aises_6_4

<style>
    .visionbox{
    border-radius: 15px;
    border: 2px solid #3585d4;
    background-color: #ebf3fb;
    text-align: left;
    padding: 10px;
    }
</style>

<style>
    .visionboxlegend{
    border-bottom-style: solid;
    border-bottom-color: #3585d4;
    border-bottom-width: 0px;
    margin-left: -12px;
    margin-right: -12px; margin-top: -13px;
    padding: 0.1em 1em; color: #ffffff;
    background-color: #3585d4;
    border-radius: 15px 15px 0px 0px;
    }
</style>

<style type="text/css">
    table.tableLayout{
        margin: auto;
        border: 1px solid;
        border-collapse: collapse;
        border-spacing: 1px;
    }

    table.tableLayout tr{
        border: 1px solid;
        border-collapse: collapse;
        padding: 5px;
    }

    table.tableLayout th{
        border: 1px solid;
        border-collapse: collapse;
        padding: 3px;
    }

    table.tableLayout td{
        border: 1px solid;
        padding: 5px;
    }
</style>

<h1 id="sec:economic-engine">6.4 The Economic Engine</h1>
<strong>What
if we allow the economy to decide what AIs will be like?</strong>
Unlike some prior technological breakthroughs (such as the development of nuclear energy and nuclear weapons), most investment in AI today is coming from businesses. Leading commercial AI developers have acquired the enormous computational resources required to train state-of-the-art systems and have hired many of the world's best researchers. We could therefore argue that AI development is most closely aligned with business or economic goals such as wealth maximization. Many believe this is good, arguing that the development of AIs should be guided by market forces. If AI can accelerate economic growth, provided we can also ensure a fair distribution of costs and benefits of AI across society, this could be positive for people's welfare around the globe. We examine the specific impacts of AI on economic growth and its distributional effects in the Governance chapter.</p>

Here, we consider the broader attractions and limitations of allowing economic growth to be the main force determining how AI systems are developed. As part of this discussion, we will briefly introduce and explain a few basic concepts from economics, such as market externalities, which provide an essential foundation for our analysis. We argue that while economic incentives can be powerful forces for prosperity and innovation, they do not adequately capture many important values. AI systems that are primarily created in order to maximise growth and profit for their developers could have a range of harmful side-effects. This highlights the need to better capture different facets of individual and societal wellbeing within the AI development process. Alternative approaches are discussed further in the following sections.</p>

<h2 id="the-free-market">6.4.1 Allocative Efficiency of Free Markets</h2>
<p>Competitive economic markets can drive efficiency and foster collective prosperity. Historically, market forces have led to specialization at a vast scale and worldwide competition, permitting the production of better goods and services at lower prices. The global economy is a complex system that allows for mass coordination. Prices, for instance, help producers and consumers find efficient trading equilibria in the face of dynamic market conditions. One can speculate that the integration of AI into the economic engine could further optimize production and enhance competition, ultimately contributing to accelerated economic growth and improved wellbeing.</p>

<strong>Under the right conditions, the free market can also create allocative efficiency.</strong>
The First Fundamental Theorem of Welfare Economics states
that—subject to some strong assumptions—an equilibrium
<em>allocation</em> of goods reached by trading on a free market must be
<em>Pareto efficient</em>. An outcome is Pareto efficient if there is no
way to make anyone better off without making someone else worse off: any
change must trade off one person’s welfare against another. If the
allocation of goods were not Pareto efficient, then individuals within
it would trade in a way that exploited the possible Pareto improvement,
reaching a state in which no mutually beneficial trades can be
made.</p>
<p>This supports Adam Smith’s famous “Invisible Hand” argument, which
suggests that when individuals pursue their self-interest within a
market, they unintentionally contribute to societal welfare. By creating
gains from trade, there is an overall improvement in living standards
for everyone.</p>
<strong>There
are many conditions required for the invisible hand argument to
hold.</strong>
Without the following conditions, the invisible hand argument—which
holds that the presence of free markets with individuals acting in
self-interest can create outcomes that are Pareto efficient—would
fail.</p>
<ol>
<li><p><strong>There must be an open market.</strong> There should be no
barriers to entry for producers or buyers, so that everyone can
participate in this market. This openness stimulates competition,
promoting economic efficiency.</p></li>
<li><p><strong>There should be no seller big enough to move prices up
alone.</strong> If there are many sellers, then anyone who raises prices
will lose consumers. There must be no monopoly power: anyone with the
ability to raise prices without being forced down by competition will
create distortions that leave consumers worse off.</p></li>
<li><p><strong>No producer should privately hold a pivotal
technology.</strong> This means that other producers should be able to
copy the production of the first-mover. While the first-mover will make
profits in the short run, in the long run the market allocation will be
Pareto optimal.</p></li>
<li><p><strong>No buyer should be big enough to move prices down
alone.</strong> Similar to the condition for producers, no buyer should
be big enough to force producers to take lower prices than others would
offer for them. Such buyers would create distortions that might, for
instance, force producers out of business, ultimately harming
everyone.</p></li>
<li><p><strong>There must be perfect information for everyone.</strong>
Producers and consumers must have access to perfect information, such as
about product quality and pricing. If consumers don’t know, for
instance, that a seller’s product is defective or that other sellers are
offering lower prices, then markets cannot achieve efficiency.</p></li>
<li><p><strong>Preferences must be non-satiated.</strong> The first
welfare theorem does not hold for all types of preferences; one
technical restriction is that consumers should always prefer more of a
good. If someone gains no further value from extra consumption, their
trading equilibria may not be Pareto efficient.</p></li>
<li><p><strong>There must be no externalities in consumption.</strong>
When one consumes a good, there must be no effect—positive or
negative—on anyone else. Second-hand smoke has a negative externality on
anyone nearby but this is unaccounted for in the price of cigarettes,
creating inefficiency.</p></li>
<li><p><strong>The state enforces property and contract laws.</strong>
Most economic theory assumes that contracts are enforceable, and that
individuals and corporations have protected property rights. Without
these, trading would be difficult to achieve and much more costly,
making it more difficult for everyone to achieve optimal
outcomes.</p></li>
</ol>
<p>Given the stringency of these conditions, it is obvious that they will not always hold in practice and that there may be market failures. Unregulated markets do not always create efficient outcomes: instead, unregulated markets often see informational asymmetries, market concentration, and externalities. Unfortunately, AIs may exacerbate these market failures and increase income and wealth inequality, creating disproportionate gains for the wealthy individuals and firms that own these systems while decreasing job opportunities for others.</p>

