The White House Office of Management and Budget (OMB) recently released a draft memorandum outlining new requirements for federal agencies using artificial intelligence (AI). OMB’s goal with the memo is to ensure agencies responsibly govern and manage risks from AI systems, especially where AI impacts individual rights and safety.
The memo—which came in response to President Biden’s 2023 AI executive order, as well as several recently-enacted laws—calls on agencies to appoint chief AI officers, develop AI strategies, and implement minimum practices for rights-impacting and safety-impacting AI systems by August 2024. These minimum practices include conducting AI impact assessments, which are evaluations conducted to assess the consequences of deploying AI technologies in various contexts.
While the memorandum is a good first step, federal agencies can improve their approach to AI safety by formally integrating quantitative risk analysis methods into their AI impact assessments. Although one might not normally expect new requirements related to AI to result in mortality, the reality is that any regulation or policy that imposes costs affects mortality risk. Therefore, risk analysis has a role to play in AI safety efforts.
As I explain in a recent book chapter coauthored with Salisbury University Professor Dustin Chambers, risk analysis involves estimating the magnitude of risks stemming from policy actions, so policymakers can direct limited public resources toward reducing the most harm.
Risk vs. risk
There are often risk-risk tradeoffs that result from regulations. Just like taking an aspirin reduces the risk of a headache but can increase the risk of stomach problems, when one risk is reduced another often increases. Preemptively identifying unintended consequences as part of careful regulatory analysis is part and parcel of good governance.
An example of a risk-risk tradeoff transpires when incomes are reduced to pay for regulations. This limits the ability of employees, customers, and shareholders to mitigate risks privately. Mortality risk analysis compares a policy or regulation’s direct risk reductions to these countervailing risk increases stemming from income losses and other sources.
The analysis may sound complicated, but it is fairly simple to conduct. To start, agencies can begin by estimating the cost-effectiveness of their policies. For example, a $500 million regulation that saves 10 lives has a cost-effectiveness of $50 million per life expected to be saved.
The “value of an induced death” (VOID) concept is a threshold level at which income losses are sufficient to cause one death in the population. At lower income levels, people might forgo home security systems, healthier lifestyles or other amenities that mitigate their risk of death.
While estimates in the literature vary, $75 million is probably in the vicinity of a reliable estimate of the VOID. If a regulation’s cost-per-life saved exceeds this level, the regulation or policy can be expected kill more people than it saves.
Complicating matters, both the costs of regulation and the VOID are changing over time, so it is possible for a regulation to reduce risk immediately, but to increase risk at some point in the future (as well as vice versa—increase risk today but reduce it at some later point).
In order to conduct mortality risk analysis properly, an analyst must determine the opportunity cost of the action and how it is evolving over time. Government analysts usually estimate a regulation’s compliance costs, i.e., what is spent. Typically, they stop there. But these expenditures represent accounting costs, not economic opportunity costs.
To convert accounting costs into opportunity costs, the analyst must consider what the most likely use of funds would have been in the absence of the regulation. At the most basic level, funds could either have been put toward investment and capital expenditures or they could have been used for consumption instead.
Just like how the marginal dollar an employee earns often goes into her 401(k) account, it’s reasonable to assume most of what is spent on compliance would have been invested. Analysts must also identify a “hurdle rate,” or rate of return, that displaced investments likely would have earned over time.
I’ve argued in the past that a hurdle rate of 5 to 6 percent is reasonable, though arguably higher hurdle rates should be applied to government projects, in part because government policies tend to be irreversible once embarked upon.
With all this background in mind, here is a simple example of how mortality analysis works:
Let’s say a regulation imposes costs of $1 billion dollars today in order to save 50 lives after five years. Perhaps the benefits take a while to achieve because implementation of the rule takes time, but the expenditures made by businesses on compliance happen more or less immediately.
At a VOID of $75 million, the regulation is expected to result in about 13 deaths. Although the exact timing of these deaths will vary, the saved lives don’t arrive for five years, so the regulation is expected to increase risk more or less immediately. This situation reverses, however, after five years, when the 50 lives are expected to be saved. Now mortality risk is on balance reduced.
But this is not the end of the story. Next, let’s say 75 percent, or $750 million, of the $1 billion accounting cost comes in the form of displaced investment, while $250 million represents displaced consumption. The $750 million in displaced investment would have been growing at some rate of return over time. Let’s say the hurdle rate for projects is 7 percent. Moreover, let’s say as incomes across society rise, the VOID also grows at 2 percent a year.
Since the regulation only saves 50 lives on a one-time basis, there is a subsequent risk reversal after about 10 years due to the growing opportunity costs. At this point, the regulation’s costs are sufficient to induce deaths that fully offset the 50 lives the regulation is expected to save. Thus, this regulation increases mortality risk immediately, reduces risk during the period from about 5 to 10 years after enactment, and then increases risk permanently thereafter.
Obviously, the characteristics of specific regulations will vary, and the timing of both saved and lost lives is difficult to predict. Some regulations save lives on an ongoing basis every year, while many AI regulations never save any lives at all because that is not their intended purpose. Even so, those regulations still impose costs that will have corresponding mortality opportunity costs to account for.
Regulatory agencies have long neglected the opportunity costs of their rules, including health opportunity costs. OMB’s memo on AI safety, along with the AI impact assessments likely to be carried out in response to the memo, presents a perfect opportunity to change course. It is high time federal agencies start accounting for the mortality costs of their regulations.