TABLE OF CONTENTS INTRODUCTION 2373 I LUMPING LAPSES 2381 A. Lapses and Accidents 2381 B. Distorted Decision-Making 2384 C. A Disaggregating Alternative 2390 II. STACKING CARE 2397 A. Marginal Analysis and Causation 2397 B. Unifying Untaken Precautions 2399 1. Negligent Defendants and Efficient Accidents 2400 2. Cliffs and Ledges 2403 C. Strategic Slicing 2414 III. REPEATING RISKS 2422 A. Compounding and Offsetting 2422 B. The Bridge of Due Care 2427 C. Matching Up Accidents 2433 IV. EXTENSIONS AND CONNECTIONS 2437 A. Setting the Viewfinder 2438 B. Bundling and Strategizing 2439 C. Consequential Line-Crossings 2441 CONCLUSION 2444 INTRODUCTION
Tort law deals in lumps. It responds not to the innumerable finegrained acts of risk creation that each of us performs every day, but rather to large, discrete, harmful events--"accidents." (1) And it responds to those events in a binary way, converting unruly facts into an on/off judgment about liability. (2) The fact that tort law operates at the accident level rather than at the level of risk creation presents some complications, at least if we understand tort liability as significantly directed at providing appropriate incentives for action. (3) The accident, on this view, serves as a window into risk-creating (and risk-abating) behavior, and liability represents a rough-and-ready way of addressing that behavior--a kind of accounting shortcut that focuses on a realization event. How then should law isolate and evaluate the sample of risk-related behavior connected to the accident?
This Article provides fresh traction on this foundational question by examining the underappreciated role of evaluative aggregation in the liability determination. I focus on three aggregative choices: (1) how much behavior to compile for purposes of assessing due care; (4) (2) how to stack together units of precaution in examining a defendant's shortfall; (5) and (3) how to factor the actual or imagined repetition of an interaction into liability judgments. (6) These aggregative choices carry decisive weight by determining how large a slice of an injurer's conduct tort law will capture within its viewfinder, and how tight the causal connection must be between the shortfalls observed there and the accident at hand.
Consider, for example, a point emphasized in work on "lapses" (7) or "compliance errors" (8): It is impossible for human beings to be perfectly consistent in taking precautions that must be repeated over and over in real time, such as alertly scanning the road while driving. (9) A single moment of inattention that produces an accident might be part of a larger pattern that represents as much care as any person could reasonably be expected to exercise. (10) Alternatively, the accident-causing shortfall might be a representative draw from an urn of chronically unreasonable conduct. Tort law does not distinguish between these cases because it takes as the relevant unit of analysis the single accident-causing moment, not a larger behavioral sample that might provide corroborating or mitigating evidence about the actor's overall level of care. (11) This does not necessarily mean that tort law should look at broader behavioral patterns--indeed, there are reasons to question that prescription--but it does mean that the choice to focus on a single moment in assessing negligence carries consequences.
