those are, with the goal of using that feedback to prioritize the newsfeed. Today it turns out that Facebook is doing that with a short, two-question survey:

  • Do you recognize the following websites? (Yes/No)
  • How much do you trust each of these domains? (Entirely/A lot/Somewhat/Barely/Not at all)

Will this simple, but lacking in nuance and impartiality, survey be the solution to rating trustworthiness? I doubt it. In the part of his post discussing false and misleading news items, Samidh Chakrabarti expressed awareness that anything Facebook does might not be enough. Said Mr Chakrabarti: “ Even with all these countermeasures, the battle will never end. Misinformation campaigns are not amateur operations. They are professionalized and constantly try to game the system. We will always have more work to do.”

While yes, there will always be work to do, maybe, given the extent of the problem and its far-reaching negative implications, it’s time to go beyond user-driven/community monitoring and machine learning and bring in actual, professional humans to be the editors. This may be Facebook’s intent in hiring “10,000 workers, including academics, subject matter experts and content moderators” – a team to understand and classify the content shared on Facebook. If that’s true, that, along with significant changes in the newsfeed, may lead to less fake and divisive posts shared which might lead to a better experience. While normally I’d end by saying that this will be interesting to follow, today I add to that the wish to see real change and results before more democratic institutions are harmed. Here’s to a better 2018.

The Mostly Random What It All Boils Down To 2017 Awards

There are new apps coming out every day and though I try a few new ones each month, I keep coming back to a few, at most 20, that I use on a regular basis. While some are amazing, some seem to be coasting on their existing reputation and the knowledge that we’ve come to rely on them. This year I also started using two apps more frequently, after realizing that they were just better than the competition.

Here, in no particular order, are my 2017 app awards:

My most used, not including the myriad messaging apps.

Best time-saver: driving with Waze. I used to think Google Maps was just as good as Waze but as I started driving longer drives in rush hour, I realized just how much of a difference Waze makes. Just this week I was able to test this with a friend when we ended up needing to travel the same 5.5 mile trip. I used Waze, which ended up taking me through a few side streets to the relatively clear highway while my friend was routed through a non-highway main street. For me, the journey took 15 minutes. My friend arrived 6 minutes later, which is an incredible difference for such a short journey. Is this because Waze is constantly optimizing the route while Google Maps picks the shortest one at the getgo and sticks with it? Either way, when road conditions are tough, i.e. every commute in the Bay Area, Waze is the better choice.

Caveat: there some situations where Waze isn’t great. First, in emergency situations. Take  Take this month’s fast spreading Southern California fires. When Waze noticed that some highways in the usually busy LA area were empty, they decided to route drivers through them. This doesn’t seem like an unsolvable problem but it is one that should be addressed soon. Second, as we reach peak highway congestion, with backups lasting for hours, it makes sense that Waze route traffic to city streets. That said, it doesn’t make residents of those streets happy and they’ve fought back by adding barricades and creating false reports. It will be interesting to see how this conflict plays out in the next year.

Finally, Waze launched support for car pool lanes earlier this month, but I haven’t yet had a chance to try it during rush hour. This feature can be great if Waze can now create a route based where the lanes exist and what times they operate.

Best use of machine learning to make me happy: Spotify Music. I first realized that music recommendations made it to “good” from “tolerable” was with Pandora a few years ago, but was loathe to give up my entire music collection, painstakingly collected on Google Music. Now, Google Music isn’t bad: it has mood playlists based on Songza as well as playlists based on popular music in different genres. I was also super impressed by its ability to create a really good playlist based on a single, selected song.

Yet, switching to Spotify opened a brand new world of recommendations and listening to new music that I don’t think I would have been exposed to otherwise. Their curated playlists are good but their personalized discovery playlists are incredible. I’ve had a chance to listen to and like music that wasn’t even on my listening radar. Within only a few months, Spotify knows my feels.

Extra bonus points: Spotify works well with Waze on iOS, with both respecting each other’s audio needs. This doesn’t work well with every audio – navigation combination so it’s commendable.

Finally, I’d love to see more voice-driven commands for both Spotify and Waze as I use them while driving. I’d rather not look at my phone screen for any reason, even if it’s just to glance at a map. It would be much better to ask for the information I need.

Most wow while getting the job done: Google Maps. Maybe not my winner for navigation, but certainly a boon both for saving local favorites and exploring a new city in preparation for a visit. Maps’ offline mode is also a great help, especially when traveling abroad without mobile data. I was also extremely impressed with the new ways Maps has been building layers of data to find new areas of interest. This blog post is worth your while on how Maps is getting better all the time.

Most unnecessary anxiety: all the messaging apps. Also, app problem I most wish solved in 2018. Today I have direct messaging set up on no less than 6 apps: Facebook Messenger, Google Hangouts, Google Voice for text, Slack DMs with different groups, WhatsApp with (mostly) international contacts, and, for extra fun, Twitter DMs. That’s way, way too many messaging apps yet there isn’t one I can turn off without losing important contacts which aren’t on any other service or without missing out important messages from less important people. Maybe we could all just go back to text messaging? Yes?