<h2 id="market-failure">6.4.2 Market Failure</h2>
<p>In this section, we consider three common types of market failures
that are especially relevant in the context of AI: <em>information
asymmetries</em>, <em>monopoly power</em>, and
<em>externalities</em>. We also discuss the idea of <em>moral hazard</em>, which can be applied to the development of AI systems which can generate profit for their owners but could prove harmful to others.</p>
<h3 id="information-asymmetries">Information Asymmetries</h3>
<strong>Information
known to only some can create market failures.</strong>
Information asymmetry captures the idea that buyers and sellers have
different information regarding the product they are trading. For
instance, buyers are keenly aware of the product’s quality and
specifications, and sellers know their true willingness to pay for the
product. Information asymmetry isn’t inherently problematic and is, in
fact, often a positive aspect of market dynamics. We trust specialists
to provide valuable services in their respective fields; for instance,
we rely on our mechanics to know more about our car’s inner workings
than we do.</p>
<p>However, issues arise when an imbalance of information is exploited
disingenuously. A classic example of this can be found in the used car
market. A dealer may be aware that a car’s axle is rapidly wearing out,
a defect that isn’t immediately noticeable to a potential buyer. By
withholding this information, the dealer could sell the car at a price
higher than its true value. The buyer, left in the dark, may end up
facing unexpected repair costs soon after purchase. This is also
referred to as <em>adverse selection</em>: when one party in a
transaction uses their access to private information to their advantage.
Here, the car dealer is using their private knowledge of the car’s
condition to make a sale at an unfair profit.</p>
<strong>AIs may exacerbate informational asymmetries.</strong>
AI holds the power to both create and exploit information asymmetries in unprecedented ways. This capacity can be employed in beneficial ways, like providing highly personalized services or excellent recommendations. However, it can also be misused, leading to situations in which those using AIs can manipulate consumers. AI-powered analytics allows companies to create sophisticated profiles of consumers that can uncover deep insights into individual personalities and behaviors. <span
class="citation" data-cites="rust2005psychometrics">[12]</span> Big tech companies already use social media and device activity to understand an individual's preferences and vulnerabilities better than ever before. This knowledge can be used to shape targeted advertisements or manipulations that are more effective. While these are not market failures in the technical sense, they can lead to exploitation and other undesirable outcomes.</p>
<p>Though AI may magnify the potential for information asymmetries, such
strategies have long standing precedents in non-AI contexts. Predatory
lending is a common practice where lenders, often equipped with more
information than borrowers, use deceptive practices to encourage
individuals into accepting unfair loan terms. These tactics tend to
target lower-income and less-educated individuals who might not have the
time or background to understand the fine-print of what they’re signing,
or the resources to find legal counsel. AI can further increase the
power imbalance; for instance, AI can be used to predict who is most
likely to accept these unfair loan terms based on their digital
behavior, leading to even more targeted predatory lending. AI can both
amplify existing issues and present new challenges.</p>

<h3 id="oligopolies">Oligopolies</h3>
<strong>Markets can tend towards concentration, leaving consumers worse off.</strong>
While competition can create productive efficiency, some firms can avoid competitive pressures; for instance, utilities that are the exclusive provider of power or gas in a region might be able to raise their prices without losing many consumers to competitors due to lack of alternatives. When a single company (monopoly) or a small group of companies (oligopoly) has a high level of control over a market, consumers are often left with limited product options and high prices. To preserve the benefits of competition, governments implement regulations such as antitrust laws, which they can use to limit market concentration by preventing mergers between companies that would give them excessive market power.</p>
<strong>High
initial investment requirements can impede newcomers from entering the
market.</strong>
Historically, first-movers in capital-intensive industries have a
competitive advantage, such as rail companies that own large quantities
of railway networks. In such industries, it is difficult for other firms
to enter the market due to high up-front costs; in many cases, an
existing firm can block other firms from entering, such as by keeping
prices below a sustainable level. AI development might be similar—at the
very least, firms within it might have some monopoly power that allows
them to avoid many competitive pressures.</p>
Developing a large model requires paying substantial fixed costs
up-front, including expenses for computing power and datasets essential
for training. However, once these initial investments are made, the
subsequent cost per user for deploying and maintaining these models is
considerably lower. Since the average cost per unit decreases as the
number of customers increases, it becomes substantially more
cost-effective for a single AI company to provide access to many people
than for multiple companies to independently develop and maintain
similar models.</p>
<p>Additionally, only a few companies might have access to enough resources
and data to create the best AI models. In the AI Fundamentals chapter, we discussed how
scaling laws demonstrate that improving AI performance has required
access to costly resources like high-performance processors and vast,
high-quality datasets. High resource requirements for developing
advanced AI can limit market entry, stifling competition. Early capability advantages for a leading firm can have positive feedback loops, which could enable them to raise more capital and pull ahead of many competitors. This raises concerns that only a few powerful entities will have access to and benefit from AI technologies.</p>
<strong>The use of AI might create or strengthen oligopolies.</strong>
AI developers are striving to have their models achieve
superintelligence: models that are able to carry out a wide range of
tasks across various domains better than humans. If someone did manage
to create such an AI, they might have a decisive advantage across large
swathes of industry, being sufficiently versatile to become an expert in
many or every market. It may become difficult or impossible for smaller
firms to carve out niche market spaces. The advanced capabilities of
general AI may outpace specialized models in diverse domains, making it
more difficult for new entrants to gain market share. Large firms
equipped with powerful AI systems could wield an enormous amount of
power, potentially leading to less competition, higher prices, and
slower innovation, hurting both labor and product markets.</p>