The consequences mount when we consider the possibility that certain durable technologies or mechanical processes might either substitute for or complement fast-eroding human precautions like paying attention while driving. (12) The emergence of autonomous vehicles represents an especially salient current example. A machine with a known error rate will seem to naturally invite an aggregate analysis, with attention focusing on whether the overall pattern of outcomes could have been cost-effectively improved. If not, then using the machine may appear nonnegligent, even though the machine might hiccup now and then and cause an accident, just as a generally cautious human might lapse. If the two cases are treated differently, certain technologies might be overused or underused depending on the degree to which their deployment demands sustained human attention. (13) It is possible to address this problem without widening tort law's behavioral viewfinder, but doing so requires recognizing the unsung role of aggregation choices in shaping liability, and hence incentives. (14)
Another margin for evaluative aggregation in tort law involves stacking together or breaking apart a set of precautionary steps that the actor in question chose to forgo. Suppose, for example, a cricket ball sails out of an unfenced cricket field at an altitude of seven feet ten inches and brains a pedestrian. (15) At trial, the pedestrian's estate shows that putting up an eight-foot fence--one high enough to have prevented this accident--would have cost only $1000 but would have saved an expected $1200 in accident costs over the useful life of the fence. This might look like an open-and-shut win for the plaintiff, at least if the jurisdiction follows the cost-benefit approach to negligence captured in the Hand formula. (16)
But not so fast. There is nothing inevitable about treating the full eight feet of fencing as an indivisible unit when analyzing the cricket club's behavior. Once we disaggregate that single, lumpy, all-or-nothing choice into incremental choices about fence heights, the simple case starts to look less airtight. Suppose the first six feet of fence height are really worthwhile, delivering $1100 in accident savings while costing only $800 in lumber and labor, but the last two feet require an extra $200 in construction costs only to save a marginal $100 in accident costs. (17) Even though it was negligent for the cricket club not to build a six-foot fence, it would not have been negligent for the club to stop at six feet--and a six-foot fence would not have stopped the fateful ball. On this account, the club's negligence (its failure to build the optimal six-foot fence) did not cause the accident. The correct doctrinal result would seem to be no liability. (18)
Yet again, not so fast. We must still consider how these aggregation decisions interact with an important architectural feature of tort law--the fact that liability falls to zero at the point of due care under a negligence standard. (19) One implication of that architecture has been well recognized: the possibility that a defendant would face a behavior-distorting cliff of liability if held to account not only for the harm that occurs because he is negligent but also for all of "the harm that occurs when he is negligent." (20) But there is another implication of the negligence regime that has been widely ignored, though it also bears on whether the cricket club should be let off the hook. The law plunks a flat ledge of zero liability across the entire range of conduct falling beyond due care--a range of conduct that, in fact, generates different real-world accident rates. (21) This ledge also has distortive effects, and these distortions will ultimately require us to reverse our earlier intuition about whether to analytically disaggregate the unbuilt cricket fence. (22) The precaution-aggregation choice turns out to be a crucial lever for optimally adjusting the required causal relationship between negligence and harm. (23)
Finally, tort law contains some important puzzles that only become visible when a particular interaction is repeated many times (whether as a matter of fact or as a conceptual exercise). Deciding whether and how to "scale up" the liability analysis constitutes another domain for implicit aggregation choices. Repetition can reveal distortions and injustices that are muted at the individual-accident level--or it can do the opposite, washing out apparent a-nomalies. (24) Because statistical risks and expected payoffs become more meaningful and tractable under large-number conditions, results that appear intolerable at close range--a large chunk of liability for a relatively trivial act of negligence, say--may look more acceptable once we zoom out to capture a larger set of similar interactions. (25) Here, it becomes important to consider whether insurance or specialized doctrines can synthetically replicate large-number conditions for individuals. (26)
Repetition may instead compound rather than counterbalance systemic shortfalls in liability patterns. Consider the negligence of doctors in subspecialties where patients routinely face high background risks of death. A doctor who negligently treats a population of patients who are overwhelmingly likely to die in any event will always be let off the hook by a more-likely-than-not standard, even if her negligence caused, say, twenty out of one hundred observed deaths. Although scholars have proposed various approaches to this well-recognized problem, (27) one underappreciated alternative deserves attention: instead of asking whether a given harm was more likely than not caused by the doctor's negligence, ask whether the harm was more likely to have been caused by the doctor's negligence than the other harms in the conceptual set generated by repetition of the interaction.
Here, we can take a page from "thresholding" in image manipulation. (28) When a grayscale image is converted into black and white, a continuous variable (shading) must be translated, pixel by pixel, into binary results. (29) A globally applied threshold will produce unacceptable results where, for example, light or shadow falls across a portion of the image, making all pixels in a given region darker or lighter than the threshold that works best elsewhere in the image. (30) The key to successfully picking out foreground from background is to see which pixels are local standouts--hence, thresholding methods examine shapes, clusters, and pixel neighborhoods in order to determine the appropriate local...