Technically, we're applying a single layer perceptron to the problem, with weights and biases taken from our beliefs. You don't have to understand the previous sentence at all. Let's begin by exploring these beliefs.
1. From First Principles
Let’s imagine your job consists of nothing else than solving mathematical equations, and you work along 99 colleagues doing the same (think “computers” from the 1940s). Given AI capabilities reach 50% success rate on 8 hour long tasks in your domain (let’s call it a workday), and the costs of doing so become less than your wage, your employer becomes financially incentivised to implement automation, let 50 of you go, and hire a few to oversee the AIs (a distinctly different job). In the case that AIs reach 99% success rate, your whole department will consist of one person solving the equations, and a few doing the oversight.
There are some adoption factors though. The AI system may do the same amount of calculations you do slower, at the same pace, or faster. Except for the much slower scenario, we can incorporate this fact simply in the cost component. Then comes the question of subtasks. Say all the tasks AIs fail in can be subdivided into two subproblems, and they fail only in one of those. That would reduce the work that still has to be done by a human to half. This is the edge case of the “AI enhanced workforce”, where people using AI are capable of doing more work than the ones not using it. If 3 out of 4 workers improve productivity by a third, the fourth worker becomes unnecessary, given flat demand. On the other hand, implementing an AI supported workflow may have an upfront transformation cost, that may slow down adoption. And there are other adoption factors than pure cost: workplace social interactions and connections, opposition, unionisation can slow down the process.
2. What Does a Job Look Like and What Are the Risks?
Most jobs are significantly broader than the task of solving mathematical equations. If we can dissect jobs into mostly independent dimensions, we may be able to better compare human skills and capabilities to AIs. It’s easier to see what an AI can and can’t do on a narrow task. For example, we could decompose jobs based on these human capabilities:
This is somewhat arbitrary, we could add more granularity, or add further human skills, constraints or even values. These also have some overlapping, but thats not a problem. I argue that these five dimensions cover much of what’s important in fulfilling a job. So to see how much risk we have of automation, we may look at how AI capabilities compare to humans in these individual components. To do that, we should first find out what’s the distribution of these components for a specific job. We then may check current AI capabilities, and the trajectory of development in the domain. Given that, we can come up with a humans vs AIs score in each dimension. If we do that for every dimension, we may weight the scores with the distribution, and we will arrive at an overall risk estimation.
3. How Good Is Our Estimation?
Such a granular estimation may incorporate many of the factors described in the introduction. For example, it accounts for subtask level granularity. However, we’re also missing some aspects. The most important seems to be cost/benefit ratios: how much can be gained by the automation? That’s not part of the who-can-do-what question. Another aspect, which may be somewhat left out, is if there’s an intrinsic value in a human doing something. For example chess computers are substantially more capable than top human players, but “professional chess player” is still a thing, because most humans prefer to see humans in sports. We’re probably also missing out on crystallised intelligence: someone mastering their profession for decades is much less prone to replacement compared to beginners.
We might try to count for these factors using different weights, and modulate our job risk scoring based on that. To my knowledge there’s no good established weighting for these. In my model, I used some heuristics (a simple decision tree). This part is waiting for improvement ideas.
4. What Model?
Okay, if this reasoning sounds rational, we can do some calculations. But calculations are cognitive processing, and in this subdomain AI systems are already quite good. So here’s a prompt that describes this process. Copy this into a chat with a reasoning AI model, and ask at the end: Apply this methodology to the profession of [YOUR PROFESSION HERE]! You may add details about specific circumstances - it’s not the same when one is an investigative journalist or one writes to the obituaries section of a newspaper. I quantified human advantage on an integer scale of 1 to 10, one being no human advantage. (Humans tend to have much better instincts on such an integer scale, providing a small set of fixed choices we're familiar with from early childhood, compared to the real numbers of probabilities from [0, 1]. Also, by using integers, we quietly introduce a nonlinearity - we just created a perceptron layer with five neurons.) So the AI will come up with an estimation of the job composition, and estimations about how capable AI systems are, compared to humans, on all five dimensions. We should not leave these to the AI, but ask corrections based on what is known about the very specific job we’re reasoning about. We simply understand the composition of our roles better. We may also narrow down the human advantage estimations based on the more precisely defined skills we use. Then we might ask the AI to search for current AI capabilities, and research trajectories on those narrower scopes.
5. The Results
Given this process, we reason step by step through our job security. We might ask the AI to modify the results according to our views about external adoption factors, and also about our estimations of plausible timelines. Interpreting the results is still somewhat arbitrary, but it will incorporate our best judgements across a reasoning process, mixed with near state of the art information retrieval from the world. The results are also somewhat stable: it won’t be too easy to cheat undetected, if we wanted to. However, we can gain useful information from looking at the reasoning process, and tweaking the model. We will see that we have more advantage in some skill dimensions, and less in others. This can work as a guide, as having more of those in our job description will improve our resilience.
6. Closing Words
I’m very curious about your experience and your thoughts about this process. Please share them!
I also wrote a shorter article on the EA Forum about how this came about. There are also three example calculations with notes in one page PDF files (my personal estimations from early 2025 for construction workers, software developers and dentists).
If you think this is useful, I have a Manifund proposal for turning this into a web app. I would appreciate an upvote there.
