Comments
A Taxonomy of Jobs Deeply Resistant to TAI Automation
This is a light, informal taxonomy of jobs that conceptually will display significant resistance to automation, even after the diffusion of transformative AI systems that can outperform humans in all forms of cognitive and physical labor.
This is intended to provide practical, non-technical context around plausible human jobs in a post-AGI economy. In particular, we would like this article to challenge two prevailing narratives from opposite ends of the aisle:
Some analyses have suggested that the development of TAI could mean that “human labor share could eventually fall to zero”, or that TAI indicates that basically no cognitive labor jobs would remain for humans. We believe this is broadly inaccurate, and provide specific examples of competitive advantages that enable human labor even after TAI systems become fully diffused.
On the other hand, many economists and much of the political ecosystem is operating under the assumption that AI will be primarily augmentative, not automative, and that we do not need to be overly concerned about massive labor displacement. We also believe this is a deeply limited worldview and fails to consider the possibility of TAI existing and having broadly automative capabilities. Assuming that TAI will exist and can automate most cognitive and manual labor, we will illustrate how we expect even the most “TAI-resistant” jobs described to be massively impacted.
This article will illustrate a potential long-term future for human jobs after transformative AI has successfully diffused deeply into an economy.
This article will contain several assumptions:
Several things we will explicitly not tackle in this article:
This is not intended to be complete or a fully comprehensive view of jobs that will exist after an AGI transition. Rather, this is simply a snapshot of certain types of human jobs that will still see competitive advantages when compared to highly capable transformative AI systems. We hope that this article will support a more well-rounded discussion of what an economy after AGI fully diffuses will look like.
We acknowledge the limitations around defining future AI systems, and in this article will use “transformative AI” as a stand-in for “AI systems that can outperform humans in all forms of cognitive and physical labor.” Note that this does not mean such systems will be comparatively cheaper than humans, which is necessary for diffusion.
This article will not concentrate on human jobs that will remain solely because of comparative advantages and low cost of labor. For example, we might expect that in developing countries, it will take substantially longer for AGI to diffuse into the economy because the comparative costs of human labor will remain much lower for a significant amount of time. Instead, this article will focus on the near-complete diffusion of AGI into a developed economy such as the US.
Finally, this article was intended to be a light thought exercise to explore how TAI may reshape the job market. We acknowledge the significant uncertainty around what the economy will look like in the future and suggest that this article represents a plausible, but not guaranteed, view of the future.
Roles where humans will add value by providing highly-capable AI systems with the intent and specifications (derived from humans) required to create the correct output.
Examples: software developers, UX designers, IT, data scientists, project managers, marketing strategists, customer success managers.
For example: An “TAI programmer system” for a software company may not have the capability to interface with a human CEO, participate in in-person stakeholder meetings, or interface with human customers in order to understand what features or design decisions need to be improved. A skilled human programmer may be required to handle internal and external communications and describe to the AI system exactly what intent and specifications it must create.
Even if such a human does not need to touch a single line of code themselves, they must understand the structure of their software system deeply to be able to communicate accurate intent, thereby requiring some technical programming knowledge.
Competitive Advantage: Humans will maintain an advantage in interfacing with both internal and external human stakeholders. They will be able to attend meetings, interview customers, and synthesize information that is shared offline for the AI systems that do the “heavy lifting”.
Long-Term Prognosis: Eventually, performance will improve and costs will drop for AI systems such that they do the bulk of the “day-to-day” technical work. Intent Communicators will largely interface with such AI systems via plain English. Significantly fewer humans will be required to manage complex systems, and they will largely be “senior”, understand the entire system on a high level, and make highly competent decisions based on their intuition and judgement. Few opportunities will exist for “entry-level” humans in these roles.
Over time, even these “intent communication” jobs may erode as stakeholders and customers become more comfortable interfacing directly with AI systems, or more important conversations move fully online.
Automation & Demand Estimates: Perhaps ~40-80% fewer humans will be required to operate an equivalent system (e.g. the programming division of a company) that is AI-driven. This will vary widely by industry: quant funds may be significantly more automatable than retail companies, for example. However, many more “systems” will likely exist due to reduced operational costs and consequent higher demand. This may offset labor displacement.
