Prologue: A Seemingly Obvious Question That Has Rarely Been Properly Tested
Within the autonomous driving industry, “being more rule-abiding” is often treated as a near-synonym for “being safer”; and “being safer” is then quietly substituted for “being more acceptable to society.” The chain of reasoning is tidy enough to be taken as self-evident: since human driving involves impulsiveness, distraction, and violations, a system that strictly enforces traffic laws, never tires, and remains consistently cautious should, in principle, deliver higher public safety and earn stronger public trust.
Yet real roads do not always validate that reasoning. Across extensive testing and pilot operations, practitioners repeatedly encounter a subtler—and harder to quantify—form of friction: the system behaves in a formally compliant manner, but consistently creates social discomfort, hesitation, probing, and even antagonistic interaction. This friction often does not immediately surface as a crash, and it is difficult to capture with existing KPIs; nevertheless, over time it tends to materialize through a familiar pathway: degraded experience, eroding trust, tighter regulatory scrutiny, and constrained operations.
If compliance naturally leads to acceptance, why do the most compliant behaviors sometimes become the least accepted?
Starting from that question, the discussion can move beyond “how to behave like a stricter driver” toward something more fundamental—and more difficult: is an autonomous system integrating into a rule system, or into a social system?
Phenomena: Typical Scenarios Where the System “Does Nothing Wrong,” Yet Still Fails
The following examples are neither rare nor exotic; in fact, they have become emblematic “compliance failures (in a social sense)” in autonomous driving practice over recent years. Their shared trait is simple: the system does not visibly break the rules, yet it “speaks the wrong language” in the structure of interaction.
First: At Intersections Where It “Should Go but Dares Not” — Social Friction Induced by Absolute Right-of-Way
At intersections or unsignalized crossings with clear priority, an autonomous vehicle is, in theory, “supposed to proceed.” But when cross traffic, pedestrian intent, occlusions, or edge-case participants overlap, the system often extends its verification window procedurally and slips into a state of “strict safety—prolonged hesitation.” It has not violated any rule, and may even appear prudent, yet the signal it broadcasts to others is ambiguous: is it yielding, hesitating, or simply failing to comprehend the situation?
For human drivers, such ambiguity forces additional inference cost. Worse, that cost is externalized: the queue behind grows impatient, cross traffic begins to probe forward, pedestrians form divergent interpretations of the vehicle’s intent, and local interaction shifts from “coordination by shared understanding” to “a game of testing and brinkmanship.” In other words, the system may be “correct” in a legal sense, while creating non-coordinability in a social sense.
Road conflicts often arise not from “who violates the rules,” but from “whose intent cannot be read.”
Second: In Merges Where It Could Proceed but Chooses to Wait — Excessive Caution Interpreted as “It Can’t Drive”
In merges, merges onto highways, or roundabout entries, human drivers do not expect a “zero-risk gap.” They expect a negotiable rhythm: you accelerate, I slow; you signal, I confirm; you enter, I yield. This rhythm is not codified in statutes; it is an interactional convention that has evolved over time. Yet rule-driven systems—or systems configured with overly conservative safety thresholds—often refuse to enter that rhythm and instead keep waiting for an almost perfect window.
The result is paradoxical: the system becomes more “correct” in safety logic, yet more “suspect” in social meaning. Human drivers interpret sustained waiting as one of two signals: either the system “cannot judge the situation” (capability uncertainty), or it “will not participate in negotiation” (intent uncertainty). Both forms of uncertainty trigger more aggressive insertions and more forceful cut-ins, which can ultimately increase local risk exposure.
Excessive caution is not neutral; in an interactive system, it is a signal that can be exploited.
Third: “Procedural Pauses” Under Ambiguous Priority — Humans Read the Pause as Uncertainty Itself
Traffic contains many gray zones that rules cannot exhaust: temporary works, non-standard markings, buses pulling out, pedestrian groups crossing, mixed micromobility flows, and even abrupt differences in local yielding norms. In these gray zones, humans often rely on a higher-level mode of resolution: communicating intent through small movements, building agreement through legibility, and advancing the situation via probabilistic judgment rather than deterministic proof.
