89.5% of People Evacuated When They Knew the Exit — What Happens When They Don’t?

89.5% of simulated pedestrians evacuated safely when they knew where the exits were. Just 1.05% made it out when they didn’t. That stark difference isn’t from a real-world disaster drill—it’s from a new artificial intelligence simulation that models how people behave when a threat moves through a public plaza, and it reveals something unsettling: knowledge of escape routes matters more than personality, more than fear, more than crowd density. In life-or-death moments, information is the difference between survival and catastrophe.

This isn’t a traditional evacuation model. It doesn’t rely on fixed rules like “move toward the nearest exit” or “follow the crowd.” Instead, it uses large language models (LLMs)—the same AI behind tools like ChatGPT—not as chatbots, but as digital minds inside each simulated pedestrian. Each agent perceives its surroundings through a symbolic ASCII view (think: a text-based map made of characters like @ for self, A for attacker, E for exit), remembers what it’s seen, and decides what to do next based on its personality, emotions, and past experience. The result is a crowd where no two people react exactly the same way—not because of random noise, but because each has its own evolving understanding of danger.

And in this world, one factor towers over all others: usable-exit knowledge.

The Science

The framework, developed by Jian Ma and colleagues at Southwest Jiaotong University and Fujian Police College, is designed to simulate how real people might respond to a moving threat—say, an armed individual entering a crowded square. Traditional models struggle with the messy, cognitive side of evacuation: how people interpret ambiguous cues, remember partial information, and change their minds under stress. Most rely on pre-programmed behaviors—rules like “avoid congestion” or “follow others”—which may capture general trends but miss the nuances of individual decision-making.

This new approach replaces those rules with LLM-powered agents. Each agent has three core components:

1. Perception: A private, symbolic view of the world encoded as ASCII characters within its field of view. Unknown areas appear as ?, making uncertainty explicit.
2. Memory: A knowledge graph built from personal observations. Each agent remembers only what it has seen—no omniscient awareness.
3. Decision-making: A prompt-driven LLM that weighs personality, current perception, emotional state, and memory to choose actions.

Personality is modeled using the Five-Factor Model (OCEAN: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism), with five archetypes defined by specific T-scores (Table 1). For example, the “Anxious-conformist” scores high in neuroticism (72) and agreeableness (60), while the “Bold-solitary” is high in openness (68) but low in extraversion (35) and neuroticism (32). These traits shape behavioral anchors in the agent’s prompt—e.g., a high-neuroticism agent is described as someone who “hesitates and revisits choices when a situation is unclear.”

Crucially, all agents share the same decision logic and sampling settings (temperature = 0.2); differences arise solely from their embedded personas and individual experiences. This ensures that behavioral heterogeneity emerges organically, not from hardcoded rules.

The simulation takes place in a virtual public plaza with gates, trees, walls, and multiple exits. A threat—represented by the symbol A—moves dynamically through the space. Agents can see the threat directly, hear shouts from others, or infer danger from indirect cues like a person lying on the ground (D). Their decisions are validated by a physics engine that checks whether a chosen route is physically possible on their known map—meaning an agent can’t move toward an exit it hasn’t seen.

After each run, the researchers reconstruct each agent’s memory and decisions as a knowledge graph, allowing full auditability: you can trace exactly why an agent chose a particular path, based on what it knew at the time.

What They Found

The team tested eight different personality compositions across eight paired randomized blocks—a total of 10,947 decision events analyzed. The results were striking.

First, exit knowledge dominated evacuation success. Of agents who had observed and remembered a usable exit, 89.5% successfully evacuated. Among those who hadn’t, only 1.05% made it out. This isn’t a marginal effect—it’s a near-deterministic relationship. As the authors note: “usable-exit knowledge was strongly associated with evacuation success.”

Second, personality shaped how agents interpreted ambiguous threats—but only before direct sighting. When danger was inferred from memory or hearsay, personality mattered. The proportion of agents assessing the situation as dangerous varied by 26.5 percentage points across personality types. Some were quick to assume risk; others remained skeptical.

This divergence was especially clear in movement behavior. The study measured “high-urgency, low-directness” decisions—those where an agent reported feeling urgent but moved inefficiently, perhaps darting erratically or hesitating. This behavior ranged from 11.41% to 34.33% across personality compositions, a nearly threefold difference.

Proportion of High-Urgency, Low-Directness Decisions by Personality Composition

Label	Value
Composition A	11.41
Composition B	18.25
Composition C	22.17
Composition D	26.83
Composition E	30.05
Composition F	32.14
Composition G	33.67
Composition H	34.33

Source: Figure 7b, Ma et al. (2026). Range: 11.41% to 34.33%.

