Why AI will not free people from work or lead to social equality

This is follow-up of my post “Why everyone is probably wrong about AI“, which I have split into two articles. A common prediction is that AI will make white-collar workers redundant or most of their jobs obsolete. Consequently, wealth inequality will fall. There is a sort of schadenfreude in seeing the tables turn against the elite. This comes off as wishful thinking.

What happened when computers became ubiquitous? Did wealth inequality go down due to automation? No, all the smart people got computer jobs and became wealthy that way. Now smart people are getting AI jobs and being paid similarly handsomely. Until robots can play sports or act in movies, top actors and athletes will continue to command huge paychecks, too. The AI ‘talent wars‘ has seen the largest of tech companies, notably Meta, write blank checks, not to acquire a company, but for a single person. If history is any guide, AI will not be a force for egalitarianism, but rather that the usual differences of human capital and the disparities these engender will remain intact.

It’s hard to think of a technology in which the relative status of the mediocre and untalented was elevated. Smart, well-connected people will just adapt to whatever the latest technology or hype-cycle is, because they are already starting at a major advantage anyway. People who are already smart and talented can quickly learn new skills and adapt–that is what makes them smart and talented in the first place.

After the ’08 crisis, bankers were quickly rehired as the stock market bottomed and recovered, but the housing market, especially homebuilders, took much longer, leading to prolonged unemployment for many blue-collar workers. Same for retail workers and other affected sectors. Shopping malls were hit especially hard by the decline of the commercial real estate market. By contrast, the tech sector boomed, with Facebook/Meta going public by 2011. Then came Covid, which was a double-whammy, again leading to mass unemployment for retail workers. By comparison, white-collar workers adapted easily by working from home.

Predictions of AI destroying jobs also are a dime a dozen. @ESYudkowsky writes:

I’m unconvinced this will happen. For several reasons: First, although AI does many thing well, its still falls short at some areas, like adversarial situations. Imagine trying to rank a book on the New York Times best-seller list, or trying to reverse engineer a competing algorithm, such as for high frequency trading or search engine optimization. Compared to repetitive tasks, these require having to anticipate what the adversary does and reacting in real-time. For example, Google and ad-blocking companies are at constant war. It’s hard to imagine how either side can outsource this task to AI without being at a disadvantage, as the algorithms on either side are always changing. Google is constantly changing its code to thwart ad-blockers, and ad-blockers have to modify their code in turn.

Second, if AI makes employees more productive, then this can justify the cost of keeping them from the perspective of the employer. It’s reasonable to assume that people who are smarter will be capable of using AI more effectively, as learning ability is correlated with IQ, allowing them to pull even further ahead. Tying this with social equality again, it’s those who are in the ‘fat middle’ of the distribution in terms of talent or skills who have the hardest time adapting to sudden economic or technological change, as seen with Covid , mentioned above, or the ’08 crisis. The plight of the ‘Rust Belt’ and the ongoing opioid epidemic in the Midwest makes this clearly evident.

Old technologies also have a tendency of hanging on. An example that immediately comes to mind is how Substack is booming despite also coinciding with the rise of LLMs, starting around 2022. One would naively assume that writers are especially vulnerable given that LLMs excel at the very sort of thing writers are good at (hence the epidemic of students using them, to the frustration of teachers and administrators), but these programs are unable (yet) to reproduce the voice, nuance, and branding of top writers. The ‘big four’ publishers and major media outlets continue to hang on, still defying the predictions and insistence of their imminent obsolescence.

This also ties in with adversarial situations. Even though AI can write well, trying to stand out among many other competing writers is a separate and even harder problem that requires different skills. If I ask Chat GPT “how do I get more traffic to my article or blog,” it will give me a generic list of suggestions, which is not that helpful. In this case, it’s like a slightly smarter version of Google. This may be fine for simple or easy situations, but large companies are not hiring consultants to use Google.

A common intellectual blind spot is thinking it has to be all-or-nothing, and that one must choose between the obsolescence of a technology or its survival. Quality human-made content can thrive alongside automated ‘AI slop’. As the post-2022 Substack boom and the thriving online literary scene shows, there will always be a market for ‘snobs’ or connoisseurs who seek high-quality content. But there is also a market for slop for a much less discriminating audience, such as anything crypto or finance related, which is well past the point of return.

It’s increasingly common to come across AI-generated articles where the headlines don’t match the actual data. For example, a headline like “Bitcoin falls below $100k” even though the price is currently $119,000. This mismatch is a telltale sign of AI generation, but the targeted audience either doesn’t notice or care. The same goes for AI-generated images and videos. The type of unsophisticated investor who is credulous enough to believe fake news of Warren Buffett endorsing AI, is not going to be deterred by AI slop:

As for liberating people from work, one thing that stands out is how rather than eliminating work, AI only shifts it around. People are doing the same amount of work, but only different types of work. The quantity of work still fills the allotted time despite AI. Rather than coding, developers are setting up run-time environments for AI, designing workflows, or modifying and combining AI-generated code. Model context protocols (MCPs) are a sort of middleman between the LLM and the desktop, yet any attempt at removing abstractions leads to new ones. The interplay between GPUs and AI is another type of job. App development is still time consuming and hard despite AI. There is also the learning curve of using the AI.

Consider the viral video below of a girl using AI. She’s still doing work at the computer:

In fact, she has two computers, as if one was not enough. Getting from the conceptual stage to the finished product, even with the help of AI, seems to still involve many intermediate steps. You have to learn how agents and MCPs work. You have to choose the right LLMs for the specific task in mind, as there are at least a dozen competing products to choose from, and costs can quickly spiral if you have many subscriptions running concurrently. There is also issues like rate-limiting and other constraints, which only become obvious after one has invested time and money. You have to know how to set it all up, and learn the interplay between all these ‘building blocks’. There are many active Reddit threads of people asking for help on using these programs. So much for freeing people from work.

In the ’90s to early 2000s, websites were typically coded by hand in HTML using a text editor. The mid 2000s saw the rise of site-builders and graphical editors, in which ready-made templates and ‘dragging and dropping’ made it unnecessary to have to write much code. Did building websites become faster? Not really, because instead of writing code, that same time was spent modifying templates or other aspects of the design process. Greater expectations placed on developers meant a bigger workload despite time-saving technology. Although websites looked nicer or had more features, the process of building them was still time consuming and hard.

This is not to say that AI will not make at least some tech jobs obsolete. The 2022-2023 ‘tech freeze’ is evidence of possible structural changes, although it’s hard to squarely blame this on AI, versus a sluggish economy and the post-Covid-stimulus hangover. Even if AI eventually leads to less work and more free time it doesn’t address the deeper challenge: how to live a meaningful and fulfilling life. In fact, those who seem to lead the most fulfilling lives, such as top YouTubers, podcasters, and CEOs, are really busy. When people retire or win the lottery, the initial excitement often fades, leaving behind a void of unstructured time that can feel empty.