<h3 id="externalities">Externalities</h3>
<strong>Externalities
are consequences of economic activity that impact unrelated third
parties.</strong>
An externality is a side effect that stems from an economic activity,
impacting individuals or groups who are not directly involved in that
activity. Because of externalities, market prices for goods or services
may not fully reflect the costs that third parties, who are neither the
consumers nor the producers of that market, bear as a result of the
economic activity.</p>
<p> A classic example of a negative externality–—a harm to a third party—is
the case of pollution. Consider the Sriracha factory in Irwindale,
California, where jalapeño peppers are ground and roasted. Residents of
Irwingdale claimed that odors from the factory caused lung and eye
irritation and created an overall unpleasant smell in the town. The
factory, by producing these odors, was imposing a negative externality
upon the town’s residents, but since the townspeople received no
automatic compensation for this inconvenience, this was not reflected in
the price of Sriracha.</p>
<strong>We
can resolve externalities with litigation, property rights, and
taxation.</strong>
In 2013, locals sued the Sriracha factory: this legal action led the
factory to install new filters to reduce pollution. Litigation can be an
effective tool to resolve externalities by forcing compensation.
Economic theory suggests that bargaining over the externality can also
create efficient outcomes; for instance, if the property right to the
air was understood to belong to the townspeople, and that the factory
would have to stop polluting or compensate the townspeople at an
acceptable rate for their inconvenience <span class="citation"
data-cites="lafleur2013coase">[13]</span>.<p>
A third commonly used resolution to externalities is taxation. Smoking cigarettes imposes negative externalities on those near the smoker; governments will thus impose taxes on the sale of tobacco, which both raise revenues for the state to run social programs and discourage smoking by increasing the price of cigarettes to better reflect its true total cost. Determining the most effective method for resolving each externality is a topic of ongoing debate among economists. These debates extend to the externalities of AI on society. Potential policy responses are further discussed in the Governance chapter.</p>
<strong>The
development and deployment of AI systems can lead to negative
externalities.</strong>
The high energy consumption of advanced AI is a significant negative
externality on the environment. Training advanced AI models requires vast computational resources, consuming a significant amount of energy and contributing to greenhouse gas emissions. Emissions can lead to climate change, a cost borne by society at large, rather than only by those who pollute---unless the externality is corrected, such as by charging companies for the carbon they emit. Beyond emissions, several other issues such as worker displacement, the potential for misuse or accidents, and the risk of the loss of control of advanced systems present serious externalities as well.</p>

<h3 id="moral-hazard">Moral hazard</h3>
<p><strong>Moral hazards occur when risks are externalized.</strong>
Moral hazards are situations where an entity is encouraged to take on
risks, knowing that any costs will be borne by another party. Insurance
policies are a classic example: people with damage insurance on their
phones might handle them less carefully, secure in the knowledge that
any repair costs will be absorbed by the insurance company, not
them.<p>
The bankruptcy system ensures that no matter how much a company damages
society, the biggest risk it faces is its own dissolution, provided it
violates no laws. Companies may rationally gamble to impose very large
risks on the rest of society, knowing that if those risks ever come back
to the company, the worst case is the company going under. The company
will never bear the full cost of damage caused to society due to its
risk taking. Sometimes, the government may step in even prior to
bankruptcy. For example, American big banks took on massive risks in the
lead up to the 2008 financial crisis, but many of them were considered
"too big to fail", leading to an expectation that the government would
bail them out in time of need <span class="citation"
data-cites="acharya2016end">[9]</span>. These dynamics ultimately
contributed to the Great Recession.</p>
<p><strong>Developing advanced AIs is a moral hazard.</strong>
Throughout this textbook, we have outlined great risks from advanced
AIs. However, while the potential costs to society are immense, the
maximum financial downside to a tech company developing these AIs is
filing for bankruptcy.<p>
Consider the following, admittedly extreme, scenario. Suppose that a
company is on the cusp of inventing an AI system that would boost its
profits by a thousand-fold, making every employee a thousand times
richer. However, the company estimates that their invention comes with a
0.1% chance of a catastrophic accident leading to large-scale loss of life. In the likely case, the average person
in the economy would see some benefits due to increased productivity in
the economy, and possibly from wealth redistribution. Still, most people
view this gamble as irrational, preferring not to risk extinction for
modest economic improvements. On the other hand, the company may see
this as a worthwhile gamble, as it would make each employee considerably
richer.</p>
<p><strong>Risk internalization encourages safer behavior.</strong> In
the above examples of moral hazards, companies take risks that would
more greatly affect external parties than themselves. The converse of
this is risk internalization, where risks are primarily borne by the
party that takes them. Risk internalization compels the risk-taker to
exercise caution, knowing that they would directly suffer the
consequences of reckless behavior. If AI companies bear the risk of
their actions, they would be more incentivized to invest in safety
research, take measures to prevent malicious use, and be generally
disincentivized from creating potentially dangerous systems.</p>

