Catherine M. Sharkey1Catherine M. Sharkey is the Segal Family Professor of Regulatory Law and Policy at New York University School of Law. She thanks Zachary Garrett (NYU School of Law 2023) for providing excellent research assistance.
* * *
A part of the series, Personalized Law.
In Personalized Law: Different Rules for Different People, Professors Omri Ben-Shahar and Ariel Porat imagine a brave new tort world wherein the ubiquitous reasonable person standard is replaced by myriad personalized “reasonable you” commands.2Ben-Shahar’s and Porat’s book reaches far beyond tort law (capitalizing on their collective expertise ranging from contract law, consumer protection law, and remedies to criminal law) to offer a “framework” to evaluate “how far personalization ought to and could be taken.” But my focus here is on the potentially transformative shift for tort (negligence) law, from the imposition of the reasonable person standard, which “insists that individuals be judged according to the standard of an external reasonable actor, representing some aggregate community measure,” to “tailor[ing] [the standard of care] to the specific actor’s tendency to create risks and her ability to reduce them” such that “[t]he reasonable person standard, traditionally derived from an aggregate relevant pool, would be replaced by the ‘reasonable you’ standard—a personalized command that is based on information about this actor’s specific characteristics.” More is said in the book about personalized standards of care than about personalized damages.3This part of the book builds on their prior work on negligence law, where Ben-Shahar and Porat argued: “Relative to a regime of uniform standards, personalization leads to more efficient precaution and has the potential to alleviate the excessive-activity distortion inherent in negligence rules. Currently, negligence law incentivizes actors to become more skilled at harm reduction but not to reduce their riskiness when possible. This latter effect can be tackled . . . if personalization is designed correctly.”
When it comes to damages, Ben-Shahar and Porat seem torn. Corrective justice goals point in the direction of personalized damages, while deterrence goals favor the damage-uniformity approach. In the end, the authors’ assessment of personalized damages is tepid at best, perhaps even antagonistic: “A substantive legal rule that awards uniform damages for loss of life or limb, regardless of one’s income or earnings, has much merit. It would reject the notion that victims differ in a relevant manner, and would require no personalization” (emphasis added).
Ben-Shahar’s and Porat’s asymmetrical embrace of personalized law—full stop for standards of care, near rejection for damages—raises four issues, not sufficiently taken up in the book. First, the authors equivocate too much with regard to the purposes of tort law; ultimately, if and when forced to choose, law-and-economics deterrence-based theory holds the most promise for modern tort law. Second, the damage-uniformity approach clearly dominates the status quo of “crude” personalization. Third, via a deterrence lens that eschews “misalignments” in tort law, a personalized standard of care necessitates personalized damages. Fourth, the true benefit of an ideal personalized damages regime might be further uncovering the root cause of racial and gender disparities in status quo tort damages. Paradoxically, ideal personalization might then reinforce the damage-uniformity approach.
I. Deterrence as Modern Tort Law’s Primary Goal
Ben-Shahar and Porat make an overarching claim that “compared to uniform law, personalized law promotes more effectively the law’s underlying goals. Any goals, of any law. . . . Across all of law, personalized rules could achieve better results with less adjudicative error, ultimately improving access to courts and justice.” As applied to the two primary goals of tort law, “[i]f tort law seeks to deter unsafe behavior, personalized commands would reduce both accidents and prevention costs,” and “[i]f its goal is to compensate victims according [to] a specific make-whole principle, personalized rules would guarantee more accuracy in the assessment of damages.”
But this Panglossian view—highlighting the advantage of personalized standards of care through a deterrence lens and the advantage of personalized damages through a corrective justice lens—obscures the tension that arises upon consistent application of a deterrence lens to both standards of care and damages. Ben-Shahar and Porat only briefly address this paradox, noting that “[t]he necessity of well-specified goals is a major hurdle for personalized law[ ] but also one of its blessings[,]”and conceding that, while “[i]t is hard to specify goals upfront, in the abstract[,] . . . it is also important to write candidly, without obscuring the objectives.”
Taking this to heart, I wish that the authors had adamantly embraced an efficiency/deterrence perspective without hedging or equivocating at critical junctures with gestures to corrective justice. To my mind (as I argued in Modern Tort Law: Preventing Harms, Not Recognizing Wrongs), law-and-economics deterrence-based theory holds the most promise for judges facing the primary challenges posed by modern tort law. Hence, I will proceed to hold up an exclusive deterrence lens to Ben-Shahar and Porat’s consideration of personalized damages.
