Example, Early Results from Generative AI and Behavioral Economics Testing

As a follow-on post to summer 2023 exploratory work that is happening with the Behavioral Economics Research and Education (BERE) Lab, we’ve started to compile early results. Here are some test result summaries of different AI platforms based on the conjunction fallacy test (Linda problem). Note that platforms vary based on degree of live access to the internet and incorporation of slower System 2 thinking influences (although these characteristics are also confounded with platform implementation). Here we test ChatGPT 3.5, Bing Chat AI (based on GPT 4), and Google Bard.

Interesting questions to reflect on:
– How do AI platforms differ?
– Which gets things right?
– Which do you trust?
– To what extent will AI adoption get impacted by use case, accuracy, and trust?

The Behavioral Economics Research and Education (BERE) Lab by Stephen Shu

For the past two summers, I have personally volunteered to spend time helping a limited number of students pursue interests in behavioral economics and build their resume of experiences. For this summer, I will expand my efforts somewhat, although I hope to eventually find a more sustainable and scalable model in terms of funding, operations, and potential synergies with other organizations.

Here’s the natural extension of what I’ll be doing for the summer of 2023.

The Behavioral Economics Research and Education (BERE) Lab by Stephen Shu is an effort geared toward helping college students and young professionals with either the empirical or applied practice of behavioral economics. BERE efforts are in support of open science and the advancement of education. Where possible, students or young graduates may be supported by grants, and BERE welcomes opportunities to help students and young graduates obtain grants or corporate sponsorship.

Students and young graduates may pursue exploratory replication studies, expansion research studies, and corporate-focused research (for use in educational settings). Where possible, students are encouraged to develop empirical or professional skills (e.g., R or Stata statistical software, Python programming, communications, writing).

The research theme for 2023 includes exploratory work around the intersection of generative artificial intelligence (AI) and behavioral economics, such as similarities and differences between AI and human decision making across different platforms.

Understanding Core Principles of Behavioral Economics Can Help Us Become More Thoughtful Designers

A student recently asked me whether design thinking was different from behavioral economics thinking. In a nutshell, I believe the disciplines complement one another and should not be viewed as separate islands. That said, insights from behavioral economics and psychology can help us to become more thoughtful designers of products, customer experiences, etc.

One important behavioral area to consider is the role of memory in judgment and decision-making processes. Last week at Cornell Tech I facilitated a brief discussion of a new startup that was trying to address the issue of helping people to have perfect memory (as opposed to ever forgetting things). To what extent is this the perfect idea? What are some considerations from behavioral science?

To help feed that discussion, I relayed some results from a study by Eric Johnson and colleagues that provides substantiation for one theoretical role of memory in the decision-making process. Their study was conducted in the context of people valuing a commodity item, a coffee mug. In the classic example of the “endowment effect”, which is that people value things more when they possess an item (e.g., are given or endowed with a mug), people endowed with a mug, valued mugs at $6.01. People who were not endowed with a mug, valued mugs less at $3.72.

However, things got more interesting when the researchers manipulated the natural, unguided memory retrieval process. They manipulated the process by reversing the order in which people thought about things. Before having people value the mug, they essentially asked sellers to think about negative aspects of the mug, and they asked buyers to think about positive aspects of the mug. The endowment effect essentially vanished with sellers now valuing the mug at $5.05 and buyers valuing the mug at $4.91.

So memory retrieval order matters. If we go back to the startup example, and we have artificial intelligence (AI) based products remembering stuff so that we can make decisions, how should designers determine what to present to us first, knowing that presentation order may influence our decisions? Could presenting too many memories cause decision paralysis? If one is to pursue a product like the one described, the design choices are not trivial.

The bottom line is that hopefully we can improve our design reasoning by remembering to factor in insights from behavioral science.

Reference: Johnson, E.J., Häubl, G. and Keinan, A., 2007. Aspects of endowment: a query theory of value construction. Journal of experimental psychology: Learning, memory, and cognition, 33(3), p.461.

ChatGPT Makes Some Decisions in Similar Ways to People Although There Are Interesting Differences

Let me first preface this post by saying two things: 1) the state of research in this area is very young (e.g., most citations in 2022 and 2023), and 2) my summary will be at risk of oversimplifying things and missing some nuance.

The students I have at Cornell Tech are really sharp and energized. In one class, an interesting question was raised about whether AI could be used as part of the testing process, such as to A/B pre-test interventions. To try to get a better understanding of this space, I sought to do a little research on to what extent AI decision making resembles human decision making. So today I shared findings from a working paper that I recently read (Chen et al., 2023). The paper covers 18 common human biases relevant to operational decision-making (e.g., judgments regarding risk, evaluation of outcomes, and heuristics in decision making, such as System 1 versus System 2 thinking).