<h2 id="inequality">6.4.3 Inequality</h2>
<strong>Most of
the world exhibits high levels of inequality.</strong>
In economics, inequality refers to the uneven distribution of
economic resources, including income and living conditions. The <em>Gini
coefficient</em> is a commonly used statistical measure of income or
wealth distribution within a country. It is a number between 0 and 1,
where 0 represents perfect equality (everyone has the same income or
wealth), and 1 signifies maximum inequality (one person has all the
income or wealth, and everyone else has none). Looking at Gini
coefficients, 71% of the world’s population lives in countries with
increasing inequality over the last thirty years <span class="citation"
data-cites="UN2020inequality">[5]</span>.</p>
<strong>Inequality
in the United States is particularly striking.</strong>
Figure 6.4 shows that the Gini coefficient in the
US has trended significantly upwards from 1969 to 2019. Over 50 years,
the pre-tax Gini coefficient has increased by 30%, while the post-tax
Gini coefficient has risen by 25%, suggesting that despite
redistributive taxation policies, the US income gap has widened
substantially. (For reference, this change in Gini coefficient is the
same size as moving from Canada to Saudi Arabia today. <span
class="citation" data-cites="UN2020inequality">[5]</span>) This increase
in the Gini coefficient is evidence of a growing inequality crisis. An
associated fall in social mobility—the ability of an individual to move
from the bottom income bracket to the top—cements inequalities over
generations.<p>
</p>
<figure id="fig:gini">
<img src="https://raw.githubusercontent.com/WilliamHodgkins/AISES/main/images/gini_v2.png" class="tb-img-full" style="width: 80%"/>
<p class="tb-caption">Figure 6.4: Inequality in the US (as measured by the Gini coefficient) has risen dramatically over the
last five decades, even adjusting for taxation. </p>
</figure>
<strong>The
distribution of gains from growth is highly unequal.</strong>
Nearly all the wealth gains over the past five decades have been
captured by the top 1% of income earners, while average
inflation-adjusted wages have barely increased <span class="citation"
data-cites="desilver2018workers">[6]</span>. A RAND Corporation working
paper estimated how the US income distribution would look today if
inequality was at the same level as in 1975—the results are in Table 6.1. Suppose my annual income is $15,000
today. If inequality was at the same level as in 1975, I would be paid
an extra $5,000. Someone else earning $65,000 today would instead have
been paid $100,000 had inequality held constant! We can see in the table
below that these increases in inequality have had massive effects on
individual incomes for everyone outside the top 1%.<p>
</p>
<br>
<div id="tab:income">
<table class="tableLayout">
<caption>Table 6.1: Real and counterfactual income distributions for all adults with income, in 2018 USD.
<span class="citation" data-cites="time2020trillion"></span></caption>
<thead>
<tr class="header">
<th style="text-align: left;">Percentile</th>
<th style="text-align: left;">Actual Income in 1975</th>
<th style="text-align: left;">Actual Income in 2018</th>
<th style="text-align: left;">Income in 2018 if Inequality Had Stayed
Constant</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td style="text-align: left;">25th %</td>
<td style="text-align: left;">$9,000</td>
<td style="text-align: left;">$15,000</td>
<td style="text-align: left;">$20,000</td>
</tr>
<tr class="even">
<td style="text-align: left;">Median</td>
<td style="text-align: left;">$26,000</td>
<td style="text-align: left;">$36,000</td>
<td style="text-align: left;">$57,000</td>
</tr>
<tr class="odd">
<td style="text-align: left;">75th %</td>
<td style="text-align: left;">$46,000</td>
<td style="text-align: left;">$65,000</td>
<td style="text-align: left;">$100,000</td>
</tr>
<tr class="even">
<td style="text-align: left;">90th %</td>
<td style="text-align: left;">$65,000</td>
<td style="text-align: left;">$112,000</td>
<td style="text-align: left;">$142,000</td>
</tr>
<tr class="odd">
<td style="text-align: left;">95th %</td>
<td style="text-align: left;">$80,000</td>
<td style="text-align: left;">$164,000</td>
<td style="text-align: left;">$174,000</td>
</tr>
<tr class="even">
<td style="text-align: left;">99th %</td>
<td style="text-align: left;">$162,000</td>
<td style="text-align: left;">$491,000</td>
<td style="text-align: left;">$353,000</td>
</tr>
<tr class="odd">
<td style="text-align: left;">Top 1% Mean</td>
<td style="text-align: left;">$252,000</td>
<td style="text-align: left;">$1,160,000</td>
<td style="text-align: left;">$549,000</td>
</tr>
</tbody>
</table>
</div>
<br>
<strong>Inequality carries
serious implications.</strong>
For those worst off in an unequal society, the implications are
severe. Beyond obvious problems like an inability to access essentials
and maintain a basic standard of living, there are additional concerns
in domains like health. A widening wealth gap often corresponds with a
health gap, where those with fewer resources have poorer health outcomes
due to less access to quality healthcare, lower quality nutrition, and
higher stress levels. For instance, life expectancy often varies
dramatically based on income in unequal societies <span class="citation"
data-cites="wilkinson2009spirit">[8]</span>.</p>
<strong>Everyone,
not just the poorest, suffers in an unequal society.</strong>
One of the most robust findings is that unequal societies have higher
levels of crime <span class="citation"
data-cites="kelly2000inequality">[9]</span>. When inequality is high,
so too are levels of social tension, dissatisfaction, and shame, which
can contribute to higher crime rates. This can lead to a cycle where the
fear of crime drives further inequality, as wealthier individuals and
neighborhoods invest in measures that segregate them further from the
rest of society, further increasing inequality and crime rates. A more
detailed discussion of this relative deprivation can be found in the
section on Cooperation and Conflict. Inequality is also related to other
signs of societal sickness: worse physical and mental health, increased
drug use, and higher rates of incarceration. Strikingly, inequality is a
strong predictor of political instability and violence as well. While it
may seem like those with more wealth are insulated from the negative
effects of inequality, in reality they suffer indirect consequences as
well.<p>
</p><figure id="fig:homicide">
<img src="https://raw.githubusercontent.com/WilliamHodgkins/AISES/main/images/inequality_crime_v2.png" class="tb-img-full" style="width:80%"/>
<p class="tb-caption">Figure 6.5: Countries with higher income inequality tend to have higher homicide rates.</p>
</figure>
<figure id="fig:political">
<img src="https://raw.githubusercontent.com/WilliamHodgkins/AISES/main/images/inequality_stability_v2.png" class="tb-img-full" style="width:80%"/>
<p class="tb-caption">Figure 6.6: Countries with higher income inequality tend to have higher rates of political instability.</p>
</figure>
<strong>The causes of inequality
are disputed.</strong>
Renowned economist Thomas Piketty, in his book “Capital in the 21st
Century,” posits that inequality results from the difference between the
interest rate investments in capital receive (r) and the rate of
economic growth (g) <span class="citation"
data-cites="piketty2014capital">[10]</span>. Piketty's key unorthodox claim is that capital is a ``gross substitute’’ for labor. On this view, as capital owners generate more wealth and capital, they are increasingly able to use capital to replace labor---imagine this as robots outcompeting human workers. Piketty argues that if the interest rate, which is the rate at which capital (or ``robots’’) can self-replicate, is greater than than the overall rate of economic growth, and thus faster than labor’s wage growth, then capital owners can become substantially and continuously richer than workers. This contrasts with the standard view that capital is a ``gross complement’’ to labor, with both labour and capital needed to produce goods. According to the standard view, increases in capital lead to higher labor productivity, which makes workers more efficient and valuable. Increased productivity raises wages; thus, increases in capital benefit both workers and capital owners.</p>