By contrast, a common autonomous response is “stop and wait until certain.” That may seem reasonable, but in social interaction, pausing is a high-intensity signal: it indicates the system cannot form a stable judgment, and that its next action is difficult to predict. Surrounding participants instinctively widen safety margins, shift strategies, or even adopt “pressuring maneuvers” to force the system to reveal its decision boundary. A single procedural pause can move a local scene from “default cooperation” to “explicit bargaining.”
Public unease is often not caused by rising crash probability, but by uncertainty becoming conspicuous.
The Core Contradiction: In Human-Dominated Traffic, Safety Is Not Equivalent to Compliance
If we attempt to explain these frictions solely through “crash rates,” “violation rates,” or “interventions,” we will arrive at a misleading conclusion: if nothing bad happened, the system must be safe; if it is safe, it should be accepted. The problem is that this metric framework silently assumes that road safety is determined by the quality of rule execution. Yet in a traffic system dominated by human drivers, safety more often emerges not from clause-level compliance, but from stable alignment of group expectations.
In the language of game theory, traffic order is an equilibrium formed through repeated interaction; it relies on common knowledge and a signal–response negotiation mechanism. Human drivers do not execute rules line-by-line; they execute a higher-level behavioral script: how to express intent in a way others can interpret, how to embed one’s actions into others’ expectation distributions, and how to maintain acceptable coordination efficiency under imperfect information.
Accordingly, “safety” in many critical scenarios is closer to a social attribute: am I understandable, am I predictable, and can I be incorporated into others’ decision models? When a system is legally compliant yet fails to satisfy that attribute consistently, it may be “safer” in a statistical sense while appearing “less trustworthy” in lived interaction.
Traffic is not a collection of rules; it is a mechanism for coordinating expectations.
Why This Is a Systemic Challenge, Not Something You Fix by “Tuning Parameters”
One can tune parameters to make a vehicle more assertive or more conservative, and one can add rules to cover more edge cases. But the persistence of the frictions above comes from structural mismatches—not from a threshold being slightly off.
Mismatch One: Law Is Discrete, Static, and Enumerable; Real Interaction Is Continuous, Dynamic, and Not Exhaustible
Traffic laws provide boundary conditions—an enumeration of “must not”—rather than a generative mechanism for “how to negotiate in this situation.” Law constrains worst behavior; it does not define optimal interaction. On real roads, the hardest problems are rarely “is it permitted,” but “can my intent be read.” This is a continuous-time, continuous-space interaction inference problem that cannot be fully represented by a finite rule set.
Mismatch Two: Human Behavior Is a Distribution; Rule Systems Tend Toward Deterministic Trees
Human driving presents strongly distributional behavior: for the same scenario, multiple “acceptable” action paths exist, and choices vary with culture, density, driving style, and even time pressure. Rule-driven systems naturally project the world into deterministic “condition–branch–action” structures. As a consequence, they either pause repeatedly in gray zones or reveal rigidity when rules conflict. Humans detect this rigidity quickly and convert it into a social judgment of “capability uncertainty” or “intent uncertainty.”
Mismatch Three: Humans Are Not Executing Rules; They Are Executing “Expected Behavioral Patterns”
This is the most important—and most frequently overlooked—point. Human driving actions typically serve two functions at once: progressing the vehicle state, and broadcasting intent to others. Many micro-actions (a light brake tap, a slight creep forward, early lane positioning) are not performed to satisfy a statutory clause, but to improve legibility and negotiability. Put differently, humans on the road are executing a “socially legible strategy,” not a set of “audit-friendly actions.”
In a social system, compliance is the floor; legibility is the order.
