Your bus breaks down over some old train tracks, and you and your fellow passengers have taken it upon yourselves to move it. Most of the passengers, particularly the largest and strongest, are very concerned about being hit by a freight train.
While this is a perfectly reasonable fear, you think that the semi-truck barreling down the road is of greater importance.
You mention that nobody knows if there is a train (although some people seem incredibly confident about its existence), but just about everyone can see the truck when you point it out. The other passengers tell you that, unlike the train, the semi won’t destroy the bus, and might actually move it off the tracks. You reply that they seem to be underweighting the damage the truck will cause to the bus, and they accuse you of distracting people from the far more weighty task of saving the bus from the train, as the tracks and the road are perpendicular. A third passenger points out that nobody knows whether any passenger is capable of moving the bus, and that there are economic and diplomatic incentives to nail the bus down anyway. You point out that the metaphor has gotten a bit ridiculous, and the semi continues its approach.
The rationalist and EA communities have focused much of their efforts on AI alignment, an admirable project which has likely been responsible for much of the progress in and public concern for AI safety. Unfortunately, the prospect of economic disruption has not received the same attention. Everyone agrees that it’s going to happen. (In the field of computer science, it is already underway.) Everyone agrees that its effects will be widespread and deleterious. However, very few individuals or organizations have been working on this problem. The purpose of this article is to give a quantitative estimate of the first-order effects and a qualitative evaluation of the second-order effects of AI on the American labor force.
I began with a list of eight-hundred thirty-one labor sectors compiled by the U.S. Bureau of Labor Statistics[1] and asked Anthropic’s Sonnet 5.5 to assign a “disruption potential” score to each labor sector on a scale from one to five. I defend my use of AI by noting my own lack of qualifications; my own assessment would have largely been based on speculation. The scale was defined by me as the following:
Level 1: No disruptive potential. Employment, productivity, and function change negligibly. Example: Aides for special needs children have little use for AI, as their function relies entirely upon human connection.
Level 2: Some disruptive potential. AI is often used as a tool, but job function largely stays the same. Employment is reduced, but the industry is still largely human-dependent. Example: Realtors will use AI to write contracts and find suitable houses, but clients will hire a human to show them the house and give them personal advice.
Level 3: Moderate disruption. Job function is significantly modified by AI, but the fundamental premise of the job remains the same. Employment is significantly reduced. Example: Graphic designers will use AI tools for much of their work, and many smaller businesses will not hire them and use AI tools directly. However, the premise of designing visually appealing content remains the same, and larger operations will still hire humans, though in much lower numbers.
Level 4: Severe disruption. The fundamental premise of the job changes, and employment is reduced to a small fraction of its former size. Example: Logistics planners will be reduced to checking over, rubber-stamping, and accepting liability for AI-generated plans, rather than managing things themselves.
Level 5: Complete disruption. AI renders the industry obsolete. Nearly all employees must reskill. Example: Translators of well-known languages, such as English and Chinese, will be completely replaced by AI, which can appropriately adjust translations for tone and regional peculiarities. Human translators are exclusively employed for ceremonial purposes.
Results
These rankings assume AI either remains similar to its current state or improves modestly and no “freight train”-esque catastrophes occur. I then totaled the people employed in sectors in each disruption level:
At first, the above graph seems quite optimistic. Only five percent of the labor force, a mere eight million people, will have their careers completely annihilated! The picture becomes significantly worse when one considers what the disruptive potentials actually mean to most people; the second category, in which the sector experiences layoffs and changes in significant day-to-day operations, would be considered moderately disrupted, and the last three categories, in which the sector’s employment is demolished and day-to-day operations are completely different, as severely disrupted.
So, given that AI doesn’t improve too much, it will upend one-third of the U.S. labor force. The future looks much grimmer when the sectors are broken down by income and education level. Income level is given as the median national yearly earnings for each sector, and education level is given as the level of education typically required for an entry-level position in each sector. Both statistics were provided by the USBLS[2]. Six sectors (actors, dancers, musicians and singers, disc jockeys, “entertainers and performers, all other,” and fishing and hunting workers) did not have median yearly incomes given and were therefore excluded from the income-level estimates.
Note: Raw disruption potentials were not given as some bars were not visible.
Notice that the lowest-income (<40k) and lowest-education (no formal education) sectors are most likely to be minimally disrupted, and that a very small portion of high-income (100k+) sectors escape disruption. Also notice that most people are either employed in sectors with a low typical education (no formal education or a high school diploma) or a typical education consisting of a bachelor’s degree. Below, I give a breakdown of disruption across income and education levels; as before, the left y-axis gives employees in the millions and the percentage of that portion of the labor force.