II. In Favor of Damage-Uniformity Over Crude Personalization
Tort damages are individuated. A jury’s or judge’s assessment of damages turns on particular plaintiffs’ harms or losses suffered. Factors like income level and socioeconomic status that can vary greatly from individual to individual play a determinative role in the calculation of both economic and noneconomic awards in tort suits.
But such individuated damages constitute “crude personalization” in the sense that “[t]ort law, when awarding damages for income-reducing personal injuries and loss of life, often uses charts and tables to calculate victims’ projected lifetime incomes. These compensation charts are meant to be personalized, but only in a very coarse and problematic manner, distinguishing people primarily by age, gender, and race.”
This conventional use of race- and gender-specific economic data in calculating tort damages leads to unjustifiable and discriminatory discrepancies in the amounts awarded, particularly to young women and minorities whose expected future wages rely considerably on race- and gender-based tables. Such crude personalization fuels a racialized and gendered deterrence gap by disincentivizing actors from taking precautions against harms to blacks and women.
Uniformity presents itself as a solution. In Valuing Black and Female Lives: A Proposal for Incorporating Agency VSL into Tort Damages, I advocated for the adoption of a uniform “value of statistical life” (VSL) as a measure of tort wrongful-death damages to eliminate the perverse incentives for defendants to channel their most risk-laden behavior toward minority communities. Citing this work, the Restatement of the Law Third Torts: Remedies calls for “an end to race and sex discrimination in the measurement of tort damages” by admonishing that “[c]ourts should not allow expert testimony or other evidence that a plaintiff’s earning capacity is higher or lower because of the plaintiff’s race, ethnicity, or gender.”4Restatement of the Law Third Torts: Remedies § 19 (“Lost Earnings and Earning Capacity”) cmt. e & Reporters’ Notes cmt. e.
Here, Ben-Shahar and Porat seem to agree that damage-uniformity is warranted: “it is important that deterrence should be spread equally across communities, not diminished in poor communities where poorer people collect lower damages and thus may be regarded as more ‘affordable’ victims for potential tortfeasors.” But surely more must be said about both the feasibility and desirability of a “personalized law” regime that combines personalized standards of care with damage-uniformity—especially in light of the efficiency considerations for personalization across the board.
III. Personalized Standard of Care and Uniform Damages: A New Misalignment?
A paradox emerges: Personalized Law apparently argues for taking the status quo—uniform standards of care coupled with (crudely) personalized damages—and flipping it to produce a system of personalized standards of care with uniform damages. But this seems like replacing one “misalignment” with another, which is all the more perplexing given that, in prior work, Porat coined the phrase “misalignment” in tort and even explored as a prime example “the misalignment that results from lost income being a major component in damages awards for bodily injury.”
The “alignment principle” in negligence law posits that “the risks taken into account by courts when setting the standard of care are the same risks considered when imposing liability and awarding damages.” By way of the Hand Formula, “to create efficient incentives under all circumstances, the rule of negligence is that as long as B < PL, the negligent injurer who failed to take precautions will bear liability for the entire harm. . . . Negligence law thus aligns the standard of care with compensable harms.” Misalignments occur when there is a disconnect between the notion of a duty of care owed by a defendant to all persons (regardless of individual characteristics, including demographics) and defendants’ incentives to take care only to the point at which the marginal expected cost of precautions (B) equals the marginal expected damages (P*L). Further, as Porat explained:
This misalignment is acute in all those cases where the injurer could know in advance that his average potential victim’s income is different from the average income. Such cases are not rare. Doctors can easily learn whether their patients are high- or low-income; employers can distinguish between high- and low-income employees; occupiers of land often know the level of income of the invitees on their premises; and polluters are often aware of the income of their potential victims.
Porat made a compelling case that, from an efficiency perspective, “both the standard of care and damages should be identical for all victims, and the misalignment between the two under prevailing law would be inefficient.” Moreover, “if [the lives and limbs of high-income and low-income victims] are of equal value, there should be identical standards of care and damages imposed for both groups.”