Most meh but used every day: Google Calendar. For the life of me, I couldn’t survive without calendar on my phone but it’s just standing still in terms of new feature development. How about smarter integration with mobile scheduling apps like Meetup and Eventbrite? How about helping set up meetings with several participants? How about smarter search of events, future and past? Better understanding of what events are important and what aren’t? Let’s add more smarts to Calendar this year, or please, can someone build a better Sunrise?

That’s it for me for 2017. May 2018 bring us the apps we need in this world.

YouTube Kids and the corruption of recommendation algorithms

Recommendation engines have been around for years, at least since Amazon started correlating shared purchases and suggesting products (or was it only books back then?) with “since you bought this, you might like this.” It was a good-enough recommendation algorithm that helped shoppers sift through endless options to find what was relevant for them.

The goals for today’s recommendation engines haven’t changed much from those early years: find the user things they like in order to either sell them more stuff or keep them on the site for a longer time to show them more ads, also known as “engagement.” Yet while today’s recommendation engines have the same goals, they’re coming up with completely different methods and results. On one hand we have Facebook’s tinkering with their Newsfeed algorithm, where trying to increase engagement has negative results such filter bubbles and promotion of extreme content. On the other we have Spotify’s amazing discovery playlists, such as the Daily Mix, that almost always delight me in their selection of new-to-me music. In between we have Netflix, which sometimes gets it right, and more selective stores, like Nordstrom, that do a decent job of suggesting products others bought. For most of these the recommendation engine is a black box for users, and its effects are measured religiously.

Yet with all recommendation engines, especially when user-generated content is involved, it’s a game between the platform, that decides what to recommend, and content providers, who try to influence that decision. What got me thinking about just how intensely this game is played is this very detailed post from a few weeks ago about how independent producers are gaming YouTube Kids. “Someone or something or some combination of people and things is using YouTube to systematically frighten, traumatise, and abuse children, automatically and at scale,” said the author, James Bridle. At the time, the post seemed too extreme and I waited for other analysts to weigh in. This week, John Biggs  at TechCrunch said conclusively: YouTube isn’t for kids. “YouTube is a cesspool of garbage kids content created by what seems to be a sentient, angry AI bent on teaching our kids that collectible toys are the road to happiness. YouTube isn’t for kids. If you give it to kids they will find themselves watching something that is completely nonsensical or something violent or something sexual. It’s inevitable.” This is as condemning as it comes.

YouTube’s and YouTube Kid’s reach is incredible. Earlier this week, Ofcom published a study on “Children and Parents: Media Use and Attitudes Report” in the UK and came up with these numbers:

Younger children especially watch a lot of YouTube Kids.
Source: Ofcom

  • YouTube is the content provider the highest proportion of 12-15s say they ‘ever’ watch – 85%.
  • YouTube is the only content provider, of the 14 examples, which is used by a majority of 12-15s to ‘often’ watch content.
  • Use of the YouTube website or app increases with the age of the child, accounting for 48% of 3-4s, 71% of 5-7s, 81% of 8-11s and 90% of 12-15s. Use of YouTube has increased since 2016 by 11 percentage points for children aged 3-4, by 17 percentage points for 5-7s and by eight percentage points for 8-11s. [Note: YouTube Kids was launched in February 2015.]
  • Half of YouTube users aged 3-4 (48%) and a quarter (25%) aged 5-7 only use the YouTube Kids app rather than the main YouTube website or app.

These numbers are high but not entirely surprising. Parents trusted Google when it said YouTube kids was child-friendly: “the app makes it safer and easier for children to find videos on topics they want to explore.” Its availability on every device has made it easy to access from practically everywhere. But with that level of trust, could YouTube Kids have done more to monitor content?

Last week, YouTube issued a response that, based on the comments on it, didn’t do enough. Of the five changes, only two went beyond guidelines: “tougher application of our Community Guidelines and faster enforcement through technology.” Also, ads will be removed from “inappropriate content.” What they didn’t do is allow parents to block specific providers or, as some requested, to whitelist channels or providers. The new restrictions don’t go far enough in allowing parents to control what their children watch and to block content they don’t want them to watch.  

What this proves, beyond the eternal axiom that we just can’t have nice things, is that once a service allows user-created content, it’s finding it extremely difficult to monitor that content. Beyond that, many creators, not only on YouTube, have figured out how to game the recommendation algorithm and get their content in front of viewers. I wish YouTube had taken a stronger stance with this, like blocking all creators aside from a few hand-picked ones until they can figure it out. Maybe that won’t be the most profitable choice, but it might be best from the product perspective. Until then, I have to agree with Mr Biggs: YouTube is not for kids.