If AI serves as a gross substitute for labour, investment in AIs, with an effective interest rate (r) higher than the overall growth rate (g), will permit capital owners to continue accumulating capital, outcompeting workers and increasing inequality. While on the standard view, this would be a fundamentally new phenomenon, Piketty would argue that this is the exacerbation of a centuries-old trend. Such a scenario would contribute to growing inequality and negatively impact the livelihoods of workers. Issues of automation through AI and its broader societal consequences are discussed further in the Governance chapter.</p>

<h3 id="growth">6.4.4 Growth</h3>
<strong>Growth
is widely considered essential to a healthy society.</strong>
Some have claimed that societies should place less importance on economic growth, whether to reduce environment impacts or for other reasons.  However, looking at the world over the last 200 years provides strong evidence that economic growth can lead to vast improvements in human welfare in the domains of health, education, and more.</p>
<figure id="fig:preston">
<img src="https://raw.githubusercontent.com/WilliamHodgkins/AISES/main/images/preston_v2.png" class="tb-img-full" style="width: 80%"/>
<p class="tb-caption">Figure 6.3: Increases in GDP per capita strongly correlate with increases in life expectancy.</p>
<!--<figcaption>Increases in GDP per capita strongly correlate with-->
<!--increases in life expectancy.</figcaption>-->
</figure>
<p><strong>We have strong reasons for encouraging economic growth.</strong> The Preston curve is compelling evidence of a positive correlation
between a country’s gross domestic product (GDP) per capita—a measure of
the total production of goods and services–—and health outcomes. This
relationship illustrates that countries with higher GDP per capita have
better health outcomes, and particularly that poorer countries stand to
benefit immensely from improvements in GDP. This also holds across time:
average global life expectancy was just below 40 years at the start of
the 20th century, whereas today, with a much higher global GDP per
capita, the average person expects to live for 70 years. Nobel laureate
Amartya Sen suggests that one pathway through which growth improves
health is by reducing poverty and increasing investments in healthcare
<span class="citation" data-cites="sen1999economics">[3]</span>.<p>
<strong> Growth makes it easier for societies to support a wide range of values.</strong> The benefits of economic growth extend beyond physical health. There is
a strong association between economic prosperity and enhancements in
freedom and education <span class="citation"
data-cites="heckelman2000economic">[4]</span>. Prosperous societies can
afford stronger institutions to safeguard democratic freedoms and human rights. As societies become wealthier, more resources can be allocated to support cultural institutions, 
artists, and creative endeavors. Economic growth can enable Pareto improvements: changes that improve people’s lives without leaving anyone worse off, like growing a pie 
to give everyone bigger slices. Instead of debating whether to spend limited resources on a new hospital, school, or cultural center, encouraging economic growth can allow us to 
build all three. The increased wealth effectively allows us to accommodate multiple values, protecting value pluralism.</p>

