Hello Dear Readers and welcome back to The Intelligent Friend!
I know, it's been a while since you've read this newsletter. A lot has happened in these weeks and I've been a little overwhelmed by some personal and professional changes.
But here we are, ready to get back to the fray! In these weeks I've read and studied a lot of research topics related to AI and I'm really glad to restart to write regularly on my newsletter. I would like to thank in the most sincere and direct way all the people who in these weeks have written to me, commented on my posts, shared them or subscribed to this newsletter (welcome to those who read one of my issues for the first time!). As usual, I won't go into more detail, but I wanted to explain the reason for this break and communicate my enthusiasm.
In The Intelligent Friend, of course, I will continue to explore the social and psychological aspects of AI through scientific papers as usual.
There will be though very little changes:
The newsletter will come out once a week, on Sunday (as always), and for now, it will be completely free and open to everyone;
Instead of talking only about the psychological and social aspects, we will broaden the spectrum a bit by trying to understand in general what research says about how AI influences the way we think, reason and behave. So we will also understand how AI can "augment" our abilities or have side effects;
You will find a format identical to the usual one, in which I will explain the theme of the paper, but I will also try to broaden the reflections with other references to make the discussion more interesting (always maintaining a reasonable length so that you can read it carefully without getting too tired!);
I will include a section in which I will continuously recommend issues and authors here on Substack. The community I found here is wonderful and I can't say enough about it.
Well, ready? I hope you are as excited as I am about this super restart of The Intelligent Friend.
Today we will explore a great topic through a truly enlightening paper: recommendation algorithms and the prediction of our choices. We will try to understand if, in fact, there is an effect (and what effect?!) on what we decide from the "big algorithms". Happy reading and welcome back!
The paper in a nutshell 🔦
Title: Predicting consumers’ choices in the age of the internet, AI, and almost perfect tracking: Some things change, the key challenges do not. Authors: Gal & Simonson. Year: 2021. Journal: Consumer Psychology Review. Link.
Main result: the analysis of the results of several studies shows how algorithms still have great difficulty predicting our choices that are new or unusual.
Intro
It's raining outside. You've dressed comfortably. You've grabbed a blanket. You've made yourself a cup of hot tea or wine (for those who prefer). It's Friday, and everything is ready for a viewing of a romantic comedy or a scary thriller on the couch at home. You open Netflix. Go to the home page. Almost unconsciously, among the recommendations "for you", you notice that a comedy that tells the story of how two apparently very different people get to know each other and become increasingly attracted to each other could be the choice for you. Netflix, in short, has "recommended" an intriguing film. You start it with a button and enjoy your well-deserved Friday rest.
You've found yourself in this scenario, right? How many times have we chosen a film based on the recommendations of Netflix, Disney+, Prime Video, Apple TV and so on. And, indeed, what would the platforms be without recommendations specifically for us?
Recommendation algorithms, AI-generated messages, and the platforms that make up social media rely on our past choices to increasingly try to understand what to show us as the next funny video on Tik Tok, psychological thriller on Disney+, or buy us as the next historical novel on Amazon. And this has led to serious concerns among consumers: are algorithms becoming so precise that they not only know and direct our choices toward something we like, but influence us to the point of making us watch, buy, consume more, or perhaps vote for the candidate they would like with more conviction? In reality, an affirmative answer to this question requires a (truly surprising) discovery. So, let's dive in.
I know your choice
The paper featured today is not an experimental study, but a Literature Review. For those who have never heard of it, very briefly and simply, it is a type of manuscript published in scientific journals in which the authors review the studies that have been published on a certain topic trying to understand at what point the results obtained are, what are the useful insights and what (in their vision) scientists should investigate as the next relevant elements.
Recommendation algorithms have made great strides in helping marketers, organizations, and platforms deliver effective messages targeted to us based on our past behavior.
Intuitively, the process is very simple: if 4 out of 5 movies I watch are comedies, Netflix might suggest I watch another comedy. Not only that, but if 3 out of 5 feature Al Pacino, it might suggest other movies, even non-comedies, in which the great American actor stars. In addition to an intuition-driven effectiveness sensor, studies actually tell us that this type of approach works: for example, having prior information about who visited a product website enhances the accuracy of predicting who is likely to make a purchase after seeing an ad for that product1. Furthermore, algorithms could improve their performance through demographic information2: if Disney discovered that "Inside Out 2" is very popular with Generation Z, it could recommend it more to users of that specific generation.