Roles where humans have some sort of competitive advantage in interfacing with other humans: for example, roles that require deep empathy, emotional intelligence, or human connection.
In this category, human presence can play a meaningful part of someone’s experience and has unique value. All of these roles historically have been centered on in-person human engagement.
In the future, these will be roughly sorted into jobs that can be performed online, or virtualizable roles (e.g. therapists), and jobs that mostly cannot be performed online (e.g. sports coaches), or embodied roles. There will be significant gray areas between these categories.
Competitive Advantage: It will be a challenge for AI systems to effectively mimic human empathy, emotion, eye contact, and presence – especially with in-person robotics. Even when they are able to mimic humans, there will be a certain level of persistent distrust and alienation from human customers when interfacing with AI systems. Humans will maintain some built-in advantages on these traits for a significant period of time post-TAI.
These are interpersonal roles that were traditionally performed in-person, but have migrated to have significant online presence. AI systems will eventually become highly competitive at providing online services, though they will still have significant weaknesses in face-to-face interactions.
Examples: travel agents, life coaches, therapists, consultants, online tutors.
Long-Term Prognosis: There will be widely available online & automated versions of these roles. Access to an online AI version of these tools will be widely available and extremely cheap - perhaps 5x - 50x cheaper than an in-person option.
At first, the AI versions of these might be limited (e.g. text-based, slow input / output lag, not as nuanced). They may be limited or mediocre compared to top humans. Over time, the AI versions will advance to becoming as effective as a human interfacing with you over a laptop. At some point in the future, most AIs available will consistently be indistinguishable from the best humans, and quite often better than them.
Only the most effective humans will retain a competitive advantage to stay on par or better than AI systems in the long run, either through a human touch, in-person meetings, or both. As a result, humans will become a luxury good: unnecessary and highly-priced for the majority of use cases. However, there will be a persistent set of excellent humans who will see significant demand in these categories, leveraging their unique human traits to drive business.
Automation & Demand Estimates: Significantly increased demand as access becomes an order of magnitude cheaper. For instance, perhaps 80%-90% of therapy requests will go through AI systems. Simultaneously, AI therapy will be 5-50x cheaper than human therapy. Despite these cost reductions, demand for human interpersonal specialists may remain high if people strongly value the “human element” beyond what AI systems can provide.
These are interpersonal roles that are primarily performed face-to-face, rather than online. As a result, purchasing the robotics required to replace these tasks must fall below the marginal cost of hiring a human, creating a significant infrastructural barrier.
Examples: nannies, primary school teachers, social workers, sports coaches, religious leaders, political translators, hospice caregivers, DC lobbyists, sommeliers.
Additionally, a robotics system inherently will have significant limitations compared to a human in these roles: e.g. the inability to make human eye contact, or to communicate human empathy or emotions as effectively.
It will be a very long time before an in-person AI robotics system can match human standards for these traits. An AI robotics system that attempts to do so will likely fall deep into the uncanny valley. Even if it perfectly mimicked human behavior and presentation, simply the knowledge that it is an AI will cause distrust and alienation. Any such AI robotics systems will face a massive uphill battle to become productionizable and cheap at scale.
Automation & Demand Estimates: Strong inherent resistance to automation due to human competitive advantages. Relatively minimal automation until certain thresholds of robotics costs are achieved. Physical supply chain constraints slow adoption even after thresholds are reached. Demand likely increases due to increased human labor supply and other augmenting factors (e.g. an aging population necessitating more healthcare specialists).
Roles where humans are unwilling to hand over major / final decision-making to AI systems, and which require human oversight. This could be for ethical, safety, or regulatory reasons, or simply have to do with human trust in AI systems, regardless of their performance.
Examples: judges, legislators, executives, military commanders, law enforcement, financial regulators, leadership roles, safety supervisors, government officials.
