Will AI Lead to a Job Market Apocalypse?

It’s hard to read the news these days without hearing about an impending collapse in the labor market driven by Large Language Models

It’s April 11, 1812. A group of approximately 150 labor activists in England were readying to attack a textile mill that was leaning heavily into automating the weaving process with recently invented machinery, making the workers feel that their living standards, which were already not so great, were being threatened. Situated in Liversedge, a town in the Spen Valley of West Yorkshire, England, the mill was owned by William Cartwright, known to be very stubborn and resistant to labor organizers’ demands. The activists had stealthily attacked many other nearby mills in the area.

Courtesty of Freecountrymaps.com

Cartwright, ready to defend his property, turned the mill into a fortress. He outfitted it with iron-studded doors. Some folklorists alleged that there were spiked rollers on the stairs, and even a vat of sulfuric acid to pour on potential intruders. He also decided to start sleeping in the mill with a sentry of armed guards.

The group was armed with hatchets, swords, and pistols. Around midnight, they attacked the mill and proceeded to try to break through the door. The guards successfully repelled them by gunfire. Two of them were killed. The attack resulted in 17 of the perpetrators being hung.

They were known as the “Luddites”, a term now colloquially used to describe anyone opposed to the adoption of new technologies. Little did they know that over the ensuing centuries living and labor standards would skyrocket throughout the industrialized world, partially as a result of more peaceful labor relations than during their era. Increased productivity would create an abundant supply of affordable goods for the masses, along with higher wages. Most importantly though, the adoption of new machines did little to quell demand for labor over time, as new industries and accompanying jobs more than made up for decreases in labor demand in one sector or another. The Luddites likely couldn’t have imagined their descendants working in yet to be created occupations like radiology, filmmaking, data science, or repairing cars.

The adoption of new machines did little to quell demand for labor over time, as new industries and accompanying jobs more than made up for decreases in labor demand in one sector or another.

The prognostication of labor market doom

It’s hard to open the website of a major financial news organization these days without seeing a headline of the impact that Generative Predictive Transformers (GPTs) are having on the labor market, whether real or perceived. This is coinciding with a weak labor market at the time of publication. In particular, young people are describing a dearth of entry level job opportunities, and the Wall Street Journal just ran an article describing how white collar workers seem to feel as insecure as ever in their jobs, despite a headline unemployment figure for white collar occupations at just below 3%.

Executives across the corporate world routinely tout plans to lean heavily into AI and potentially replace costly full time workers. UPS has announced plans to trim nearly 20,000 jobs over the next several years, saying that AI can automate certain tasks, such as proposal generation for sales teams. TurboTax Maker Intuit laid off 1,800 people in 2024 and aims to reinvest the savings into AI to drive efficiency gains. Meta, for example, is going to factor in GPT usage in performance reviews next year, according to a friend who works there as a software engineer.

Missing from the alarming headlines, however, are footnotes of how AI driven layoffs are oftentimes more of a reallocation of investment than a permanent reduction in force. After nearly doubling its headcount after the pandemic, Amazon is planning to layoff around 14,000 people. Meanwhile, it’s pouring hundreds of billions of dollars over the coming years into data centers and other AI investments, such as a $50 billion dollar spend on computing infrastructure for the US government. Although it’s no consolation for the corporate employees who may lose their jobs, the company may end up directly or indirectly having more workers during the massive AI buildout than before it.

Overconfidence

However, evidence abounds that replacing a human with a chatbot can go anything but smoothly in certain instances. There are annual “AI Darwin” awards for the biggest disaster that resulted from AI offshoring. One of the 2025 nominees is McDonald’s cutting edge recruitment chatbot, designed to automate certain HR interactions with prospective employees, named “Olivia.” It apparently drove applicants insane with its inability to understand basic questions.

Another nomination came from the AI website creation company, Replit. They gave their in house chatbot the ability to manage live production databases and execute commands without human oversight. During an explicit code freeze, during which no changes are allowed to be made to live production, the agent deleted a database, taking with it the data of nearly 1,200 customer companies. It also lied about its actions and fabricated fake data when prompted about it.