<h2 id="beyond-economic-models">6.4.5 Beyond Economic Models</h2>
<p>However, in measuring societal wellbeing, we must recognize the shortcomings
of traditional economic metrics. Here, we will discuss the disconnect
between economic output and social value, why relying on economic models
of welfare economics can be inadequate in describing human goals, and
how more holistic measures of happiness and economic prosperity may give
us a clearer sense of true societal wellbeing.</p>
<h3 id="economic-output-and-gross-domestic-product">Economic Output and
Gross Domestic Product</h3>
<strong>Economic
indicators measure what we can quantify, not necessarily what we care
about.</strong>
Indicators like Gross Domestic Product (GDP) measure the monetary
value of final goods and services–—that is, those that are bought by the
final user–—produced in a country in a given period of time. When
economists discuss growth, they are typically referring to increases in
GDP. Measures of productive output like GDP are useful in gauging a
country’s economic health, but they fail to capture the value of
socially significant activities that aren’t priced in the market.</p>
<strong">Socially
important tasks are not captured by GDP.</strong>
Many essential roles in society, such as parenthood, community
service, and early education, are crucial to the wellbeing of
individuals and communities but are often undervalued or entirely
overlooked in GDP calculations <span class="citation"
data-cites="jones2016gdp">[14]</span>. While their effects might be
captured–—education, for instance, will increase productivity, which
increases GDP–—the initial activity does not count. The reason is
simple: GDP only accounts for activities that have a market price.
Consequently, efforts expended in these socially important tasks,
despite their high intrinsic value, are not reflected in the GDP
figures.</p>
<strong>Technologies
that make our lives better may not be measured either.</strong>
Technological advancements and their resultant value often fail to be
reflected adequately in GDP figures. For instance, numerous open-source
projects like Wikipedia provide knowledge to internet users worldwide at
no cost. However, because there’s no direct monetary transaction
involved, the immense value they offer isn’t represented in GDP. The
same applies to user-generated content on platforms like YouTube, where
the main contribution to GDP is through advertisement revenue and most
creators aren’t compensated for the value they create. The value viewers
derive from such platforms vastly outstrips the revenue generated from
ads or sales on these platforms, but this is not reflected in GDP.</p>
<strong>There
might be a similar disconnect between GDP and the social value of
AI.</strong>
As artificial intelligence systems become more integrated into our
daily lives, the disconnect between GDP and social value might become
more pronounced. For example, an AI system that provides free education
resources may significantly improve learning outcomes for millions, but
its contribution would be largely invisible in GDP terms. Similarly, an
AI may substantially increase GDP by facilitating high-frequency trading
without doing much to increase social wellbeing. This growing chasm
between economic metrics and real value could lead to policy decisions
that fail to harness the full potential of AI or inadvertently hamper
its beneficial applications. Recognizing this gap is a vital step
towards devising better ways of measuring and encouraging the socially
beneficial use of AI.</p>
<strong>The
proxy-purpose distinction is especially important for AI.</strong>
Imagine a future where an AI system is tasked with maximizing GDP,
often seen as a proxy for wellbeing. The system could potentially
achieve this goal by promoting resource-intensive industries or
fostering a work culture that prioritizes productivity over wellbeing.
In such a scenario, the GDP might increase, but at the cost of essential
considerations like environmental sustainability and human happiness.
Therefore, relying solely on economic indicators could lead to decisions
that, while effective in the short term, might harm society in the long
run.<p>
Understanding the limitations of economic measurements is an important
step towards a safer use of AI. It helps us question what we should
optimize. The aim of economic policy should not just be to maximize
economic output but also to promote overall societal wellbeing. AI
systems may thus need to take into account multiple factors, like
equality, sustainability, and personal fulfillment, rather than only
economic indicators. Such an approach could pave the way for a future
where AI contributes positively to steering society without compromising
the values we hold dear.</p>
<h3 id="models-of-welfare-economics">Models of Welfare Economics</h3>
<p><strong>The most basic form of welfare economics maximizes social surplus.</strong>
Social surplus is a measure of the total value created by a market:
it is the sum of the consumer surplus and the producer surplus. The
consumer surplus is the difference between the maximum price a consumer
is willing to pay and the actual market price. Conversely, the producer
surplus is the difference between the actual market price and the
minimum price a producer is willing to accept for a product or service.
By maximizing the total surplus, welfare economics seeks to maximize the
social value created by a market.<p>
For instance, imagine a scenario where consumers are willing to pay up
to $20 for a book, but the market price is only $15. Here, the consumer
surplus is $5. Similarly, if a producer is willing to sell the book for
a minimum of $10, the producer surplus is $5. The social surplus, and
hence the social value in this market, is $10: $5 consumer surplus plus
$5 producer surplus.</p>
<strong>We should
be wary of generalizing from simple models.</strong>
The core model of welfare economics has its limitations. Notably,
welfare economics is concerned with the maximization of surplus, but is
indifferent to its distribution. This might not align with societal
notions of fairness and equality. For example, an AI optimized to
maximize profits might model consumers well enough to enable perfect
price discrimination: allowing firms to sell each good at exactly a
consumer’s maximum willingness to pay, converting all the consumer
surplus into producer surplus, but leaving the sum total of “social
surplus” unchanged.<p>
However, social surplus is not the only thing we care about. Since
utility functions with respect to money are concave, we care about how
rich consumers and producers are to begin with. If, as is usually the
case, consumers are poorer than the owners of a company, then
transferring $5 (of surplus) from consumers to producers decreases total
utility. The real gains for the firm may be small compared to the losses
for the consumers.<p>
As AI systems become more integral to our economies, we must be mindful
of these complexities. A narrow focus on maximizing economic surplus
could lead us to promote AIs which, while efficient in a purely economic
sense, might have harmful consequences for society.</p>