As noted before, the highest-income sectors are nearly all disrupted (only 4.5% of employees avoid significant disruption), and the lowest-income sectors are least affected; the middle three income levels (40k-100k) face similar amounts of disruption, with about a third being undisturbed and half being completely upended.
Disruption levels across education levels points to another major structural issue. Sectors requiring an associates degree or some college, as well as sectors requiring a bachelor’s degree, are the sectors which best lend themselves to upwards mobility and the maintenance of the middle class. These sectors face much higher rates of severe disruption than than sectors that typically require graduate education, which are generally pursued by the wealthy, and much higher rates of general disruption than sectors that do not require any collegiate education, which are generally not high-income sectors. Notice that sectors with a typical education of an associates degree or a small amount of college are most likely to be rendered obsolete; these jobs are often undertaken by those attempting to escape poverty. The consequence of this uneven disruption is that AI will make upwards mobility significantly more difficult and will further shrink the middle class.
Another concerning trend is that sectors that typically require bachelor-level or higher degrees also have the highest rates of general disruption, while sectors with no typical college education are far more likely to be minimally disrupted, so there will be less incentive to go to college and attain higher education. While a college education is not necessary to be knowledgeable or intelligent, reducing the incentives for people to learn or educate themselves by any means will also have deleterious effects.
Second Order Effects
The second-order effects of such high disruption are much harder to predict, especially with such a rough estimate. The most obvious consequence is that unemployment will shoot through the roof; 5% of the labor force will be completely unemployable in their current sector as AI renders it obsolete. Added to the approximately 4.5% unemployment at the end of 2025, unemployment rates are comparable to the 2008 financial crash (~9%)[3]; factoring in the fact that another 60% of the labor force works in a sector that will be disrupted, expect unemployment rates to reach or eclipse that of the Great Depression, when unemployment peaked around 25%[4].
Another second-order effect is the concentration of wealth and power; as the middle class shrinks and upwards mobility becomes harder to achieve, wealth inequality will increase drastically; not only will it be harder to stay out of poverty, with all the layoffs and decreases in income which the layoffs allow (high supply of potential employees and low hiring demand), but it will be harder to escape poverty. Wealth is largely concentrated within the companies that make or control AI. This is not a situation where wealth inequality increases but overall standards of living improve; most Americans are jobless or in poverty, so standards of living decrease compared to pre-disruption society.
A third effect to consider are industries which are not directly impacted by AI, but are heavily disrupted by the social and economic changes AI causes. Among the largest ten sectors with negligible disruption potential are “home health and personal care aides” (#1), “janitors and cleaners” and “maids and housekeeping cleaners” (#3 and #9), “waiters and waitresses” (#4), and “cooks, restaurant” with “first-line supervisors of food preparation and serving workers” (#8 and #10). In fact, many of the sectors facing minimal disruption from AI itself are reliant upon large numbers of wealthy and middle-class individuals purchasing frivolous goods and services. While a bit of the loss resulting from the new lack of customers will be made up by richer customers buying more, disruption will still be massive; a man with the wealth of ten middle-class Americans might buy a more expensive wardrobe, but he won’t eat ten chain-restaurant meals or ride ten Ubers.
In essence, if modestly improved AI is allowed to disrupt our economy unrestricted, America will suddenly have a glut of newly unemployed and impoverished citizens amidst increasing concentration of power and wealth. Further, nations whose populace’s well-being is not integral to wealth generation lends itself to dictatorships (for a further discussion, see CGP Grey). What happens after is difficult to predict; I find it relevant to mention that, in addition to rational and empathetic beings, humans are apes, which are famously tolerant of perceived hoarders and thieves.
Epistemic Status
Do I actually believe this? I tend to assume that most humans are intelligent, empathetic, and at the very least self-interested, so regulatory bodies and society in general will probably adjust and intercede to prevent the situation from completely deteriorating. However, the probability of the worst-case scenario is not negligible enough to be ignored, and the most probable outcome isn’t exactly peaches and cream either. My personal projection, considering other factors such as climate change and sub-replacement birth rates, is that we are headed towards three to five decades of massive social disruption and semi-sporadic violence, with social and economic change on par with the Industrial Revolution. I don’t have a complete solution for this problem, but I’m sure someone smarter than me will.
very few individuals or organizations have been working on this problem.
My main issue with this is the potential saturation of workers on the problem. Once Amodei wrote down the problem and a fan internalised it, there is no other work which could further convince the fan. Once anyone proposes a solution, everyone dies the solution is merely to reach and convince those responsible.