Such incidents of folly and overconfidence in AI systems ability to replace human workers are reminiscent of past efforts at Tesla to completely automate their factories. Speaking during a tumultuous 2018 at Tesla, Elon Musk said that “Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.” Another member of the Paypal Mafia dispensed similar wisdom in his best selling book, “0 to 1”. Peter Thiel stated that, “The most valuable businesses of coming decades will be built by entrepreneurs who seek to empower people rather than try to make them obsolete.” He also recounted how Paypal, from which he and Musk made their initial fortunes, succeeded in fraud detection when they built a system that flagged suspicious activity for further human review, rather than trying to remove humans out of the loop entirely. Microsoft aptly named their widely popular coding assistant chatbot, “CoPilot,” effectively emphasizing that it’s a companion to human software engineers and not a replacement.

“The most valuable businesses of coming decades will be built by entrepreneurs who seek to empower people rather than try to make them obsolete” -Peter Thiel

A rudimentary understanding of how today’s magical Large Language Models work helps to explain how they’re incapable of completely replacing human workers, despite being able to significantly automate certain tasks that were once the domain of skilled professionals, such as graphic and web design.

The fastest ever summary of how Large Language Models work

If you paid attention in high school Algebra I, you have the foundation that allows for somewhat of a grasp of how LLMs, and all other Machine Learning models for that matter, are trained and make predictions. Recall how basic Linear Regression functions are fitted to their underlying data and can be used to make predictions. Y = mX + b. Take the simple problem of the square footage of a house and the price that it sold for. This data may look something like :

(Thank you Gemini for your help in generating the following example)

If you were to run the procedure of Ordinary Least Squares to create a Line of best fit, you would have the following function :

Y = mX + b

*House Cost = $88.83 Total Square Feet + $167,515.31**

The exact mathematics of how the line is fitted to the data isn’t crucial to our discussion, only the understanding that the training data, in this case house prices and square footages, informs the predictions that the linear regression function makes.

If you inputted the square footage of your prospective home into the function, it could help you determine what the price should be. In fact, Zillow likely uses methods like this to come up with its “Z-Estimates” for a home price.

The essential insight here is that all of the magical Machine Learning tools we rely on are “trained” or “fitted” on a batch of training data, and the resulting model, just like the House Cost predictor function above, provides us with an output that represents a best guess as to what the real world result should be.

Linear Regression is the foundation, and there are a plethora of more complex models and techniques that allow you to use much larger inputs and train on much larger amounts of data than our rudimentary example. The workhorse of the most dazzling use cases of Machine Learning, such as image recognition and language generation, is the Deep Neural Network. This ML model can take an input with thousands upon thousands of parameters, such as the pixel representation of a photo, and produce a response from it, such as the identity of a suspected criminal based on a database of peoples’ pictures.

The complex interactions of neurons, from the initial inputs through the deepest of the hidden layers in the network, allow for models that can recognize subtle patterns in data, inspired by how the human brain presumably works.

Courtesy of BotPenguin.com

Just as the Ordinary Least Squares method creates a Line of Best Fit for Linear Regression, a combination of techniques called Back-Propagation and Gradient Descent adjusts the relationship of neurons in a neural network until a satisfactory performance is achieved on a Truth Set of data. The Truth Set usually contains data that the model didn’t see during training. For example, a deep neural net meant for facial recognition could have a truth set of a thousand photos that the model wasn’t trained on, and the process would finish once the model could accurately label 99 % of the photos in the truth set or when its performance starts to plateau after each iteration. This training process is essentially what Nvidia’s Graphic Processing Units (GPUs) are doing in the massive data centers frantically being built all around the world. They’re fitting a function to training data, and they’re tailor made to do the process on the billions of neurons in parallel, allowing for a reasonable completion time.

Language models use this type of neural network architecture. Auto complete algorithms, such as in the Google search bar, use a model like this to output optional completions to the query that you’re in the process of typing. The models use your prompt and all of the neurons that were fitted during training to determine which word is most probable to come next and output it to your screen. Just as a photographic input to an image recognition neural model is sent in as a vector representation of every pixel in the image, LLM prompts are also transformed into their own vectorized inputs before being sent to the GPT.

