What Are the Jobs Unlikely to Be Replaced by AI?
April 5, 2022 posted by Franklin Ren, largely revised on Oct 30, 2023
Ever since the AlphaGo moment, there has been a tremendous amount of discussion about the possibility of AI replacing humans for most of our current jobs. This anxiety of people losing their jobs has escalated like never before after 2022 when image-generation tools like Stable Diffusion and text-generation tools like ChatGPT pushed the availability of AI to another level. There is already news about parents stopping their children from learning visual arts as they believe those jobs will be replaced by AI when their children grow up. I am also asked by a lot of friends not from the AI community about what kinds of jobs are safer in today's world. Here are some of my thoughts. This is not a theory or a formula - but rather a general guidance for evaluating the “safeness” of your jobs during the rapid growth of the AI industry.
Accessibility of training data
Most of our current AI models still require a large amount of domain expertise data to train. Hence, the accessibility of training data in your job domain could be an important factor that determines whether it is easy for people to train an AI model that possibly replace you. Almost all successful AI applications (AlphaGo, AlphaStar, AlphaFold, ChatGPT) come from domains where there is already a substantial amount of domain expertise data online to be readily used for AI training. AI in domains where training examples are harder to collect tends to evolve more slowly. For example, child daycare robots evolve much slower than chatbots which tell children stories. The imitation learning of the former might require millions of child daycare videos which you could find nowhere online. If learned from scratch without any imitation first, the robot could hurt the children or itself, which makes it less deployable. In contrast, chatbots could simply train from millions of fairy tales online before being deployed.
Degree of customization
The degree of customization of your job is another important factor. If the product or service of your job is highly customized, it is harder for an AI to quickly generalize to the degree you want. For example, when facing AI paintings like Stable Diffusion or Midjourney, manga drawers' roles are less threatened compared to poster drawers. One key reason is that many manga have specific storylines and the painting in the manga is highly customized, even a small detail will make the story completely different. Besides, most manga drawers' works have a strong personality which attracts a particular set of customers. In contrast, posters are usually vaguer and more generalized, where a slight change in detail usually doesn't change the entire topic. That will make it much easier for AI to generalize.
.png)
.png)
.png)
Acceptable level of risk
How bad can it be when you mess up your job? This is another crucial indicator of how possible a job to be replaced by AI. If one mistake in your job will cause catastrophic results, then it is less likely that your job will be replaced by AI. For example, despite those automatic tools, pilots, surgeons, and accountants are not largely replaced by AI as the nature of those jobs needs almost 100 percent accuracy. Even 99.99 percent accuracy (which is considered to be very good for current AI tools) is bad enough for those jobs. One general guidance here is that AI generative content (e.g., image generation, chatbot) tends to be less risky than AI predictions (e.g., physical object detection, stock price prediction). If AI makes a mistake in the former one, we could just throw it away and try again, but if AI makes a mistake in the latter one, more likely there will be some direct bad consequence that we cannot recover from. Another general guidance is that NLP (Natural Language Processing) tends to be less risky than CV (Computer Vision) since NLP usually doesn’t require an interaction with the physical world. If something wrong happened to your Amazon personal recommendation, it is not a big deal, but if something wrong happened when you sitting inside a Tesla self-driving car, that could be horrible.
Cost of an employee
There are lower incentives for employers to use AI to replace lower-paid jobs. Easy-peasy.
Power in the society
A factor that stays highly neglected. Some jobs are theoretically replaceable by AI, but if that job community has a larger power in society, things could go differently. Many people predict that the recent text-generation AIs like ChatGPT have the potential to replace lawyers in the US, or civil servants in China, but I don't agree with that. While whether or not ChatGPT-generated law document is accurate enough is still questionable, a more essential factor is that the American Bar Association is a community with strong political influence in the US. Only 19 of the total 46 US presidents had not been a lawyer. With such strong political influence, it is hard to imagine they will push some bills that will allow AI to replace themselves. Yes, they might use AI to assist their work. But replacement? Maybe not. Things about civil servants in China are similar. All government officials in China started with civil servants, giving this group substantial political influence.
Skill shift
Whilst I mentioned in the previous paragraphs there are certain jobs unlikely to be replaced, (e.g., doctors, lawyers), there is indeed a high possibility of skill shift in those jobs. That means, some skills previously essential for certain occupations might not be relevant in the AI era, while other skills might become increasingly important. For example, since ChatGPT is so good at memorizing professional materials, in the future law schools might prefer students who perform better in debates than those who perform better in memorizing cases, and medicine schools might prefer students who are more talented in experimental work than those who earn higher scores on tests. Those kinds of skill shifts are something you definitely should take into consideration when you choose your job.
Low-skilled vs High-skilled
Here is the tricky part. There is a widely believed myth that for many jobs, AI could only replace those low-skilled workers, so high-skilled workers don't have to worry about anything. Something they missed out on is that there is hardly anyone who is born to be high-skilled in certain jobs. Most people, except for a few geniuses, start their work from a very premature level, feed themselves, and years after become proficient in their domains. If an occupation becomes one in which only high-skilled workers can make a living with it, most people won't consider entering it for the first time. Eventually, that occupation could become the playground for only a few top geniuses. Examples of such jobs include professional athletes, classical musicians, and theoretical physicians. Will painters join this group after Stable Diffusion and Midjourney? Maybe. What we know for sure is that there is already a large group of painters going on strikes to call for a ban on using artwork for AI model training. Many of them are big names. Maybe they know way much better than any outsiders what will happen to the entire industry if this is not stopped.