aises_8_2

<h1 id="growth">8.2 Economic Growth</h1>

<h1 id="ai-population">8.2.1 AI and Population</h1>
<p>There are many different factors that feed into the rate of economic growth, and AI has the potential to amplify several of them. For instance, deploying AI systems could artificially augment the effective population of workers, improve the efficiency of human labor, or accelerate the development of new technologies that improve productivity. While it is generally accepted that AI will boost economic growth to some degree, there is debate over the exact magnitude of the impact it is likely to have. Some researchers believe that it will speed up growth to an unprecedented rate, which we refer to as “explosive growth,” while others think its impact will be limited by other social and economic factors. We will now explore some of the arguments for and against the likelihood of AI causing explosive growth.</p>

<p><strong>Population growth may drive economic growth by accelerating technological progress.</strong> The worldwide economic acceleration observed in recent centuries has been variously attributed to the unique conditions of the industrial revolution, the technologies developed in 18th and 19th century Europe, and the growth in total population over time. Population growth is emphasized most by the semi-endogenous theory of economic growth. It holds that since economic growth causes population growth by reducing bottlenecks on population growth and population growth causes economic growth by providing a large labor force (including an increasing number of researchers driving technological progress), there is a positive feedback loop and so population growth is the key factor to consider when looking at the increase in the economic output over time.</p>

<p>According to this theory, human population growth was determined for many thousands of years by the availability of food. As agricultural technologies were developed, food became easier to produce, which allowed for more population growth. Since larger populations have more opportunities to innovate and develop better technology, some economists argue that this process loops back into itself recursively, producing a faster-than-exponential development curve over the long run.</p>

<p>This acceleration ultimately slowed down in the mid-20th century. The semi-endogenous theory explains this slowdown as a result of the independent decline in the population growth rate, arguing that demographic changes such as falling birth rates uncoupled productivity and population growth. This is one explanation for why economic growth did not explode in the late 20th and early 21st centuries. According to this line of reasoning, lifting the population bottleneck would once again enable the multi-thousand-year trend of accelerating growth to continue until we exhaust physical resources like energy and space.</p>

<figure id="fig:homicide">
<embed src="https://raw.githubusercontent.com/WilliamHodgkins/AISES/main/images/growth_v2 updated.png" class="tb-img-full" style="width: 80%"/>
<p class="tb-caption">Figure 8.1: Economic output could grow much faster than past trends if not constrained by population.</p>
<!--<figcaption>Growth of economic output in two scenarios: growth slowdown,-->
<!--and AI-driven growth explosion</figcaption>-->
</figure>

<p><strong>AIs may fuel effective population growth.</strong> If AIs can automate the majority of important human tasks (including further AI development), this could lift the bottleneck on labor that some believe is the primary obstacle to explosive growth. There are some reasons to think that AIs could boost the economic growth rate by substituting for human labor. As easily duplicable software, the AI population can grow at least as quickly as we can manufacture hardware to run it on—-much faster than humans take to reproduce and learn skills from adults. This replication of labor could then boost the effective workforce and accelerate productivity.</p>

<p><strong>AIs may accelerate further AI development.</strong> If AIs become proficient at altering themselves to enhance their own capabilities, then we could see accelerated AI development through recursive self-improvement. At each step, this may make them more efficient at performing tasks and producing goods, as well as make them better at self-improving. It is worth noting that the growth of AI capabilities driven by human design is already much faster than the growth of human capabilities driven by biological evolution; over the entire course of evolution from humans’ last common ancestor with chimpanzees, human brains grew roughly 4 times in size, whereas over the decade after AlexNet, the largest machine learning models increased in size by the same amount roughly every 16 months. If AIs were able to effectively automate their own R&D, then this could speed up the improvements in their own performance even more.</p>

<p><strong>AI population and self-improvement could form a positive feedback loop.</strong> Taken together, a growing number of AIs and accelerated AI development through recursive self-improvement could form a positive feedback loop between AI population and technology production, mirroring the loop with human population proposed in the semi-endogenous theory. However, since neither the self-replication nor self-improvement aspects of this loop are subject to biological constraints, the AI population feedback loop could theoretically be much faster than the one involving humans. Assuming we do not see a slowdown in AI development or face other bottlenecks like energy production, this self-amplifying cycle could lead to unprecedented economic growth.</p>