In today's paper, Gal and Simonson (2021) identify (among others) two major types of recommendation algorithms3:
Collaborative filtering: in one variant, known as a “user-user” algorithm, recommendations are made to a consumer based on items purchased or liked by other consumers with similar purchasing histories. An alternative approach, the “item-item” algorithm, recommends items based on similarities to products they previously purchased or liked. In this case, item similarity is often determined by how many users have purchased both items.
Content-based systems: this type creates recommendations by matching a user’s profile with item profiles, suggesting items similar to those the user has previously purchased or viewed: “for example, users might be recommended songs by artists or from genres that they previously listened to”.
These processes are naturally based on enormous amounts of data ("big data") that make these methodologies more sophisticated. Now, the real interesting point to reflect on is that although these algorithms have an enormous capacity to orient us, they have important limits that we often overlook, especially in behavioral effects, that is, when they try to try a specific behavior.
And these limits are represented above all by the surrounding environment of the choice that occurs.
As you know I usually do, to illustrate this, let's take an example: imagine that an algorithm is trained to identify a person's favorite brands based on a purchase made. Following the authors’ example, let’s imagine, for example, a toaster. However, a person who bought the toaster could not have been influenced by a deep conviction, but by contextual factors. Imagine for instance a notification on a discount or e-commerce app (such as Amazon itself). An immediate need for a toaster for a dinner with your boss a day later with the product suggested in various reviews. Consequently, recommending additional appliances from the same brand may not lead to particularly relevant suggestions for that user.
Where there is a limit, I read your brain
Now, as we know, the predictive techniques on which platforms and recommendations are based have improved a lot over the years, especially thanks to developments in machine learning. In particular, there is a specific category of algorithms that has received attention and to which part of the literature analysis of the authors of today's paper is dedicated: algorithms that are based on psychological characteristics.
Basically, to simplify, they are tools through which platforms are able to analyze the contents that are published, infer psychological characteristics or orientations and adapt targeted messages to the characteristics.
In short, a very high level of sophistication, and this information can be detected thanks to the texts we publish, the images we post, the brands we like on social media4567.
Although their effectiveness and ability to improve the analyses and results of various organizations remains undisputed, several scholars have wondered how much these influences actually translate into actual behaviors. From here, some truly interesting studies were born and which, especially from a methodological point of view, have stimulated a fervent debate in the scientific community.
I imagine, for example, that one of the cases that came to mind while reading these lines is that of Cambridge Analytica8. For those who don't know, the "Cambridge Analytica case" exposed how personal data was used to psychologically profile and target voters with tailored ads, raising concerns about the power of political microtargeting and its potential impact on the 2016 election. One interesting study related to this type of dynamic is the one by Matz et al. (2017)9.
The authors conducted a large-scale experiment and reported experimental findings suggesting it was possible “to influence the behavior of large groups of people by tailoring persuasive appeals to the psychological needs of the target audiences”. Specifically, by placing ads aligned with users' personality profiles for over 3 million Facebook users, they observed increases in clicks and conversions (i.e., sales) of up to 50% for a beauty product compared to non-tailored ads.
This fact might scare many of you. But let's try to go deeper. Gal and Simonson (2021) specify that such a result should actually be "taken with a pinch of salt", for several reasons:
First, personality traits are generally “modest predictors of behavior”10. For example, in the context of voting behavior, personality traits explain only about 5 percent of the variance in individuals' left-right political leanings11;
Second, personality profiles derived from online data - such as user "likes" or text analysis - are typically validated against self-reported measures, considered the “gold standard” for personality assessment12;
Furthermore, Matz et al. (2017) reported a click-through rate of approximately 0.3% and a conversion rate of approximately 0.01% (390 purchases from 3.1 million ad impressions) for personality-matched ads. The absolute conversion rate, in the view of the authors of today, was minimal, and even the gain from personality-matched ads (about 100 additional conversions) remained small.
Of course, as mentioned, these counter-arguments represent analyses that do not constitute absolute validity, but it is interesting - as well as important - to also reflect on this type of vision in the context of often complex phenomena.
Less is more?
Among the other very interesting studies that I mentioned, there are those who have tried (in a way that can be traced back to what the authors of today's paper said) to understand how much algorithms were actually more effective than other models in predicting our choices in certain contexts.
In several contexts, researchers have observed that simple models often perform nearly as well as advanced machine learning methods in predicting behavior.