Your bus breaks down over some old train tracks, and you and your fellow passengers have taken it upon yourselves to move it. Most of the passengers, particularly the largest and strongest, are very concerned about being hit by a freight train.
While this is a perfectly reasonable fear, you think that the semi-truck barreling down the road is of greater importance.
You mention that nobody knows if there is a train (although some people seem incredibly confident about its existence), but just about everyone can see the truck when you point it out. The other passengers tell you that, unlike the train, the semi won’t destroy the bus, and might actually move it off the tracks. You reply that they seem to be underweighting the damage the truck will cause to the bus, and they accuse you of distracting people from the far more weighty task of saving the bus from the train, as the tracks and the road are perpendicular. A third passenger points out that nobody knows whether any passenger is capable of moving the bus, and that there are economic and diplomatic incentives to nail the bus down anyway. You point out that the metaphor has gotten a bit ridiculous, and the semi continues its approach.
The rationalist and EA communities have focused much of their efforts on AI alignment, an admirable project which has likely been responsible for much of the progress in and public concern for AI safety. Unfortunately, the prospect of economic disruption has not received the same attention. Everyone agrees that it’s going to happen. (In the field of computer science, it is already underway.) Everyone agrees that its effects will be widespread and deleterious. However, very few individuals or organizations have been working on this problem. The purpose of this article is to give a quantitative estimate of the first-order effects and a qualitative evaluation of the second-order effects of AI on the American labor force.
I began with a list of eight-hundred thirty-one labor sectors compiled by the U.S. Bureau of Labor Statistics[1] and asked Anthropic’s Sonnet 5.5 to assign a “disruption potential” score to each labor sector on a scale from one to five. I defend my use of AI by noting my own lack of qualifications; my own assessment would have largely been based on speculation. The scale was defined by me as the following:
Level 1: No disruptive potential. Employment, productivity, and function change negligibly. Example: Aides for special needs children have little use for AI, as their function relies entirely upon human connection.
Level 2: Some disruptive potential. AI is often used as a tool, but job function largely stays the same. Employment is reduced, but the industry is still largely human-dependent. Example: Realtors will use AI to write contracts and find suitable houses, but clients will hire a human to show them the house and give them personal advice.
Level 3: Moderate disruption. Job function is significantly modified by AI, but the fundamental premise of the job remains the same. Employment is significantly reduced. Example: Graphic designers will use AI tools for much of their work, and many smaller businesses will not hire them and use AI tools directly. However, the premise of designing visually appealing content remains the same, and larger operations will still hire humans, though in much lower numbers.
Level 4: Severe disruption. The fundamental premise of the job changes, and employment is reduced to a small fraction of its former size. Example: Logistics planners will be reduced to checking over, rubber-stamping, and accepting liability for AI-generated plans, rather than managing things themselves.
Level 5: Complete disruption. AI renders the industry obsolete. Nearly all employees must reskill. Example: Translators of well-known languages, such as English and Chinese, will be completely replaced by AI, which can appropriately adjust translations for tone and regional peculiarities. Human translators are exclusively employed for ceremonial purposes.
Results
These rankings assume AI either remains similar to its current state or improves modestly and no “freight train”-esque catastrophes occur. I then totaled the people employed in sectors in each disruption level:
At first, the above graph seems quite optimistic. Only five percent of the labor force, a mere eight million people, will have their careers completely annihilated! The picture becomes significantly worse when one considers what the disruptive potentials actually mean to most people; the second category, in which the sector experiences layoffs and changes in significant day-to-day operations, would be considered moderately disrupted, and the last three categories, in which the sector’s employment is demolished and day-to-day operations are completely different, as severely disrupted.
So, given that AI doesn’t improve too much, it will upend one-third of the U.S. labor force. The future looks much grimmer when the sectors are broken down by income and education level. Income level is given as the median national yearly earnings for each sector, and education level is given as the level of education typically required for an entry-level position in each sector. Both statistics were provided by the USBLS[2]. Six sectors (actors, dancers, musicians and singers, disc jockeys, “entertainers and performers, all other,” and fishing and hunting workers) did not have median yearly incomes given and were therefore excluded from the income-level estimates.
Note: Raw disruption potentials were not given as some bars were not visible.
Notice that the lowest-income (<40k) and lowest-education (no formal education) sectors are most likely to be minimally disrupted, and that a very small portion of high-income (100k+) sectors escape disruption. Also notice that most people are either employed in sectors with a low typical education (no formal education or a high school diploma) or a typical education consisting of a bachelor’s degree. Below, I give a breakdown of disruption across income and education levels; as before, the left y-axis gives employees in the millions and the percentage of that portion of the labor force.
