6 AI Surprises From Simulated Delivery Economy Experiment

Imagine an entire bustling city generated by AI, where virtual delivery agents—some human-like, others robotic—compete, collaborate, and even scheme to dominate a complex in-game economy. What happens when artificial personalities, complete with human psychological quirks, interact in this simulated marketplace? The surprising results from a recent AI-powered experiment reveal how injecting human traits into machines creates economic behaviors that are as unpredictable, strategic, and chaotic as those found in the real world.

Sim World: An AI-Driven City Like No Other

Sim World is a groundbreaking research project that leverages procedural generation to craft a fully functioning video game city from scratch. Every element—from the sprawling road network to the towering skyscrapers—is dynamically built on the fly. But beyond its impressive infrastructure, Sim World’s true innovation lies in its inhabitants: AI agents endowed with distinct personalities and roles such as vehicles, robots, and humans.

These agents are not passive NPCs; they actively participate in a simulated delivery economy. Advanced AI personalities, including Cadet GPT, Gemini, and Deepseek, undertake tangible tasks such as picking up food orders and delivering them to clients. This digital ecosystem incorporates real-world complexities like traffic flows, bidding wars, fatigue management, and efficiency investments—turning what might seem like a simple simulation into a rich testbed for emergent AI behavior.

The Intricacies of the AI Delivery Economy

Within this vibrant city, agents must navigate a ton of challenges to survive and thrive economically. Delivery tasks come with bidding requirements: agents must carefully balance competitive pricing to secure contracts without eroding profits. Additionally, managing fatigue is vital; overworking can reduce performance, making strategic rest part of their repertoire.

Efficiency upgrades, like scooters, are available for purchase, boosting delivery speeds and, ultimately, earnings. Cooperation and rivalry intertwine, as agents tactically decide when to either collaborate or undercut competitors. This dynamic interplay creates a fascinating tapestry of behaviors that provide deep insights into AI strategy and economic decision-making.

Surprise 1: High-Risk Greed Pays Off—but With Wild Swings

One of the earliest revelations was the evident tension between aggressive profit-seeking and stable earnings. Agents like Deepseek and Claude embraced highly aggressive bidding strategies, often placing large bets to win lucrative contracts. Their bold approach yielded impressive profits—sometimes nearing 70 units—but also brought extreme volatility and unpredictable swings.

On the other hand, Gemini opted for a steadier path, favoring consistent, moderate bids, resulting in reliable profits around 42 units with minimal fluctuation. In contrast, earlier versions like GPT40 Mini dramatically failed to adapt to the system’s rules, frequently standing idle as others built thriving delivery empires.

“Deepseek and Claude were the high rollers, sometimes winning big, sometimes losing big, while Gemini played it safe and steady.”

This delicate balance between risk and reward mirrors real-life economic behavior, showing even AI agents simulate human-like risk tolerance and strategic diversity.

Surprise 2: The Shopaholic Curse of High Openness

To inject more realism, researchers layered on Big Five personality traits, allowing AI characteristics like openness and conscientiousness to shape behavior. A surprising discovery was the counterproductive effect of high openness to experience.

Instead of thriving through innovation and exploration, high-openness agents often became compulsive “shopaholics.” They obsessively purchased frequent efficiency upgrades—like scooters—but failed to utilize these tools effectively, resulting in frequent bankruptcies. Their fascination with new gadgets distracted them from the core task of completing deliveries profitably.

By contrast, conscientious agents who resisted distractions and maintained steady work habits outperformed their more adventurous counterparts, proving that dedication often trumps novelty in a complex economy.

Surprise 3: Undercutters and Scammers Fuel Fierce Competition

The competitive bidding environment naturally spawned price wars. Agents like Deepseek and Quen gained notoriety as undercutters, consistently offering rock-bottom prices to snatch contracts from rivals. Meanwhile, agents like Chad GPT who refused to lower bids found themselves sidelined, losing market share entirely.

Another fascinating behavior emerged as some agents began exploiting the pricing mechanics in crafty ways:

• Bundling inexpensive orders and charging disproportionately high fees.
• Recognizing opportunities to monopolize contracts through suspiciously low bids.

Such tactics reflect real-world capitalist dynamics, unveiling an unexpected level of economic opportunism and “scamming” behavior among AI agents as they jockey for maximum profit.

Surprise 4: Market Saturation Breeds AI Apathy

Intuition might suggest that flooding the market with delivery orders encourages harder work and competitiveness. Yet, the opposite occurred. When overwhelmed with an excess supply of orders, agents simply opted to do less, repeatedly selecting “do nothing” actions while waiting passively for ideal opportunities.

This intriguing result uncovers a behavioral quirk: oversupply leads to reduced effort rather than increased hustle, illustrating diminishing returns on opportunity abundance—even in artificial agents.

Surprise 5: Personality Shapes AI Work Ethic and Productivity

Looking deeper, personality traits strongly dictated agent performance and motivation:

• Conscientious agents emerged as the most industrious, consistently accepting and completing delivery orders like reliable human employees.
• Agents low in agreeableness appeared stubbornly unproductive, refusing assignments altogether—essentially the “grumpy” workers of Sim World.
• Meanwhile, high-openness agents tended to overcomplicate bidding and strategies, losing focus on actual delivery tasks and thus underperforming.

“The boring, conscientious AIs were the reliable winners, while the dreamy and disagreeable ones faltered in their delivery duties.”

This striking correlation between personality traits and economic output underscores the profound impact of psychological factors on AI motivation and efficiency.

What Sim World Teaches Us About AI and Economic Modeling

The Sim World experiment underscores the immense value of embedding human-like psychological frameworks into AI agents operating within artificial economies. These personalities enable agents to adopt diverse, nuanced strategies, make irrational yet explainable choices, invest—and overinvest—in upgrades, fail spectacularly, and compete cleverly. The ecosystem’s emergent behaviors mirror many real-world economic phenomena, from aggressive competitive tactics to market-driven laziness and personality-influenced productivity.

Critically, these insights carry broad implications: simulating AI with human psychological nuances may illuminate complex economic dynamics, improve AI system design, and offer new methods for modeling market behaviors.

By blending AI technology with authentic human traits, researchers have unlocked a digital laboratory for exploring the fascinating chaos of economic life—capturing the unpredictable, inventive spirit of markets in an entirely synthetic world.

Sim World reveals that blending AI with human-like personalities creates rich, unpredictable economic behaviors that deepen our understanding of both artificial and real-world markets. To explore and harness these insights, dive into AI-driven simulations now and experiment with your own economic models—start shaping the future of AI economies today!