<p><strong>AI automation may create explosive growth.</strong> Some researchers have argued that if we add AIs to standard models of economic growth (such as the Solow model) developed using economic theory and past data, we find that AIs could trigger a dramatic surge in economic growth. Some studies suggest that such AIs could spark unprecedented growth, causing the world economy to grow at rates exceeding ten times the current growth rate. If this were to transpire in its most extreme form, it could result in an unprecedented acceleration of scientific and technological advancement, reshaping our economy and the trajectory of human history over a period of perhaps only a few years. Growth of this magnitude would be unlike anything in human history: for the past 10,000 years after the agricultural revolution, total world output grew at 0.1% per year, steadily increasing over the last 1000 years to single-digit percentage points.</p>

<p>However, other researchers have argued that there are potential physical constraints, as well as social and economic dynamics, that could prevent AI from driving explosive growth. We will now explore some of these arguments.</p>

<p><strong>Non-accumulable factors may bottleneck production.</strong> The argument for explosive growth relies on the assumption that some factors, such as AI labor and its outputs, are indefinitely accumulable, such that they can be repeatedly re-invested and lead to ever-increasing returns. However, this may not always be true. For example, there is a limited supply of land that can accommodate physical infrastructure required to run AIs. There is also a fixed amount of energy that the Earth receives from the sun each day, representing a theoretical limit on the rate at which AIs could operate. While we have not yet reached the upper bounds of either land or energy available, they do imply that economic growth is unlikely to continue indefinitely. Whether or not we see AI-driven explosive growth depends on whether or not this happens before the limits of non-accumulable factors become constraints.</p>

<p><strong>R&D may be harder to accelerate than expected.</strong> A key ingredient of the economic feedback loop described above is the idea that R&D activities will improve technology, and thus improve efficiency of work and production. However, it could be the case that, in AI development, the “low-hanging fruit” of new ideas that yield substantial improvements have already been taken, and that there will be diminishing returns on future R&D, even if done by an increasing population of AIs. This could weaken the effect of additional research activities and slow down the feedback loop, potentially to such a degree that it could not fuel explosive economic growth.</p>

<p><strong>AI adoption and its impact could be slow and gradual.</strong> Some researchers argue that the greatest impact of a new technology may not manifest as an intense peak during the early stages of innovation. Rather, the productivity gains may be delivered as a slower increase continuing over a longer period of time, as the technology is gradually adopted by a wide range of industries. This could be because the technology needs to be adapted to many different tasks and settings, humans need to be trained to operate it, and other tools and processes that are compatible with it need to be developed. For example, although the first electric dynamo suitable for use in industry was invented in the 1870s, it took several decades for electricity to become integrated within industries. It has been argued that this is why electricity only boosted the US economy significantly in the early 20th century. Similarly, a slow process of diffusion could also smooth out AI’s impact on today’s economy.</p>

<p><strong>Regulations and human preferences may prevent complete automation of the economy.</strong> Even if AIs could theoretically automate all tasks and create explosive growth, social factors may prevent this from happening. For example, fears about risks associated with AI may prompt regulations that restrict the technology’s use, perhaps requiring that certain services, such as medical or legal services, be performed by human professionals. Other regulations seeking to protect intellectual property might limit the amount of training data available, thus inhibiting the growth of AI capabilities.</p>

<p>Besides regulations, humans’ own preferences may also limit the fraction of tasks that are automated. For example, we can speculate that humans may prefer certain services that involve a high degree of social interaction, such as those in healthcare, education, and counseling, to be provided by other humans. Additionally, people may always be more interested in watching human athletes, actors, and musicians, and in buying artwork produced by humans. In some cases, these jobs may therefore evade automation, even if it were theoretically possible to automate them, just as there are still professional human chess players, despite the fact that machines have long been able to beat Grandmasters.

<p><strong>Economic growth may be bottlenecked by non-automated tasks.</strong> To generate explosive growth, AI automation may need to be extremely comprehensive; even if most tasks were automated, the non-automated ones may become constraints on the growth rate of productivity. For example, if AI exhibited above-human performance on cognitive tasks but its ability to move in the physical world lagged behind due to slow progress in robotics, this would cap the extent to which AI could accelerate economic growth. This is because physical tasks involved in manufacturing products and moving them around are likely to remain important; even if AI made design more efficient, output could only increase up to the limit imposed by how quickly the physical tasks could be performed.

