Have you ever looked back at a decision and thought, "What was I thinking?" Maybe you stayed in a job too long, bought something you didn't need, or made a choice that seemed perfectly logical at the time but turned out to be anything but.
Traditional economics and game theory would tell you that you made a mistake. But according to a rich body of research spanning decades, the mistake might actually lie with the theory itself. Let’s take a journey through the minds of the behavioral economists and game theorists who fundamentally reshaped how we understand human choice.
The Founders: Simon, Selten, and the Birth of "Good Enough"
The story begins in the 1950s, when economist Herbert Simon dropped a truth bomb that seems obvious today but was radical at the time: the human mind’s capacity for solving complex problems is inherently bounded. We are not supercomputers. We don't have perfect foresight, and we certainly don't always make optimal choices.
Shinji Teraji (2018) did extensive work clarifying Simon’s four core principles, which form the bedrock of this new perspective:
- Bounded Rationality – our cognitive limits shape our decisions.
- Satisficing – we seek satisfactory outcomes, not the absolute maximum (Simon's famous concept of "good enough").
- Sequential Search – we look for options in stages, not all at once.
- Adaptive Behavior – we change our strategies based on the environment.
Building on this, Gigerenzer and Selten (2001) took Simon’s ideas and ran with them, coining the concept of an "adaptive toolbox." They argued that humans don’t engage in heavy optimization or complex probability calculations. Instead, we rely on fast and frugal heuristics—simple, intuitive rules of thumb that get the job done quickly and efficiently. Selten (1998) would later lay the experimental groundwork for these ideas, identifying core behavioral features like presentation effects (caused by superficial analysis), reciprocity-based strategy construction, and "learning direction theory" as fundamental to observed human choice.
The Architects of Modern Models: How We Actually Quantify Irrationality
Theoretical work is great, but how do behavioral game theorists actually model bounded rationality? A major review by Kuzmanović (2020) highlights that researchers relax the strict assumptions of Nash equilibrium (which requires perfect best-responses and mutually consistent beliefs) in two primary ways:
- Quantal Response Equilibrium (QRE): Instead of always picking the absolute best move, Kuzmanović (2020) and Asl (2025) explain that QRE assumes players make "noisy" mistakes. They choose strategies with a probability proportional to their payoffs—meaning better options are more likely, but errors are expected.
- Cognitive Hierarchy & Level-k Models: These models, also covered by Kuzmanović (2020) , assume players occupy discrete levels of reasoning. A level-0 player acts randomly. A level-1 player assumes others are random and acts accordingly. A level-2 player assumes others think one step ahead. Asl (2025) elaborates that these models characterize players by their iterative reasoning depth, from naive to perfectly rational.
But the innovations don't stop there. Evans (2021) introduced a unifying "Quantal Hierarchy" model that uses two parameters to relax both best-response and mutual consistency simultaneously. The beauty of Evans' work is that it can actually recover Level-k, QRE, or even classic Nash equilibrium as limiting cases, giving researchers a single flexible framework.
Meanwhile, Xavier Gabaix (2011) took a different route with his "sparsity-based" approach. Gabaix argued that agents actively seek simplification; they build a partial representation of the world, ignoring irrelevant payoff dimensions. This ingenious method finally yielded tractable, closed-form predictions in games where none were previously available.
Adding a physical twist, Asl (2025) introduced a "Boltzmann weight" formalism. By parameterizing bounded rationality as a "temperature" per player, Asl created a model that smoothly interpolates between utility maximization (cold, precise) and completely random, equiprobable choices (hot, chaotic), while providing a full probability distribution over all joint strategies.
The Experimenters: What the Data Actually Says
Theory is one thing, but these researchers put their money where their mouths are through rigorous experimentation:
- The p-Beauty Contest: Kuzmanović (2020) and Gabaix (2011) both point to this classic game, where participants pick a number between 0 and 100 to get closest to ⅔ of the average. If everyone were perfectly rational, they'd pick 0. Yet most participants act as level-1 thinkers, choosing around 33. It’s a textbook demonstration of k-level reasoning failing to reach the Nash ideal.
- Tullock Contests: Lim (2014) ran experiments varying group size and discovered that logit-based bounded rationality models (like QRE) organize the "overdissipation" data remarkably well. Crucially, Lim found that rationality becomes even more restricted as groups grow larger—we get dumber in crowds, so to speak.
- Capacity Allocation Games: Chen (2012) found that the Nash equilibrium substantially exaggerates strategic over-ordering. Chen’s work showed that QRE-based models fit the experimental data much better, and while players become more rational through repeated play, they rarely converge to perfect rationality.
- Rock-Paper-Scissors & Prisoner’s Dilemma: Zheng (2025) observed that humans systematically depart from rational play, repeating winning moves and switching after losses (a "win-stay, lose-shift" heuristic). Zheng also noted that cooperation in the Prisoner's Dilemma consistently exceeds one-shot equilibrium predictions.
- Archetypes and Decomposition: Stahl and Wilson (1995) took a psychological approach, developing and testing a theory of five boundedly rational archetypes in 3x3 symmetric games. Their experimental evidence was striking—they rejected the rational expectations type but confirmed that the boundedly rational theory held up.
- The Invariance Problem: Jessie (2015) performed a sharp mathematical decomposition of standard models (QRE, Level-k, and Cognitive Hierarchy). Jessie found that these models exhibit mathematical invariance to nonstrategic game components. Yet, human subjects systematically respond to these very behavioral components that the models ignore—a critical reminder that our models still have blind spots.
Current Frontiers: AI, Limitations, and the Predictive Edge
So, where are we now? Interestingly, Zheng (2025) placed Large Language Models (LLMs) into identical experimental conditions. The result? LLMs reproduce familiar human heuristics, like outcome-based strategy switching and increased cooperation under repetition. However, Zheng found they apply these rules more rigidly and show weaker sensitivity to dynamic game changes, indicating they capture only a partial form of human-like bounded rationality.
We still face major structural hurdles. Asl (2025) points out that existing models often produce sets of possible outcomes rather than full probability distributions, meaning they cannot predict the relative likelihood of joint strategies or assign probabilities across multiple Nash equilibria. Furthermore, as Di (2016) notes, bounded rationality parameters remain latent and incredibly difficult to identify, making the cognitive processes underlying choice inherently challenging to model.
Despite these challenges, there is a massive silver lining. Bonau (2017) provides the ultimate vindication: across a breathtaking variety of applications, predictions that assume bounded rationality consistently outperform traditional rational models.
The Takeaway
The next time you make a "suboptimal" decision, remember the decades of work by Simon, Teraji, Gigerenzer, Selten, Kuzmanović, Asl, Evans, Gabaix, Lim, Chen, Stahl, Wilson, Jessie, Di, and Bonau. You aren't being irrational, you are being profoundly, predictably, and beautifully human. And thanks to these brilliant researchers, our economic and game-theoretic models are finally starting to reflect that reality.
Which researcher’s work resonated with you the most? Do you see yourself as a level-1, level-2, or simply a "satisficer" in your daily life? Let me know in the comments below!