But once the threat was seen directly, behavior converged dramatically. After a visual sighting of A, 99.6% of agents classified the situation as dangerous. Movement was selected in 99.8% of decisions. Personality, prior skepticism, and individual differences collapsed in the face of unambiguous evidence.

The figure above shows how urgency and movement directness varied across decisions. Most agents fell into the high-urgency, high-directness quadrant—moving fast and purposefully. But a significant minority (18.9%) were highly urgent yet indirect in their movement, suggesting panic-like behavior. This pattern was more common in certain personality types, particularly those high in neuroticism.

Evacuation time, meanwhile, was jointly influenced by three factors: spatial geometry (how far exits were), information access (whether the agent knew an exit), and affect (emotional state, driven partly by neuroticism). In other words, knowing the exit got you out; being calm and close helped you get out fast.

Why This Changes Things

For decades, evacuation planning has treated people as interchangeable particles—governed by flow rates, densities, and average speeds. This study challenges that assumption by showing how deeply individual cognition shapes outcomes. It’s not just about how many people are in a plaza, but what they know, how they think, and what they’ve seen.

The finding that information access trumps all has profound implications. It suggests that in real-world emergencies, the most effective intervention may not be better architecture or faster alarms—but better information distribution. If 89.5% of people evacuate when they know the exits, then the priority should be ensuring that knowledge is widespread, persistent, and accessible even in chaos.

Consider current public spaces: museums, train stations, stadiums. Exit signs are often small, poorly lit, or obstructed. Wayfinding is assumed to be static—something you figure out when calm, not during panic. But this simulation shows that in a crisis, most people won’t know where to go unless they’ve seen it recently and personally. Memory is fragile; perception is limited.

And here’s the twist: the agents in this model aren’t irrational. They’re not ignoring exits. They simply haven’t seen them. In the simulation, agents only know what they’ve observed. No omniscient map. No GPS. Just eyes and memory. That’s closer to reality than we’d like to admit.

The role of personality is more nuanced. It doesn’t determine survival—but it shapes how people respond to uncertainty. Some are quick to flee at a shout; others wait for proof. This has implications for emergency communication. A one-size-fits-all alarm may not work. High-neuroticism individuals may need clearer, more repeated cues; high-conscientiousness types may benefit from structured guidance.

But when the threat is visible, none of that matters. The human response to direct danger is remarkably uniform. This aligns with psychological research on threat detection: once a danger is perceived, cognitive diversity collapses into action. The brain’s survival circuits override deliberation.

What’s revolutionary here is not just the use of LLMs, but the auditable heterogeneity they enable. Unlike black-box neural networks, this framework allows researchers to reconstruct why an agent did what it did. You can trace a decision back to a specific observation, a memory update, a personality prompt. This isn’t just simulation—it’s a microscope for human behavior.

What’s Next

This work opens several paths forward. First, validation with real human data. The authors note that their model generates behavior that resembles human decision-making, but it hasn’t been tested against actual evacuation trajectories. Future work could compare LLM-agent choices with data from drills, VR experiments, or even historical events.

Second, scaling to larger, more complex environments. The current simulation is a single plaza. Real cities have layered spaces—underground tunnels, multi-level buildings, mixed pedestrian-vehicle flows. Can this framework scale? And how do social dynamics—leadership, herding, communication networks—emerge when agents can share information more richly?

Third, designing for information resilience. If knowledge of exits is the key to survival, then urban design must prioritize cognitive accessibility. This could mean:

• Larger, more visible exit signage
• Dynamic wayfinding systems that adapt during emergencies
• Pre-emptive spatial memory training (e.g., gamified orientation apps)
• Architectural features that naturally guide attention toward exits

Imagine a train station where digital displays don’t just show train times—but during an alert, pulse with directional arrows toward the nearest known exit, tailored to your location. Or a stadium where augmented reality glasses (or even smartphone AR) highlight escape routes in real time.

There are also limitations. The model assumes agents act independently—no physical pushing, no group coordination beyond shouting. It doesn’t model injuries, disabilities, or language barriers. And while the LLM provides rich decision-making, it runs in a symbolic world far simpler than reality.

Yet even with these constraints, the core insight stands: in a crisis, knowing where to go is the single most powerful advantage a person can have.

This isn’t just about plazas or threats. It’s about the design of all public life. Every building, every transit hub, every gathering place should be judged not just by how many people it holds—but by how well it prepares them to leave.

As climate change increases the frequency of urban emergencies—from wildfires to floods to extreme heat—the need for cognitively informed design will only grow. We can’t prevent all threats. But we can ensure that when they come, people know the way out.

The future of safety isn’t just stronger walls or faster alarms. It’s smarter information.