<p>It is worth noting that, although robotics has so far appeared more challenging than achieving cognitive tasks with AI, there are some initial signs that progress in the area could accelerate in the coming years. As such, robotics may not always represent a bottleneck to growth. Nevertheless, there could be many types of tasks in production processes that constrain the overall output of the system, no matter how efficient AI can make the other tasks.</p>

<p><strong>Baumol’s cost disease.</strong> Another reason why even just a few non-automated jobs could prevent explosive growth is the concept of Baumol’s cost disease, proposed by the economist William Baumol in the 1960s. This idea states that, when technology increases productivity in one industry, the prices of its products fall, and the wages of its workers rise. Another industry, which cannot easily be made more efficient with technology, will also need to increase its workers’ wages, to prevent them moving into higher-paying jobs in technologically enhanced sectors. As a result, the prices of outputs in those sectors take up an increasing share of the overall economy. Thus, even if some industries undergo rapid increases in productivity, the effect on the growth of the economy as a whole is more muted. This is one explanation for why the prices of goods such as TVs have declined over time, while the costs of healthcare and education have risen. According to this concept, if AI automates many jobs, but not all of them, its economic impact could be substantial, but not necessarily explosive.</p>
<p>While AI has the potential to significantly enhance economic growth through various routes including improving workers' productivity and accelerating R&D, the extent and speed of this growth remain uncertain. Theories suggesting explosive growth due to AI rely on relatively strong assumptions around the removal of potential bottlenecks. However, potential constraints such as physical resource limits, diminishing returns on R&D, slow AI adoption, regulatory and human preferences, and non-automatable tasks could moderate this growth. Therefore, while AI's impact on economic growth is likely to be substantial, whether it will lead to unprecedented economic expansion or be tempered by various limiting factors remains an open question.

<br>
<br>
<h3>References</h3>
<div id="ref-hinton2023doomer" class="csl-entry" role="listitem">
<div class="csl-left-margin">[1] </div><div
class="csl-right-inline"><span>“What really made Hinton an AI
doomer.”</span> [Online]. Available: <a
href="https://www.wired.com/story/geoffrey-hinton-ai-chatgpt-dangers/#:~:text=Hinton%20concluded%20that%20as%20AI,to%20be%20five%20to%2020.”">https://www.wired.com/story/geoffrey-hinton-ai-chatgpt-dangers/#:~:text=Hinton%20concluded%20that%20as%20AI,to%20be%20five%20to%2020.”</a></div>
</div>
<div id="refs" class="references csl-bib-body" data-entry-spacing="0"
role="list">
<div id="ref-Davidson-growth" class="csl-entry" role="listitem">
<div class="csl-left-margin">[2] T.
Davidson, <span>“Could advanced AI drive explosive economic
growth?”</span> [Online]. Available: <a
href="https://www.openphilanthropy.org/research/could-advanced-ai-drive-explosive-economic-growth/">https://www.openphilanthropy.org/research/could-advanced-ai-drive-explosive-economic-growth/</a></div>
</div>
<div id="ref-cotra2020forecasting" class="csl-entry" role="listitem">
<div class="csl-left-margin">[3] A.
Cotra, <span>“Forecasting TAI with biological anchors.”</span> [Online].
Available: <a
href="https://docs.google.com/document/d/1k7qzzn14jgE-Gbf0CON7_Py6tQUp2QNodr_8VAoDGnY/edit#heading=h.r808lkmw6i0l">https://docs.google.com/document/d/1k7qzzn14jgE-Gbf0CON7_Py6tQUp2QNodr_8VAoDGnY/edit#heading=h.r808lkmw6i0l</a></div>
</div>
<div id="ref-erdil2023algorithmic" class="csl-entry" role="listitem">
<div class="csl-left-margin">[4] E.
Erdil and T. Besiroglu, <span>“Algorithmic progress in computer
vision.”</span> 2023. Available: <a
href="https://arxiv.org/abs/2212.05153">https://arxiv.org/abs/2212.05153</a></div>
</div>
</div>