With characteristic clarity, Porat concluded:
If courts want to comply with the alignment principle, they should choose to either: (a) apply different standards of care to high-income and low-income victims (contrary to what they actually do), coupled with different levels of compensation (as they actually do); or (b) apply the same standard of care to high-income and low-income victims (as they actually do), coupled with the same level of compensation (contrary to what they actually do).
Fast forward now to Personalized Law where, at the juncture where Ben-Shahar and Porat embrace personalized standards of care coupled with the damage-uniformity approach, a warning bell should sound: “for proponents of efficiency, minimization of social costs, or more generally the promotion of social welfare as the sole goal of tort law, misalignment should ring a warning bell that the law is probably inefficient and should be modified.”
IV. Ideal Personalization and the Promise of AI
Even if we assume that (1) efficiency/deterrence is the best lens for torts, (2) damage-uniformity dominates “crude personalization,” and (3) the alignment principle dictates that any personalization choice should be symmetric across standards of care and damages, a final question arises: (4) What about the prospects for “ideal personalization” at both ends?
Ben-Shahar and Porat make a foolproof case that with big data, “the process whereby computers sift through enormous quantities of data to identify patterns that can predict individuals’ future behavior,” and Artificial Intelligence (AI)5With regard to AI-driven technologies: “[We have a] notion that [personalized law] would come from machine-sorted information—training algorithm[s] to sort through Big Data about people and identify personal features that are relevant to the goals of the law. . . . Kind of in [a] similar way and maybe even more advanced than what insurance companies do right now when they try to assess a driver’s risks or someone’s life expectancy for the purpose of life insurance. Once we have this data, we identify a formula—that’s what the training algorithm does—and then the screening algorithm just issues personalized commands based on individual attributes.” an information revolution is at hand, with significant implications for tort law. Indeed, as they recount, “the key to successful personalization—to any successful tailoring—is information.” And, specifically with regard to damages, “[a] primary limitation on personalized damages, both practical and principled, is lack of information.”
Ben-Shahar and Porat are on solid ground in critiquing the status quo crude personalization of damages: “In a word of small data, these remedial rules do not account for enough factors to accurately predict people’s actual losses.” They also give reason for confidence that “the technology to roll [personalized law] out in some legal domains may be at our doorstep.” But when it comes to envisioning the brave new world of (ideally) personalized damages, the picture looks sketchy. According to Ben-Shahar and Porat, “[a] fully personalized damages regime would employ a multitude of factors to measure each victim’s actual loss” (emphasis added). Not only a bit vague, but not quite revolutionary, as Ben-Shahar and Porat concede: “This is not very different from what tort law currently does, especially when damages are set to equal lost future income calculated on the basis of socioeconomic statistical data.” The role for big data is relegated to its potential to “refine and improve the implementation of this methodology, hopefully neutralizing some of its objectionable reliance on a handful of suspect classifications.”6Paradoxically, as Ben-Shahar noted, reliance on suspect classifications such as race and gender is the way out of a separate moral hazard personalization conundrum: “A problem with personalized law . . . is that if we increase the duties of care to people who have more skill in taking care, that will undermine their incentive, their motivation to improve their own skill. . . . This is a general problem with personalized law. . . . What we can do, therefore, is for personalized law to rely on personal features that people cannot manipulate, like inherent traits: what’s your race, how many years of experience you have in something, or a proxy.” And perhaps that is why, in the book, Ben-Shahar and Porat somewhat equivocate: “When a factor like sex or race is used, it must not be ‘decisive by itself’ nor used in a ‘nonindividualized, mechanical’ way. It could permissibly be ‘one factor weighed with others in reaching a decision’ so as to ‘provide for a meaningful individualized review.’”
The guiding principle—to be treated as an individual as opposed to an average member of a group (especially a “suspect category” group)—is clear:
Predictive algorithms that rely on Big Data could allow for personalized compensation without this weighty reliance on constitutionally suspect classifications like race and gender. People would not be treated as members of a group, but rather as unique specimens, each characterized by dozens (if not more) of characteristics that are correlated with lifetime income. Rather than constituting decisive factors, race and gender could be relegated to a marginal role. If desired, information about race and gender could be eliminated entirely.
But the specification of these “dozens (if not more) of characteristics” (or “multitude of factors”) and how, as a normative matter, they relate to measuring harm or losses (especially in the most difficult cases of victims with no individual earnings track record where the use of race- and gender-based tables are most prevalent) is less pellucid.