<h3 id="happiness-in-economics">Happiness in Economics</h3>
<p><strong>There is a gap between human happiness and material prosperity.</strong>
Most people would agree that the goal of social sciences should not
be to just increase material wealth. A more meaningful aim would be to
enhance overall wellbeing or happiness. However, defining and measuring
happiness can be challenging. Whether happiness is correlated with
material wealth remains an ongoing research question; other aspects of
life like physical and mental health, job satisfaction, social
connections, and a sense of purpose seem important as well.</p>
<strong>Debates
about the Easterlin paradox highlight the complexity of understanding
happiness.</strong>
While wealthier people and countries are generally happier than their
less affluent counterparts, long-term economic growth does not always
correlate with long-term increases in happiness: this is the Easterlin
paradox. Several studies have tried to explore the relationship between
happiness and economic growth <span class="citation"
data-cites="earterlin2012china">[15]</span>. While some findings suggest
a correlation, others don’t, highlighting that our understanding of the
happiness-growth relationship is still evolving.<p>
We do, however, have strong evidence that inequality is unhelpful.
People often evaluate their wellbeing in relation to others; so, when
wealth distribution is unequal, people are dissatisfied and unhappy. For
instance, the recent rise in inequality may explain why there has been
no significant increase in happiness in the US over the last few decades
despite an approximately tenfold increase in real GDP and a fourfold
increase in real GDP per capita.</p>
<strong>More
holistic economic measurements can get closer to capturing what we
value.</strong>
Due to the disconnect between economic prosperity and true wellbeing,
some economists now propose the adoption of broader metrics. These new
measures aim to capture wellbeing more comprehensively, rather than
solely focussing on economic growth. One such measure is the Human
Development Index (HDI). The HDI comprises a nation’s average life
expectancy, education level, and Gross National Income (GNI) per capita
(which is similar to GDP). Notably, the UN uses the logarithm of GNI per
capita in the HDI calculation, which accounts for the diminishing
returns of wealth: the idea that each additional dollar earned
contributes less to a person’s happiness than the one before it. In
general, economists consider a “report card” of indicators to assess a
nation’s wellbeing, rather than just depending on a single measure. By
capturing various aspects of wellbeing, this approach could provide a
more holistic and accurate representation of a nation’s quality of
life.</p>
<strong>Summary.</strong>
Traditional economic measures and models are insufficient for
measuring and modeling what we care about. There is a disconnect between
what we measure and what we value; for instance, GDP fails to account
for essential unpaid labor and overvalues the production of goods and
services that add little to social wellbeing. While economic models are
useful, we must avoid relying too much on theoretically appealing models
and examine the matter of societal wellbeing with a more holistic
lens.</p>
<h3 id="conclusions-about-the-economic-engine">Conclusions About the
Economic Engine</h3>
<p><strong>We should be wary of using AI to increase metrics that are only proxies for wellbeing.</strong> Our current economic system strongly incentivises the deployment of AI systems that optimize for economic growth which, while often a worthy goal, may not capture essential aspects of societal health such as equality or sustainability (due to unequal distribution of gains, externalities, and other market failures). While economic objectives like GDP growth are quantifiable and easy to pursue, they may not truly reflect what makes a society happy and healthy.</p>

<p><strong>It is risky to let the goals of AI systems be determined entirely by economic incentives.</strong> AI systems created by for-profit businesses are designed to maximize shareholder profits, not societal wellbeing. If we let the economy decide what AIs do by letting largely unregulated markets create AIs, we may end up with an increase in inequality and exacerbation of market failures. Using money as a proxy for social value might seem practical, but it can distort societal priorities; for instance, this system implies that the preferences of wealthier individuals hold more weight since they are willing and able to spend more.</p>

<p>The examples in this section demonstrate a divergence between economic incentives and other important societal goals and should serve as a cautionary note. In the next section, we consider an alternative view: a framework that puts happiness front and center. Perhaps if we can direct AI systems to focus directly on promoting human happiness, we might aim to overcome the human biases and limitations that stop us from pursuing our happiness and enable the system to make decisions that have a positive impact on overall wellbeing.</p>