Technically, we're applying a single layer perceptron to the problem, with weights and biases taken from our beliefs. You don't have to understand the previous sentence at all. Let's begin by exploring these beliefs.
1. From First Principles
Let’s imagine your job consists of nothing else than solving mathematical equations, and you work along 99 colleagues doing the same (think “computers” from the 1940s). Given AI capabilities reach 50% success rate on 8 hour long tasks in your domain (let’s call it a workday), and the costs of doing so become less than your wage, your employer becomes financially incentivised to implement automation, let 50 of you go, and hire a few to oversee the AIs (a distinctly different job). In the case that AIs reach 99% success rate, your whole department will consist of one person solving the equations, and a few doing the oversight.
There are some adoption factors though. The AI system may do the same amount of calculations you do slower, at the same pace, or faster. Except for the much slower scenario, we can incorporate this fact simply in the cost component. Then comes the question of subtasks. Say all the tasks AIs fail in can be subdivided into two subproblems, and they fail only in one of those. That would reduce the work that still has to be done by a human to half. This is the edge case of the “AI enhanced workforce”, where people using AI are capable of doing more work than the ones not using it. If 3 out of 4 workers improve productivity by a third, the fourth worker becomes unnecessary, given flat demand. On the other hand, implementing an AI supported workflow may have an upfront transformation cost, that may slow down adoption. And there are other adoption factors than pure cost: workplace social interactions and connections, opposition, unionisation can slow down the process.
2. What Does a Job Look Like and What Are the Risks?
Most jobs are significantly broader than the task of solving mathematical equations. If we can dissect jobs into mostly independent dimensions, we may be able to better compare human skills and capabilities to AIs. It’s easier to see what an AI can and can’t do on a narrow task. For example, we could decompose jobs based on these human capabilities:
This is somewhat arbitrary, we could add more granularity, or add further human skills, constraints or even values. These also have some overlapping, but thats not a problem. I argue that these five dimensions cover much of what’s important in fulfilling a job. So to see how much risk we have of automation, we may look at how AI capabilities compare to humans in these individual components. To do that, we should first find out what’s the distribution of these components for a specific job. We then may check current AI capabilities, and the trajectory of development in the domain. Given that, we can come up with a humans vs AIs score in each dimension. If we do that for every dimension, we may weight the scores with the distribution, and we will arrive at an overall risk estimation.
3. How Good Is Our Estimation?
Such a granular estimation may incorporate many of the factors described in the introduction. For example, it accounts for subtask level granularity. However, we’re also missing some aspects. The most important seems to be cost/benefit ratios: how much can be gained by the automation? That’s not part of the who-can-do-what question. Another aspect, which may be somewhat left out, is if there’s an intrinsic value in a human doing something. For example chess computers are substantially more capable than top human players, but “professional chess player” is still a thing, because most humans prefer to see humans in sports. We’re probably also missing out on crystallised intelligence: someone mastering their profession for decades is much less prone to replacement compared to beginners.
We might try to count for these factors using different weights, and modulate our job risk scoring based on that. To my knowledge there’s no good established weighting for these. In my model, I used some heuristics (a simple decision tree). This part is waiting for improvement ideas.
4. What Model?
Okay, if this reasoning sounds rational, we can do some calculations. But calculations are cognitive processing, and in this subdomain AI systems are already quite good. So here’s a prompt that describes this process. Copy this into a chat with a reasoning AI model, and ask at the end: Apply this methodology to the profession of [YOUR PROFESSION HERE]! You may add details about specific circumstances - it’s not the same when one is an investigative journalist or one writes to the obituaries section of a newspaper. I quantified human advantage on an integer scale of 1 to 10, one being no human advantage. (Humans tend to have much better instincts on such an integer scale, providing a small set of fixed choices we're familiar with from early childhood, compared to the real numbers of probabilities from [0, 1]. Also, by using integers, we quietly introduce a nonlinearity - we just created a perceptron layer with five neurons.) So the AI will come up with an estimation of the job composition, and estimations about how capable AI systems are, compared to humans, on all five dimensions. We should not leave these to the AI, but ask corrections based on what is known about the very specific job we’re reasoning about. We simply understand the composition of our roles better. We may also narrow down the human advantage estimations based on the more precisely defined skills we use. Then we might ask the AI to search for current AI capabilities, and research trajectories on those narrower scopes.
5. The Results
Given this process, we reason step by step through our job security. We might ask the AI to modify the results according to our views about external adoption factors, and also about our estimations of plausible timelines. Interpreting the results is still somewhat arbitrary, but it will incorporate our best judgements across a reasoning process, mixed with near state of the art information retrieval from the world. The results are also somewhat stable: it won’t be too easy to cheat undetected, if we wanted to. However, we can gain useful information from looking at the reasoning process, and tweaking the model. We will see that we have more advantage in some skill dimensions, and less in others. This can work as a guide, as having more of those in our job description will improve our resilience.
6. Closing Words
I’m very curious about your experience and your thoughts about this process. Please share them!
I also wrote a shorter article on the EA Forum about how this came about. There are also three example calculations with notes in one page PDF files (my personal estimations from early 2025 for construction workers, software developers and dentists).
If you think this is useful, I have a Manifund proposal for turning this into a web app. I would appreciate an upvote there.