There are many roles in which humans will inherently distrust AI systems to take over responsibilities, at least for a period of time. For example, there will likely be significant human biases towards human legislators or policy-makers, relying on the underlying belief that humans will best represent the interests of other humans from their region (representative democracies already function according to this model).
Other examples:
Competitive Advantage: Humans will continue to occupy these roles based on underlying distrust of AI systems, beliefs that humans will keep the best interests of other humans at heart, and concerns about the reliability and performance of AI systems in outlier situations. These concerns may steadily reduce over time, reducing human competitive advantages as well-aligned AI systems prove their reliability and lack of bias.
Long-Term Prognosis: The majority of these jobs will see strong initial resistance to automation, due to underlying human biases, lack of reliability, reluctance to give more power to AI developers and systems, and other compromising features of AI.
Over time, society will become more acclimated to the reality that highly capable AI systems will be capable of taking on these roles and making good decisions. Automation will steadily and slowly increase over time, with adoption happening at a much slower pace compared to the rest of the labor market. Significant societal debates will need to occur for many of these job roles.
Many of these roles will see society decide that they should be permanently occupied by humans. We will not permit full automation by AI systems for some of these roles, either from a legal standpoint or via societal pressure.
Automation & Demand Estimates: Relatively consistent demand, as many of these roles are not driven by the free-market. Strong inherent resistance to automation. A slow thawing of resistance over time leading to increased long-term automation. A final steady state that still sees significant human presence in these roles.
Roles where consumers find value in the conceptual belief / understanding that a person produced the output. Can relate to storytelling, empathy, or “human” narratives.
Examples: Certain singers and artists, actors, book writers (in particular, memoirs and autobiographies), creatives, painters, handcrafted artisans.
Humans have a strong demonstrated affinity for consuming compelling human narratives and non-fiction storytelling. They may also have an underlying preference for human-created art over AI-generated art, because of the human connection and emotional investment underlying such creative work.
TAI systems could conceivably become significantly better than humans in domains like fictional storytelling, technical songwriting, or digital art. However, they will not be able to create genuine narratives sourced from human experience. There will still exist consumer demand (as there is now) for compelling real-world narratives that result in music, art, books, and movies that deeply reflect the human experience.
Similarly, there will remain demand for human actors, even if such actors can be perfectly recreated by AI systems. There will remain demand for human artists, even if AI-generated art is technically superior.
Because AI systems will be able produce an overwhelming amount of creative content so quickly, there will inevitably be some human demand for fewer, more tailored, and “real” human narratives that they can connect to.
Competitive Advantage: The knowledge that some forms of art were created (largely) by humans, or at least reflect the lived experience of certain humans, will continue to be a competitive advantage for certain writers, storytellers, thought leaders, artists, actors, and more.
Long-Term Prognosis: Despite the resilience of certain Authentic Creatives, large portions of the creative industry will eventually see drastic automation from AI. Many forms of art may not rely on the human experience (for example, certain genres of EDM production are primarily based on technical skill). In domains where consumers do not care about exploring the underlying human connection, we can expect to see significant long-term automation.
Even for Authentic Creatives, there will be substantial augmentation via AI systems. A pop-star with a compelling human narrative may still use AI to generate much of their music but release it under their own name, for example. Non-fiction writing will be significantly improved and edited by AI systems. The creative process for all content will substantially leverage AI in the future.
Automation & Demand Estimates: Because AI content will be so easy to generate, we might expect that 95% - 99% of all future content will be AI-generated, flooding the market. However, content created (or at least affiliated) with humans will occupy a disproportionate amount of attention and revenue of the societal zeitgeist. A large proportion (and perhaps the majority) of “popular” content will still be sourced or affiliated with human creators. Overall consumer demand for content will be largely dependent on the discretionary income of humans.
Roles where the complexity of physical tasks is not worth the resources to develop specialized robotics for. May include some aspects of “handcrafted” value, or simply be a highly specialized trade.
Examples: custom furniture makers, violin makers, specialized equipment repair technicians, handcrafted goods (e.g. Etsy), custom tailors, restoration specialists.