Here’s a summary of differences between ChatGPT and humans:

  • Judgments Regarding Risk – ChatGPT seems to mostly maximize expected payoffs with risk aversion only demonstrated when expected payoffs equal. It does not understand ambiguity. Also, ChatGPT exhibits high overconfidence, perhaps due to its large knowledge base.
  • Evaluation of Outcomes – ChatGPT is sensitive to framing, reference points, and salience of information. No sensitivity to sunk costs or endowment effect (e.g., may not have physical or psychological ownership concept).
  • Heuristics in Decision Making – More research needed, although aspects such as confirmation bias present. Additionally, ChatGPT has the ability to generate both classic System 1 responses (incorrect answers by humans typically driven by fast, automatic thinking) and System 2 responses (correct answers, such as those by humans which typically require more slow, reflective thinking).

While the reasoning for these modes of responses is not fully known, it seems as though ChatGPT is extremely logical when it comes to things like maximizing expected value. However, perhaps due to its nature of trying to be conversational and responding to salient information provided by the user, it can be overly sensitive to framing effects.

There are surely a lot things to think about, opportunities to pursue, and research to pursue.

Reference: Chen, Yang and Andiappan, Meena and Jenkin, Tracy and Ovchinnikov, Anton, A Manager and an AI Walk into a Bar: Does ChatGPT Make Biased Decisions Like We Do? (March 6, 2023). Available at SSRN: https://ssrn.com/abstract=4380365 or http://dx.doi.org/10.2139/ssrn.4380365

Reflections on Hong Kong UST DBA Program Session

This past week I gave a talk to students part of the Hong Kong UST DBA Program regarding implementing behavioral finance initiatives in companies. The talk covered some case studies that varied in different dimensions relative to the degree of integration of science and degree of organizational complexity. I have often emphasized that organizations that want to implement behavioral initiatives need to consider dimensions of Goals, Research, Innovation, and Testing (GRIT) among other behavioral-specific considerations (e.g., choice, information, process, and personalization architecture).

However, one of the most striking parts of the discussion for me surrounded the notion of ethics, which has come up a number of times in my discussion with students.

Although I was only able to touch on two angles in my HKUST talk, for the core classes I teach, I offer at least three different lenses for thinking about behavioral economics and ethics: 1) goal alignment between the company and the end user, 2) nature of behavioral intervention design (e.g., how much control does it exert), and 3) moral foundations and considerations (e.g., care/harm, fairness).

There are clearly other considerations that could come into play (e.g., to what extent comfortable sharing behavioral intervention thinking publicly; legal versus ethical 2×2). However, it is good that students think through ethical considerations. Things aren’t always as black and white as we’d might like, so it’s important to have multiple lenses through which one can evaluate situations.

Applied Behavioral Economics Rule of Thumb #8: Recognize the Limits of Forecasting Your Preferences and Hedge by Using a Process

Based on request from a Cornell student, last night I gave a talk to the Phi Sigma Pi National Honor Fraternity. The talk was entitled, “The Future You” and was themed around different career and leadership lessons that I experienced over more than three decades of work experience from engineer to management consultant to applied behavioral economics expert.

As part of the talk, I posed the question to students as to whether they could predict where they would be or what they would like, say several decades from now.

To set the context as to how well people can predict their tastes, I described a study by Kahneman and Snell. In the study, participants were first given a sample taste of plain yogurt and asked to rate how much they liked it. Participants then committed to eating a full serving of yogurt every day for about a week. They were also asked to predict how much they thought they would like the yogurt over the next week.

During the sampling phase, people had some dislike of the yogurt and predicted that their dislike would get even worse over the course of eating yogurt for a week.

However, what actually happened? People had very strong dislike of eating yogurt on Day 1 and the trend went in the opposite direction than predicted where liking improved over time instead of worsening. By the end of the week, while people were still somewhat negative on liking plain yogurt, by Day 8 their degree of liking was higher, even higher than what they reported during their first sample taste.

If we can’t predict our tastes for something simple like plain yogurt over the course of a short period of time like a week, then what are our chances of predicting things over a long horizon or even more modest time horizons?

My takeaways were the following:

  • Forecasting is hard, even forecasting the future you.
  • Interests, preferences, and skills develop and compound over time, so invest in them.
  • Find environments where you can learn and experiment (e.g., sometimes longer drive tests can be helpful).
  • To increase the chances of success and minimize the chances of overlooking blindspots, leverage 3rd party perspectives and out-of-the-box thinking tools from time to time.