As noted before, the highest-income sectors are nearly all disrupted (only 4.5% of employees avoid significant disruption), and the lowest-income sectors are least affected; the middle three income levels (40k-100k) face similar amounts of disruption, with about a third being undisturbed and half being completely upended.
Disruption levels across education levels points to another major structural issue. Sectors requiring an associates degree or some college, as well as sectors requiring a bachelor’s degree, are the sectors which best lend themselves to upwards mobility and the maintenance of the middle class. These sectors face much higher rates of severe disruption than than sectors that typically require graduate education, which are generally pursued by the wealthy, and much higher rates of general disruption than sectors that do not require any collegiate education, which are generally not high-income sectors. Notice that sectors with a typical education of an associates degree or a small amount of college are most likely to be rendered obsolete; these jobs are often undertaken by those attempting to escape poverty. The consequence of this uneven disruption is that AI will make upwards mobility significantly more difficult and will further shrink the middle class.
Another concerning trend is that sectors that typically require bachelor-level or higher degrees also have the highest rates of general disruption, while sectors with no typical college education are far more likely to be minimally disrupted, so there will be less incentive to go to college and attain higher education. While a college education is not necessary to be knowledgeable or intelligent, reducing the incentives for people to learn or educate themselves by any means will also have deleterious effects.
Second Order Effects
The second-order effects of such high disruption are much harder to predict, especially with such a rough estimate. The most obvious consequence is that unemployment will shoot through the roof; 5% of the labor force will be completely unemployable in their current sector as AI renders it obsolete. Added to the approximately 4.5% unemployment at the end of 2025, unemployment rates are comparable to the 2008 financial crash (~9%)[3]; factoring in the fact that another 60% of the labor force works in a sector that will be disrupted, expect unemployment rates to reach or eclipse that of the Great Depression, when unemployment peaked around 25%[4].
Another second-order effect is the concentration of wealth and power; as the middle class shrinks and upwards mobility becomes harder to achieve, wealth inequality will increase drastically; not only will it be harder to stay out of poverty, with all the layoffs and decreases in income which the layoffs allow (high supply of potential employees and low hiring demand), but it will be harder to escape poverty. Wealth is largely concentrated within the companies that make or control AI. This is not a situation where wealth inequality increases but overall standards of living improve; most Americans are jobless or in poverty, so standards of living decrease compared to pre-disruption society.
A third effect to consider are industries which are not directly impacted by AI, but are heavily disrupted by the social and economic changes AI causes. Among the largest ten sectors with negligible disruption potential are “home health and personal care aides” (#1), “janitors and cleaners” and “maids and housekeeping cleaners” (#3 and #9), “waiters and waitresses” (#4), and “cooks, restaurant” with “first-line supervisors of food preparation and serving workers” (#8 and #10). In fact, many of the sectors facing minimal disruption from AI itself are reliant upon large numbers of wealthy and middle-class individuals purchasing frivolous goods and services. While a bit of the loss resulting from the new lack of customers will be made up by richer customers buying more, disruption will still be massive; a man with the wealth of ten middle-class Americans might buy a more expensive wardrobe, but he won’t eat ten chain-restaurant meals or ride ten Ubers.
In essence, if modestly improved AI is allowed to disrupt our economy unrestricted, America will suddenly have a glut of newly unemployed and impoverished citizens amidst increasing concentration of power and wealth. Further, nations whose populace’s well-being is not integral to wealth generation lends itself to dictatorships (for a further discussion, see CGP Grey). What happens after is difficult to predict; I find it relevant to mention that, in addition to rational and empathetic beings, humans are apes, which are famously tolerant of perceived hoarders and thieves.
Epistemic Status
Do I actually believe this? I tend to assume that most humans are intelligent, empathetic, and at the very least self-interested, so regulatory bodies and society in general will probably adjust and intercede to prevent the situation from completely deteriorating. However, the probability of the worst-case scenario is not negligible enough to be ignored, and the most probable outcome isn’t exactly peaches and cream either. My personal projection, considering other factors such as climate change and sub-replacement birth rates, is that we are headed towards three to five decades of massive social disruption and semi-sporadic violence, with social and economic change on par with the Industrial Revolution. I don’t have a complete solution for this problem, but I’m sure someone smarter than me will.
https://www.bls.gov/emp/tables/occupational-projections-and-characteristics.htm
https://www.bls.gov/emp/tables/industry-occupation-matrix-occupation.htm
https://www.bls.gov/charts/employment-situation/civilian-unemployment-rate.htm
https://data.bls.gov/timeseries/LFU21000100&series_id=LFU22000100&from_year=1929&to_year=1939&periods_option=specific_periods&periods=Annual+Data