Many niches of physical jobs for artisans require substantial amounts of specialization, high expertise with significant learning curves, unique physical manipulation skills, regulatory or certification hurdles, or adaptability to variability for each job.
These requirements in and of themselves may not be a strong moat against TAI automation, which could conceivably develop these skills over time. However, small market size with limited scaling potential may make it financially imprudent or implausible to develop the specialized robotics and AI systems required to automate these jobs. Even if developed, costs to deploy such low-volume robotics systems may remain higher than the marginal cost of human specialists.
Competitive Advantage: Certain artisans currently have highly technical, niche physical skills that may not make financial sense to teach robots to emulate. They may have developed relationships and trust over time with consumers that may be resilient to automation, especially in niche domains. Most importantly, developing and deploying specialized robotics at scale will continue to be expensive even after TAI exists.
Long-Term Prognosis: These jobs will vary highly by domain, ease of automation, demand, and the costs of robotics equipment to effectively automate certain tasks. They will see strong initial resistance to automation as more high-volume and high-profit physical jobs are automated first. Over time, as the cost of robotics equipment drops, they will see relatively more exposure.
However, they may also see fundamental resilience as human consumers place additional value on the “handcrafted” or “artisan” aspects of these jobs, much as handcrafted furniture or goods can ask for a higher selling price than machine-crafted equivalents.
Automation & Demand Estimates: This is difficult to predict across a wide range of diverse artisan jobs.
Roles where the complexity or diversity of physical tasks remains challenging (or prohibitively expensive) for robotics to imitate at scale.
Examples: Certain construction workers, firefighters, plumbers, EMTs, choreographers, dentists, policemen, airplane manufacturing technicians, surgeons, massage therapists, wilderness guides, janitors.
Even accounting for a future where robotics can technically automate any form of manual labor, there will likely be categories of jobs where dexterous human labor remains competitive economically with robots for a very long period of time. This will occur for two reasons:
This can be easily seen with construction workers, who operate in highly demanding, physically intense environments for relatively low pay, while conducting a wide range of tasks that require both precision and adaptability. Though many aspects of construction may see further automation (e.g. operating heavy machinery), there will continue to exist construction tasks that are cost-prohibitive to perform with specialized robotics.
Competitive Advantage: Humans are in high supply relative to robotics, and may come with significantly lower startup costs in certain domains – because they do not need to be built and paid for corporations. They may be cheaper to maintain, more adaptable to different tasks, and able to function in more extreme environments for long periods of time. They may have competitive advantages in certain skills (e.g. massage therapy) or provide interpersonal engagement benefits on top of dexterous manual labor (e.g. a wilderness guide).
Note that this may not be a permanent competitive advantage, but may simply be a question of “diffusion”: when robotics become cheap enough to replace humans.
Long-Term Prognosis: These types of jobs will see strong resistance to automation, corresponding almost directly to the cost of building, deploying, and maintaining robotics that can perform similar tasks. These costs will be dependent on the cost of raw materials, efficiency of supply chains, international trade policies, efficiency of manufacturing processes, and many more factors. Training the AI systems driving such robotics will likely be a less limiting factor, as it will be a one-time cost and have similar returns to scale as software / LLM development.
Once the marginal cost of building and deploying a comparable robot drops below that of hiring a human, we may see a rapid increase in automation in certain markets. Because many humans globally perform such tasks for extremely low salaries, we expect this form of automation to take decades to diffuse globally. However, we may see significant AI robotics adoption in certain sectors of developed countries much sooner than that.
Overall, we may see two major phases to TAI diffusing into the economy. The first will be when cognitive systems surpass human performance. The second will be when supply chains for general-purpose human-replacing robotics improve to the degree that they become cheaper than the average unskilled manual laborer (in a specific country). This second phase could take decades to reach, even if we develop TAI systems in the next few years.
Executive summary: While transformative AI (TAI) will automate the majority of cognitive and physical labor, certain job categories will persist due to human advantages in communication, trust, dexterity, creativity, and interpersonal interaction, though their structure and demand will shift over time.
Key points:
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.