Nonetheless, Ben-Shahar and Porat provocatively point us in a different direction. They are upfront about the fact that any big-data–driven “algorithm will be relentless at assuring that plaintiffs receive awards closely reflecting their true losses, which unfortunately do vary by membership in protected classes.” For example, as they lament, “[i]f members of a particular racial minority receive, on average, less than what members of other groups collect, it is not the result of biases in the law of remedies, but rather of its accuracy—it reflects the differential earnings in our society.”
The status quo “crude” personalization of damages, making use of race- and gender-based tables, reifies and perpetuates structural inequities. To the extent that the lower wage estimates for members of minority racial/ethnic groups and women reflect, at least to some degree, structural inequities and historical injustices, these are incorporated into tort damage estimates.
Racial and gender differences might arise from two very different sources. First, preferences with respect to risk might vary (as with age and income level) by race. Second, there are fundamental differences in labor market opportunities between blacks and whites and men and women. Indeed, there exists a “substantial literature on market discrimination [that] has documented that there is a persistent difference in the earnings of whites and blacks even after controlling for a broad set of individual characteristics and job characteristics.” As Professor Kip Viscusi has explained: “Although differences in preferences could be influential, such differences cannot reconcile the various empirical findings. Rather, there must also be fundamental differences in labor market opportunities for blacks and whites as well as in the structure of their offers for risky jobs.”
Two dimensions of the more revolutionary potential of big-data–driven approaches thus come into focus: first, the enhanced ability to distinguish between competing factors driving race and gender disparities in the damages realm, and second, the superior predictive abilities that inhere in a dynamic, learning-based model.
With regard to the first dimension, in a study of the use of AI to unpack racial disparities in predicted credit default rates among white, black, and Hispanic borrowers in the United States, Talia Gillis and Jann Speiss raise the potential comparative advantage of algorithmic decision-making over human decision-making:
Unlike the human decision-making context in which many aspects of the decision remain highly opaque—sometimes even to the decisionmakers themselves—in the context of algorithmic decision-making, we can observe many aspects of the decision and therefore scrutinize these decisions to a greater extent. The decision process that led to a certain outcome can theoretically be recovered in the context of algorithmic decision-making, providing for potential transparency that is not possible with human decision-making.
At the same time, Gillis and Speiss are upfront that formidable challenges remain. Not only is it that, “in the context of machine-learning prediction algorithms, the contribution of individual variables is often hard to assess,” especially in cases where “data is high dimensional,” but also “the exclusion of [a] forbidden input alone may be insufficient when there are other characteristics that are correlated with the forbidden input—an issue that is exacerbated in the context of big data.” Indeed, in their study, Gillis and Speiss found that “even excluding . . . variables that are correlated with race has limited effect in big data.”7“Despite significantly reducing the number of variables that correlate strongly with race, the disparity still persist[ed]. . . .” Correlative inputs of this sort are congruent with what Ben-Shahar and Porat describe as “substitute factors that would operate as proxies” for suspect classifications under a regime of personalized law. Ben-Shahar and Porat elaborate: “[I]n many contexts the law might prohibit the use of suspect classifications, most importantly race, as a factor for differential treatment. No matter how narrowly tailored the treatment . . . it will often be prohibited to take race into account as a treatment factor. Stripped of the use of suspect classifications, personalized law would likely identify substitute factors that would operate as proxies.” And, “[i]f race is not permissible, it could be substituted with zip code.”
With regard to the second dimension, the potential improvements from the status quo can hardly be gainsaid given the superior predictive abilities that would inhere in a dynamic, learning-based model. The race- and gender-based tables currently in use in the calculation of tort damages are inaccurate predictors of the future. The table figures are static, meaning they reflect a snapshot of the time period from which the data were collected without accounting for trends that were evolving over that time period. The use of static figures rests on the assumption that whatever discrepancies exist between demographic groups will not only persist but also remain constant into the future. Specifically, such figures do not capture the emerging legal, professional, and social norms that favor workplace equality and increase the likelihood that discrepancies in employment data between demographic groups will diminish in the future.
* * *
Catherine M. Sharkey is the Segal Family Professor of Regulatory Law and Policy at New York University School of Law. She thanks Zachary Garrett (NYU School of Law 2023) for providing excellent research assistance.
* * *