AI Plotting Global Takeover? It Struggles to Suggest a Decent Purchase!

Human + AI > AI

By Troy Lowry

“The best minds of my generation are thinking about how to make people click ads.” - Jeff Hammerbacher

There’s been a lot of concern about AI taking over the world. In large part, this is because AI can do some things, such as playing chess and folding proteins, at a superhuman level. It does these at a level so far above humans that we can’t hope to match it. Chess and protein folding , however, take place in a limited space, and they have a limited number of clearly defined rules. Physics and human interactions are far more complex, and I don’t believe AI is anywhere close to mastering them at a human, let alone superhuman, level. Until it can do a good job at suggesting items I want to buy that I didn’t already search for, I won’t waste a moment worrying about it taking over the world.

We’ve all had the experience; let’s say you are looking around the web for boots. You find the perfect pair and buy them. Afterwards, for weeks all the ads you see online are for boots. You think, I already bought my boots; show me something I might want to buy!

Why do you see these ads after you’ve bought the boots? In short, because serving ads about things you’ve searched for has a much higher chance you will click on the ad and go on to make a purchase. Because the AI serving you ads is watching what websites you go to and what you look at but does not know when you’ve purchased something. It doesn’t know when to stop.¹

Past Performance Is No Guarantee of Future Results

At its heart, AI is a statistical engine. It excels at finding patterns in large amounts of data. This is why it can identify cat photos more accurately and with orders of magnitude faster than humans.

One of the first big public demonstrations of statistical engines predicting what people would do was Netflix's recommendation engine . My feeling is that while it is better than most sites’ AI, the engine is still far from being superhuman, let alone a threat to dominate Earth. For instance, looking at it now, I see it recommending a Jerry Seinfeld stand-up routine, a few lesser-known Monty Python pieces, the movie “School of Rock,” and a CCR concert film. This is because I recently rewatched “Monty Python and the Holy Grail” as well as binge-watched the series “Seinfeld.” “School of Rock” is recommended because I just binge-watched “The Amazing Race” Season 14, on which the writer of School of Rock was a contestant, so I was searching the web about him. The CCR film is because I watched the excellent Taylor Swift documentary: “Miss Americana.”

In short, one of the best recommendation engines on the planet recommends things very close to things I recently watched but nothing I want to watch now.

Don’t get me wrong, these are good recommendations. Netflix claims that 80% of what people watch on its site comes from recommendations.² These recommendations are not, however, completely personalized. Rather, Netflix has set up 1,300 “recommendation clusters .” With almost 250 million subscribers, this means each of us is grouped with an average of almost 200,000 other people by what we like to watch. We are assigned to a group, and then Netflix recommends whatever other members of that group have watched. Your previous interactions with the site are factored into what it chooses for you, but it chooses from the cluster of other people’s choices.

Think of it this way: it has a list of possible choices based on other people who have watched similar movies, and then it orders that list based upon your previous activity. When you first sign up for Netflix, it asks you questions about which movies you will like specifically so it can put you into a cluster.³

On the other hand, I often get recommendations from friends: “Did you see ‘My Octopus Teacher’! Amazing!,” “‘John Wick’ was non-stop and awesome,” “If you like Taylor Swift, you should see ‘Miss Americana.’” None of which I would have watched had a friend not recommended them.

Netflix’s recommendations are strong, but they aren’t superhuman. If they were, I’d feel it was “reading my mind” and constantly giving me the exact thing I wanted to watch. Being based on what similar users have done has a lot of power but significant limitations. In particular, it rarely tries to broaden my horizons by offering me new genres. It treats me as one of a group instead of as a person with my own wants and needs.

Who’s Afraid of a Spreadsheet?

In my opinion, the worry about AI taking over, or even putting many people out of jobs, is misplaced, or at least premature. I am an AI evangelist; I see it transforming much of society and even more of work and academics. It is a powerful tool that, used carefully, will help usher in a better society. But I do not see it taking over the world.

Everything in Netflix’s recommendation engine could be done in an Excel spreadsheet.⁴ Is anyone afraid that spreadsheets are going to take over the world?⁵

This is not to say that AI won’t be a powerful tool used for both good and ill purposes. It is a certainty that authoritarian societies will use AI’s ability to correlate massive amounts of data and facial recognition to repress parts of their populations. We should be focusing our attention on protecting individual rights in the age of AI due to misuse by humans instead of worrying about AI taking over the world.

The day AI can recommend purchases I haven't searched for is the day I will start fearing its dominance.

1. There’s more to this story, including different companies with different data and how you are tracked across devices. A great deeper dive into that process is available at Digital Trends .
2. In other words, 20% of what is watched is from people doing searches. Given that my iPad is showing 24 recommendations on the screen right now, shows over 200 if I scroll up and down and well over 1,000 if I scroll left and right, the chances I’ll find something in there I want to watch is quite high and not an indication of superhuman predictive capabilities.
3. Netflix also uses some A/B testing where it will randomly give you one of two choices, see what you pick, and base further recommendations from there. I find this approach intriguing.
4. Ok. There’s a little license here. In Excel, it would be far too slow, by many orders of magnitude, to be useful, and Excel doesn’t natively support neural networks, matrix factorization, or probabilistic graphical models and would need to call out to libraries. That said all could, in theory, be created in Excel spreadsheets directly.
5. In a way, they already have. In many organizations, Excel spreadsheets run critical parts of the organization!