For example, Jung et al. (2020)13 found that straightforward rules could predict certain behavioral outcomes, such as whether released defendants would await future court proceedings, with accuracy close to that of machine learning models.
However, one of the most intriguing studies that caught my attention is that of Kizilcec et al. (2020)14: in a large-scale study of 250,000 students across 247 online courses over 2.5 years, the authors used machine learning to identify students who would benefit most from targeted interventions to complete their courses. They found only a small advantage: students in the individualized intervention group had a completion rate of 13.38% compared to 13.08% in the group receiving a random intervention.
Furthermore, a recent large-scale collaboration15, 160 teams built predictive models for six life outcomes (e.g., children's GPA, eviction) using data from the Fragile Families and Child Wellbeing Study. The authors concluded that, despite a rich dataset and machine learning optimization, the best predictions showed only marginal improvement over a simple baseline model.
In relation to this latest large-scale experiment, furthermore, there is a reflection full of ideas by Garip (2020), with the didactic title: "What failure to predict life outcomes can teach us". The author noted, as reported by Gal and Simonson, that “the results produced by 160 independent teams using myriad strategies are clearly not an artifact of any one method and suggest that SML [Supervised Machine Learning] tools offer little improvement over standard methods in social science data”.
The power of algorithms and other factors
Now, at this point, should we think that perhaps algorithms have a negligible impact in predicting our choices and that we should not pay attention to this type of issue? The answer is, as you can imagine, absolutely not. On the contrary.
Science has shown that recommendations are effective and especially when they are combined with social elements (such as "bestseller" badges). Research shows that, for example, adding contextually relevant information, like location data for restaurant recommendations, can significantly improve recommendation accuracy16. And this is also something that is intuitive.
But neglecting environmental elements could be a serious mistake. This is in fact the underlying vision of the authors of today's paper. The impact of understudied environmental factors could in fact constitute a flourishing area of research to be studied together with algorithms to promote positive choices oriented towards the well-being of consumers and people in general.
Of course, I would like to point out that the reflections made are also valid, above all, for those choices that the authors define as "new, non-habitual, and require new evaluations, new sources of information".
To conclude, let's go back to our initial example. Imagine that although you are about to start the comedy, someone comes to visit you and shows you an action movie instead. Or maybe you receive a notification about a new sports documentary that has been released and you change your mind at the last minute.
In short, the power of algorithms remains undisputed and will certainly grow with more data and sophistication, but the joint impact with the decision environment can be something truly surprising and robust.
P.S. I hope that these reflections have stimulated several ideas in you as they have done for me!
Takeaways 📮
Algorithms vs. Context: while recommendation algorithms are powerful, their predictions often fall short when they overlook the specific context of each choice;
Limitations of Psychological Profiling: studies show that predicting behavior based on psychological traits or online activities (like the Cambridge Analytica case) is often limited;
Simplicity Competes with Complexity: in some scenarios, simpler prediction models perform nearly as well as advanced machine learning.
Further research directions
How does the presentation of consumer reviews (e.g., tone, device used to write, or storytelling elements) impact consumer reliance on them when making purchase decisions together with algorithms?
Are consumers more receptive to personalized recommendations, which they perceive as specifically tailored to them, or to general recommendations believed to be applicable to a broader audience? Which factors change these effects?
How do "beyond accuracy" factors in recommendations - such as novelty, surprise, or curiosity - affect consumer engagement and receptivity in recommendation systems?
The Highlight
As promised, this is the section where I'd like to highlight the amazing work that several authors do here on Substack, through links to their newsletter or specific pieces I've read. Here are first some of the authors you can't miss reading on this platform, who put care, passion and dedication into every single piece:
Artificial Intelligence Made Simple by
;The Absent-Minded Professor by
;Agora by
;AI Supremacy by
;- ;
Educating AI by
;The FuturAI by
;- ;
- ;
- ;
Graves Data Insights Substack by
;Marhworlds by
;- ;
New Things by
;- ;
Learning to Read, Reading to Learn by
;- ;
- ;
Rhetorica by
;- ;
- ;
Teaching computers how to talk by
;Uncharted by
;- ;
From the next issue I also plan to dedicate a short description, but in the meantime have fun exploring and discovering!
Thank you for reading this issue of The Intelligent Friend and/or for subscribing. The relationships between humans and AI are a crucial topic and I am glad to be able to talk about it having you as a reader.
