Just Different Enough: AI Recommendations, Newsfeeds, Jokes, and Civil Society
By David Weinberger
AI Outside In is a column by PAIR’s former writer-in-residence, David Weinberger, who offers his outsider perspective on key ideas in machine learning. His opinions are his own and do not necessarily reflect those of Google.
Let’s say you’re an Orc in the Lord of the Rings and bought a copy of Morc the Orc’s Sauron: The Delights of Our Very Evil Leader from Mordor’s leading online bookseller, Harnen Books. When you return to that store, it’s likely to recommend more books praising Sauron. But if you had instead checked the book out from the public library, the kindly public librarian Pagewise the Orc is likely to have recommended a different set of books as the next for you to read.
That’s because Harnen Books’ algorithms aim at selling you more books, no matter what they are about, and history, common sense, and data analysis all show that someone who buys a book praising Sauron is more likely to buy another one praising him than one tearing him down. The bookstore thus has incentives to provide “recommendations of least resistance.”
But Pagewise, as a public librarian has a commitment to increasing and improving civic engagement, even in Mordor, so she may recommend many of the same books as the online bookstore does, but is likely to add something such as, “You know, there’s another book that I have a hunch you might enjoy: Shire Days: Life among the Hobbits. It’s by Xorc the Orc, a former Sauron supporter who does a great job presenting what our beloved Sauron (long may he reign) looks like from another point of view.”
But now let’s say the online Harnen Books suddenly becomes civic-minded and wants to take up Pagewise’s practice of recommending books that will broaden customers’ point of view. After all, with all the tumult in Middle Earth, it seems increasingly important for people to get out of their cocoons and echo chambers, understand the other side, and see if there are some common grounds for mutual understanding. (Sauron is not exactly on board with this, but Harnen Books has unexpectedly grown a spine.)
For the online bookstore to do this, it’s not enough that bookstores and libraries use algorithms that recommend any book that runs contrary to the reader’s views. For example, recommending Jolly Sauron: The Great Killers of Hobbits to a hobbit is going to be as effective as recommending Heroic Legolas: The Great Killer of Orcs to an Orc. You can’t just lob opposing books into people’s echo chambers because people are in those chambers precisely to avoid contrary opinions.
Instead, you’d have to do what Pagewise the Librarian does: Recommend items that are just different enough.
If “just different enough” were simply a quantitative measure of difference, the task would be simple. Alas, it is not. The useful types of difference vary depending on the domain and the person being given the recommendations.
For example, for book recommendations, the aim might be to broaden the reader’s horizons by suggesting works in the same genre that open up new political, cultural, or literary perspectives. If a reader has just enjoyed Isaac Asimov’s Foundation series, the librarian might recommend Heinlein’s Starship Troopers about the cost of imperial war, or Octavia E. Butler’s Dawn for reflections on war and the role of gender and otherness. But the recommendations-with-a-difference need not be in the same genre: “You might like this history of Rome that Asimov based the series on.” Or, “James Clavell’s Shogun is set in another world, but an historically real one, but is also about multiple territories struggling for power.” Or a different genre by the same author: “Have you read any of Asimov’s non-fiction? He brings the same sort of opinionated insight to his reading of Shakespeare.”
Yes, some of these recommendations may be far fetched but the point is not: Finding what is just different enough first means having insight into the relevant vectors of difference.
Finding the right difference is crucial not only for book recommendations, but also for news feeds. As is commonly feared, a news feed that aims only at keeping eyeballs locked on the streams and fingers ready to click the proffered links can reinforce the existing outlooks of readers while narrowing their sympathies. But simply inserting posts that readers don’t care about will not get readers to read those posts; you can lead a person to kale but you can’t make them eat. (It was like this before the Internet as well: newspaper readers turned the inky pages on articles about topics they didn’t care about.)
Sites that aim at building social connections should also want to bring together people who are just different enough, rather than always trying to bring together people who are most alike. This could include dating sites, social networking sites, and services that evaluate job and college admission applications.
The just-different-enough approach applies more broadly than just to recommendations. A related metric of just-enough-difference applies to humor: Jokes depend on surprise, and are often funnier the harder they are to “get.” If the stretch is too far, the joke is not intelligible. If the stretch is too short, as with a lazy pun, the joke only gets a quick snort or eye roll. Jokes require just enough difference.
So, if we can we find a way to generate results that are just different enough, it well might apply not only to prompts for the next book or video you might enjoy, but also to how news streams are composed and perhaps even to the generation of puns, jokes, and — who knows — fiction. Indeed, it might be a key for the generation of serendipity in many fields.
Neural Nets to the rescue?
Do we need to train neural network models to crack this problem?
Traditionally. straightforward statistical analyses of usage patterns often suffice for recommending books, films, and music. For example, if a pro-Mordor book shows up often in the cluster of books read by Orcs but with some minimal frequency in the cluster of books read by hobbits, then try recommending that book to more hobbits, especially if that book is highly rated by the few hobbits who have read it. Heck, you might even try recommending it to elves.
That’s fine, but it leaves discovery up to the circles of readers who are reading from outside their circle. If the echo chamber hypothesis holds — repeatedly hearing the same opinions hardens and narrows one’s views — then as social spaces become more cloistered, the range of discoverable books will narrow as well.
So, how to recommend books to the hobbits that are just different enough but that none of the hobbits have discovered yet?
Perhaps machine learning could use book abstracts, reviews, or even the full text of books to train a model to predict the existing social clustering, and especially the outliers. Identify the outlier books being read by clusters (the Sauron book being read by hobbit clusters), and analyze them looking for results based on cues in addition to the social ones. If machine learning can note patterns in the text of the books or comments about them that enable it to predict which outlier books have penetrated hostile echo chambers, then perhaps it can identify books that could penetrate those echo chambers if they were to be recommended.
If so, it would be interesting to see if we can learn what about a book enables it to succeed with groups that disagree with it. What makes a book just different enough? Of course, the machine learning model might work without yielding its secrets.
But, as you may have noticed, I am not a machine learning expert, even though I once did successfully train a model to classify photos of deserts and jungles with an accuracy just slightly north of random.
Serendipity is often seen as the solution to the problem of echo chambers. Indeed, Cass Sunstein, who popularized the echo chamber idea in a series of books starting in 2001, himself proposes increasing serendipity as an important part of the solution.
Yet, Sunstein’s explanation of the rise of echo chambers would seem to rule out serendipity as a solution. Sunstein’s argument, widely embraced, is that because people naturally prefer to read that with which they agree, if you give them an infinite buffet of reading options — i.e. the Internet — they will fill themselves up with comforting ideas. But if that’s what has caused the rise of echo chambers, then putting out yet more platters of things people disagree with will not suddenly cause people to decide to read them.
Serendipity by definition is a happy surprise. In the case of a book recommendation, the surprise can be the difference from what the reader usually reads. The happiness is that the work is pleasing despite, or because of, the difference. For serendipity to work, the work cannot be too surprising. It has to be just surprising enough, just different enough — the same as a joke.
If algorithms can genuinely engineer serendipity, machine learning can make a difference wherever serendipity counts. And that, arguably, is just about everywhere.