Reference: Kahneman, Daniel, and Jackie Snell. “Predicting a changing taste: Do people know what they will like?.” Journal of Behavioral Decision Making 5, no. 3 (1992): 187-200.

Applied Behavioral Economics Rule of Thumb #7: Attention is Limited, So Maximize the Behavioral Impact Ratio

My level of awareness was heightened this week by correspondence with the brilliant Mac Hodell (former Principal at BCG). We were exchanging ideas about definitive references for management consultants related to the visual aspect of presentations. To cut a long story short, he brought up the powerful idea of maximizing the Insight:Ink ratio on slides.

How might this work? One could add ink to a presentation slide while increasing insight dramatically. For example, instead of using generic slide titles such as “Financial Impact,” it’s more effective to use specific titles that answer the “so what” question. As such, instead of leaving readers to draw their own conclusions, use titles that clearly state the outcomes of the work, such as “Our fall study of three behavioral interventions resulted in adding $250 million in AUM.” On the other hand, subtracting unnecessary ink from a graph on a slide might also work well, such as removing grid lines or merging overlapping legend information into the graph itself. There are also the aspects of substituting, synthesizing, aligning, or even redoing the content completely.

In a study by Mavis and Yoon, they posed a question to participants as to how they would change a Lego structure so that they could put a heavy object on top of it without crushing the figurine, recognizing that each block added would cost another $0.10. What did participants suggest? The title of the journal article, spells out people’s tendencies loud and clear, “Adding is favoured over subtracting in problem solving”. People tend to have an additive bias, and this could inadvertently lead to poorer design.

How to Improve the Stability of a Lego Structure?

This concept leads to maximizing impact ratios. In the behavioral world, one technique I teach students is to look at user journeys and touchpoints with users. Is each word needed in this email copy to drive engagement? How can we maximize Impact:Text? What about thinking about the end user and that their attention and time are limited? How to maximize Impact:Time? What about this complicated feature in the product? Can we cut down on the features and maximize the Impact:Features?

In closing, the rule of thumb is to subtract if possible, add if you must, and focus on maximizing the Behavioral Impact Ratio.

Reference: Meyvis, Tom, and Heeyoung Yoon. “Adding is favoured over subtracting in problem solving.” Nature (2021): 189-190.

Applied Behavioral Economics Rule of Thumb #6: Tap Into the Power of Thinking Architecture Tools To Help People With Decisions

Yesterday I outlined five primary areas of behavioral architecture that we would cover during the academic semester. Two prominent ones we covered included choice architecture (e.g., how choices are presented, such as with respect to defaults and number of choice options) and information architecture (e.g., how information is presented, such as percent of salary versus pennies for every dollar you earn).

As a third area, I framed thinking architecture as the process by which an architect tries to encourage end users to use more slow, reflective thinking versus fast, intuitive thinking. A classic question that tries to illustrate this is the following:

A bat and ball together cost $1.10. The bat costs $1.00 more than the ball. How much does the ball cost?

It is tempting for many people to think that the ball costs $0.10 (based on fast, intuitive thinking), although the correct answer is $0.05.

One form of thinking architecture could have been to get users to follow a checklist:

  1. Write down your guess for the ball cost (e.g., $0.10).
  2. Add $1 to the ball cost and write that number down as the bat cost (e.g., $1.10).
  3. Add the ball and bat cost (e.g., $1.20)
  4. If the numbers don’t add up to $1.10, repeat step 1.

In my view of thinking architecture, we are essentially trying to slow down both the brain to try to get the mind to follow certain thinking pathways. Whereas neoclassic economics doesn’t account for thinking pathways, path dependency is everything in psychology and affects behavior.

Implementations of thinking architecture space have been less explored. However, the possibilities are endless. Some examples include:

  1. Addressing complexity (e.g., aviation pre-flight checklists for pilots)
  2. Helping to avoid common errors (e.g., blindspots such as forgotten opportunities as to how I might want to use money in retirement or risks of underinsuring myself when younger)
  3. Expanding the thinking (e.g., are there other potential ways of realizing goals during retirement)
  4. Weighing difficult tradeoffs (e.g., should the money be used to save a single life or implement an equipment upgrade)

Thinking architecture requires thoughtful design, and it is not always the easiest to implement. However, the human mind is an amazing wonder. Sometimes we need to really tap into its power.

References: Frederick, Shane. “Cognitive reflection and decision making.” Journal of Economic Perspectives 19, no. 4 (2005): 25-42.