From this perspective, autonomous driving does not face a “rule coverage” problem, but a deeper engineering objective: social embeddedness. If a system cannot be incorporated into existing human coordination and expectation structures, then even infinite rule refinement will continue to drain trust capital in the real world.
Why Public Unease Often Comes Not from Accidents, but from “Being Hard to Understand”
Public risk perception of autonomous driving is often misread as “fear of accidents.” But in many urban deployment experiences, the more immediate source of unease is that the system’s behavior lacks stable explanatory cues, making each interaction feel like a “fresh relearning.” When you cannot infer intent from another party’s actions, you instinctively widen safety margins; when widening becomes habitual, the traffic system experiences higher friction, lower efficiency, and stronger adversarial probing.
More importantly, this unease is cumulative and contagious. A single crash is a discrete event—often explainable, fixable, and eventually forgotten. But “unreadable interaction” occurs at low intensity and high frequency, and over time it crystallizes into a social narrative: perhaps it does not crash often, but it “feels uncomfortable,” it is “unpredictable,” it “drives like a novice.” Such narratives are difficult to refute with crash statistics, because they correspond to a system property at the experiential layer, not to isolated events.
Trust usually collapses not because of one accident, but because of prolonged accumulation of a “legibility deficit.”
Toward the Methodological Divergence: Why End-to-End Models Have a Structural Advantage in “Social Perceptual Consistency”
If the essence of the problem is social embeddedness, the methodological divergence becomes clearer. Rule-driven systems attempt to approximate reality through finer clauses, richer branching, and more exhaustive scenario enumeration. But reality’s critical difficulty is not enumeration per se; it is learning distributions: how humans form coordination equilibria under imperfect information, how they communicate intent through motion, and how they preserve predictability in gray zones.
The potential advantage of end-to-end models lies precisely here: rather than translating the world into explicit rule trees, they learn from large-scale human driving data the statistical distribution of “how groups typically act in this context.” This makes the system’s outputs more likely to align with interaction conventions that have already been socially validated, thereby facilitating expectation alignment at the social perception layer. In plain terms, the system becomes more likely to make moves that humans “immediately understand,” rather than moves that are “compliant but difficult to negotiate with.”
This does not imply that end-to-end models are inherently safer, nor that they can do without constraints or governance; but they do carry a structural advantage in one critical dimension: they are closer to the road’s real operating logic—roads run on distributions and expectations, not on clauses and branches.
Rule-driven systems pursue “clause correctness”; end-to-end systems have a stronger chance of pursuing “social consistency.”
Conclusion: From “Being More Lawful” to “Being Predictable”
Let us return to the opening question: is autonomous driving more accepted by society the more it “follows the rules”? From an engineering instinct, that is an attractive answer, because it collapses complexity into “the quality of rule execution.” But the road’s real operating logic suggests that this path can lead us into a long-term misconception: when we substitute social legibility with compliance, and substitute social embeddedness with rule coverage, what we are validating is no longer “whether the system can live in the world,” but whether it matches an imagined road model that is overly tidy. What autonomous driving truly needs to solve is not how to become a flawless executor of legal clauses, but how to become a traffic participant that remains understandable, predictable, and negotiable even under uncertainty.
Accordingly, public acceptance is not a problem that can be solved through messaging, education, or “getting people used to it.” It is more akin to a system metric—except that it does not show up as a single crash-rate number; it emerges from countless micro-interactions as social signals. When a system’s behavior is consistently legible, explainable, and negotiable, trust accumulates naturally. When a system is legally impeccable yet repeatedly generates social uncertainty, trust is consumed in a subtler—and far harder to reverse—manner. The reason end-to-end models have a better chance of outperforming in “social perceptual consistency” is not that they are “bolder,” but that they are more likely to learn the group behavior distributions that make traffic work, thereby embedding their behavior into the human expectation structures that already exist.
The ultimate challenge of autonomous driving is not “how to be more law-abiding,” but “how to be readable even when humans are not looking at the rules.”