Has a friend of yours sent you this newsletter or are you not subscribed yet? You can subscribe here.
Surprise someone who deserves a gift or who you think would be interested in this newsletter. Share this post with your friend or colleague.
P.S. If you haven't already done so, in this questionnaire you can tell me a little about yourself and the wonderful things you do!
Johnson, G. A., Lewis, R. A., & Nubbemeyer, E. I. (2017). Ghost ads: Improving the economics of measuring online ad effectiveness. Journal of Marketing Research, 54(6), 867–884.
Zhao, W. X., Li, S., He, Y., Wang, L., Wen, J. R., & Li, X. (2016). Exploring demographic information in social media for product recommendation. Knowledge and Information Systems, 49(1), 61–89.
Felfernig, A., Jeran, M., Ninaus, G., Reinfrank, F., Reiterer, S., & Stettinger, M. (2014). Basic approaches in recommendation systems. Recommendation Systems in Software Engineering (pp. 15–37).
Berger, J., Humphreys, A., Ludwig, S., Moe, W. W., Netzer, O., & Schweidel, D. A. (2020). Uniting the tribes: Using text for marketing insight. Journal of Marketing, 84(1), 1–25.
Hartmann, J., Heitmann, M., Schamp, C., & Netzer, O. (2020). “ The power of brand selfies in consumer-generated brand images”. Available at SSRN.
Hu, Y., Xu, A., Hong, Y., Gal, D., Sinha, V., & Akkiraju, R. (2019). Generating business intelligence through social media analytics: Measuring brand personality with consumer-, employee-, and firm-generated content. Journal of Management Information Systems, 36(3), 893–930.
Schoenmueller, V., Netzer, O., & Stahl, F. (2020a). “Polarized America: From political partisanship to preference partisanship. SSRN Electronic Journal,
https://www.nytimes.com/2018/04/04/us/politics/cambridge-analytica-scandal-fallout.html
Matz, S. C., Kosinski, M., Nave, G., & Stillwell, D. J. (2017). Psychological targeting as an effective approach to digital mass persuasion. Proceedings of the National Academy of Sciences, 114(48), 12714–12719.
Judge, T. A., Klinger, R., Simon, L. S., & Yang, I. W. F. (2008). The contributions of personality to organizational behavior and psychology: Findings, criticisms, and future research directions. Social and Personality Psychology Compass, 2(5), 1982–2000.
Furnham, A., & Fenton-O'Creevy, M. (2018). Personality and political orientation. Personality and Individual Differences, 129, 88–91.
Kosinski, M., Bachrach, Y., Kohli, P., Stillwell, D., & Graepel, T. (2014). Manifestations of user personality in website choice and behaviour on online social networks. Machine Learning, 95(3), 357–380.
Jung, J., Concannon, C., Shroff, R., Goel, S., & Goldstein, D. G. Simple rules to guide expert classifications. Journal of the Royal Statistical Society: Series A (Statistics in Society), 183(3), 771–800.
Kizilcec, R. F., Reich, J., Yeomans, M., Dann, C., Brunskill, E., Lopez, G., & Tingley, D. (2020). Scaling up behavioral science interventions in online education. Proceedings of the National Academy of Sciences, 117(26), 14900–14905.
Salganik, M. J., Lundberg, I., Kindel, A. T., Ahearn, C. E., Al-Ghoneim, K., Almaatouq, A., Altschul, D. M., Brand, J. E., Carnegie, N. B., Compton, R. J., Datta, D., Davidson, T., Filippova, A., Gilroy, C., Goode, B. J., Jahani, E., Kashyap, R., Kirchner, A., McKay, S., … McLanahan, S. (2020). Measuring the predictability of life outcomes with a scientific mass collaboration. Proceedings of the National Academy of Sciences, 117(15), 8398–8403.
Bao, J., Zheng, Y., & Mokbel, M. F. (2012 November). Location-based and preference-aware recommendation using sparse geo-social networking data. In Proceedings of the 20th International Conference on Advances in Geographic Information Systems (pp. 199–208).
So is the conclusion that advanced models really don’t work better than simple algorithms? Am I misreading that? So interesting considering how much is spent. What’s interesting is how textured recommendations have been for me from llms when I just feed some of my own data in. Like a download of my Goodreads data into Claude gives way better recommendations than Amazon. Why do you think this would be?
Thanks for highlighting me! Good to have you back.