<br>
<br>
<h3>References</h3>

<div id="refs" class="references csl-bib-body" data-entry-spacing="0"
role="list">
<div id="ref-lipton2019shareholder" class="csl-entry" role="listitem">
<div class="csl-left-margin">[1] A.
M. Lipton, <span>“What we talk about when we talk about shareholder
primacy,”</span> <em>69 Case W. Rsrv. L. Rev.</em>, vol. 69, 2019.</div>
</div>
<!--<div id="ref-pwc2017sizing" class="csl-entry" role="listitem">-->
<!--<div class="csl-left-margin">[2] </div><div-->
<!--class="csl-right-inline"><span>“Sizing the prize,”</span> PwC’s Global-->
<!--Artificial Intelligence Study: Exploiting the AI Revolution.</div>-->
<!--</div>-->
<div id="ref-wikiprestoncurve" class="csl-entry" role="listitem">
<div class="csl-left-margin">[2] </div><div
class="csl-right-inline">Radeksz, [Online]. Available: <a
href="https://commons.wikimedia.org/wiki/File:PrestonCurve2005.JPG">https://commons.wikimedia.org/wiki/File:PrestonCurve2005.JPG</a></div>
</div>
<div id="ref-sen1999economics" class="csl-entry" role="listitem">
<div class="csl-left-margin">[3] A.
Sen, <span>“Economics and health,”</span> <em>The Lancet</em>,
1999.</div>
</div>
<div id="ref-heckelman2000economic" class="csl-entry" role="listitem">
<div class="csl-left-margin">[4] J.
Heckelman, <span>“Economic freedom and economic growth: A short-run
causal investigation,”</span> <em>Journal of Applied Economics</em>,
vol. III, pp. 71–91, May 2000, doi: <a
href="https://doi.org/10.1080/15140326.2000.12040546">10.1080/15140326.2000.12040546</a>.</div>
</div>
<div id="ref-UN2020inequality" class="csl-entry" role="listitem">
<div class="csl-left-margin">[5] </div><div
class="csl-right-inline"><span>“Inequality - bridging the
divide.”</span> [Online]. Available: <a
href="https://www.un.org/en/un75/inequality-bridging-divide">https://www.un.org/en/un75/inequality-bridging-divide</a></div>
</div>
<div id="ref-desilver2018workers" class="csl-entry" role="listitem">
<div class="csl-left-margin">[6] D.
Desilver, <span>“For most u.s. Workers, real wages have barely budged in
decades.”</span> [Online]. Available: <a
href="https://www.pewresearch.org/short-reads/2018/08/07/for-most-us-workers-real-wages-have-barely-budged-for-decades/">https://www.pewresearch.org/short-reads/2018/08/07/for-most-us-workers-real-wages-have-barely-budged-for-decades/</a></div>
</div>
<div id="ref-hanauer2020top" class="csl-entry" role="listitem">
<div class="csl-left-margin">[7] N.
Hanauer and D. M. Rolf, <span>“The top 1% of americans have taken $50
trillion from the bottom 90% — and that’s made the u.s. Less
secure.”</span></div>
</div>
<div id="ref-wilkinson2009spirit" class="csl-entry" role="listitem">
<div class="csl-left-margin">[8] R.
D. Wilkinson and K. Pickett, <em>The spirit level: Why more equal
societies almost always do better.</em> Bloomsbury Publishing,
2009.</div>
</div>
<div id="ref-kelly2000inequality" class="csl-entry" role="listitem">
<div class="csl-left-margin">[9] M.
Kelly, <span>“Inequality and crime,”</span> <em>The Review of Economics
and Statistics</em>, vol. 82, pp. 530–539, Feb. 2000, doi: <a
href="https://doi.org/10.1162/003465300559028">10.1162/003465300559028</a>.</div>
</div>
<div id="ref-piketty2014capital" class="csl-entry" role="listitem">
<div class="csl-left-margin">[10] T.
Piketty, <em>Capital in the twenty-first century</em>. Belknap Press,
2014.</div>
</div>
<div id="ref-yang2018war" class="csl-entry" role="listitem">
<div class="csl-left-margin">[11] A.
Yang, <em>The war on normal people: The truth about america’s
disappearing jobs and why universal basic income is our future</em>.
Hachette Books, 2018.</div>
</div>
<div id="ref-rust2005psychometrics" class="csl-entry" role="listitem">
<div class="csl-left-margin">[12] J.
Rust, <span>“Psychometrics and artificial intelligence.”</span> UK Test
Users Conference., 2005.</div>
</div>
<div id="ref-lafleur2013coase" class="csl-entry" role="listitem">
<div class="csl-left-margin">[13] S.
Lafleur, <span>“Coase, pigou, and sriracha sauce.”</span> [Online].
Available: <a
href="https://fcpp.org/2013/12/16/coase-pigou-and-sriracha-sauce/">https://fcpp.org/2013/12/16/coase-pigou-and-sriracha-sauce/</a></div>
</div>
<div id="ref-jones2016gdp" class="csl-entry" role="listitem">
<div class="csl-left-margin">[14] C.
Jones and P. Klenow, <span>“Beyond GDP? Welfare across countries and
time,”</span> <em>American Economic Review</em>, vol. 106, pp.
2426–2457, Sep. 2016, doi: <a
href="https://doi.org/10.1257/aer.20110236">10.1257/aer.20110236</a>.</div>
</div>
<div id="ref-earterlin2012china" class="csl-entry" role="listitem">
<div class="csl-left-margin">[15] R.
Easterlin, R. Morgan, M. Switek, and F. Wang, <span>“China’s life
satisfaction, 1990-2010,”</span> <em>Proceedings of the National Academy
of Sciences of the United States of America</em>, vol. 109, pp. 9775–80,
May 2012, doi: <a
href="https://doi.org/10.1073/pnas.1205672109">10.1073/pnas.1205672109</a>.</div>
</div>
<div id="ref-sandel2012money" class="csl-entry" role="listitem">
<div class="csl-left-margin">[16] S.
M. J., <em>What money can<span>’</span>t buy: The moral limits of
markets</em>. Farrar, Straus; Giroux; Farrar, Straus; Giroux,
2012.</div>
</div>
<div id="ref-raworth2017doughnut" class="csl-entry" role="listitem">
<div class="csl-left-margin">[17] K.
Raworth, <em>Doughnut economics: Seven ways to think like a 21st century
economist</em>. Chelsea Green Publishing, 2017.</div>
</div>
</div>