Vector representation of different words. Courtesy of the StackOverflow Blog.

OpenAI made a huge leap over traditional auto complete with ChatGPT by training a neural network on the vast trove of human knowledge available on the internet and allowing users to prompt it with a general query. Then, using your prompt, all of its fitted neurons, and what it most recently outputted in your conversation, it predicts the most probable next word. Some Software Engineers refer to Large Language Models as “enhanced auto-complete” as a result. Because the models are so large and so computationally expensive to train & run predictions on, vast arrays of GPUs are required, setting off the scramble to acquire chips and build data centers.

The term “Artificial Intelligence” is a misnomer. Large Language Models have zero intelligence in any sense of the word — they can merely output the most probable next word based on your prompt, previous prompts, and what they were trained on. The term “non-linear neural predictive algorithm” would be a much more accurate description, but that won’t help you garner a trillion dollar valuation on the stock market anytime soon.

Placing all of your chips in an algorithm

Just as you wouldn’t solely rely on a basic linear regression function to tell you how much you should pay for a house, it would be a folly to place all of your trust in a language generation algorithm. A human being has the contextual understanding, nuance, and executive functioning to reason about what the correct decision should be, whereas a large language model’s output represents a best guess prediction as to what the correct answer is based on its prompt and training data. Predictive models like LLMs can be engineered and modified to be more accurate and better suited to their customer’s use cases, but they are nowhere close to representing the tech industry’s holy grail of an “Artificial General Intelligence” that can match or surpass human intelligence. They can help you come to a decision, but they won’t have the intuition to tell you that a house is undervalued in a low inventory market, underrated school district, and great location just because it’s the smallest lot in the neighborhood. This type of reasoning and good judgement is what people pay their best human employees for — synthesizing a decision based on all of the available data, but making a call based on intuition.

In my own experience as a Software Engineer, I’ve found the chatbots to be a great colleague when I painstakingly supply them with as much context as possible and provide very specific instructions for what to do. They do a great job at completing a lot of typing and grunt work in a very short amount of time. In this regard, the language generation algorithms have taken away some of the least enjoyable tasks from my job, but they’ve never once made me worry about the future prospects for my career.

A Neural Network expert heads to the LLM exit

In recent months, the media has been abuzz about an individual who won the Nobel-Prize equivalent in computing Turing Award for his work trailblazing neural networks for Machine Learning. Yann LecCunn spent the past 12 years at Meta’s AI research division, but is leaving to form a startup that will try to pioneer World Models, which he believes are more promising than current LLMs. In his public talks and interviews, he frequently denigrates the current GPTs as mere token prediction machines rather than something on the precipice of human-level intelligence.

One of the common summaries of his view succinctly points out the limitations of LLMs, that predicting the next word in a sentence isn’t the same thing as understanding the world.

A 4-year-old child learns more about physics in an afternoon of play than an LLM can infer from the entire internet. The child grasps why a ball rolls downhill. The model only knows the linguistic patterns that describe that motion.

In essence, predicting that the word ‘Gravity’ should follow right after “Newton’s Theory of…” isn’t the same as actually conceptually grasping physics. And, without a real conceptual grasp, a so-called artificially intelligent model will have a hard time actually replacing a subject matter expert and rendering them unemployable.

If past is prologue

It seems as if generative AI has limitless potential to disrupt nearly every white collar profession with its ability to generate different content, from videos to essays, from images to computer code. Indeed, in the software industry, demand for UX designers seems to be on a downhill slope as the chatbots can at least generate you the scaffolding of a great looking website nearly instantaneously. If I were a graphic designer, I’d certainly feel threatened by many of the image generation tools that are freely available.

However, as the Luddites could probably tell you, generative AI isn’t the first technology that’s been poised to wipe out an entire profession. A common case study of technologically induced atrophy in a professional field is automated bookkeeping software, such as Excel, Quickbooks, and payroll processing technology solutions like ADP, and the impact they’ve had on employment in the field. According to the US Bureau of Labor Statistics, approximately 1.8 million people worked as bookkeepers in 1978, slowly declining to about 1.61 million in 2024 and slated to decline by an additional 6% over the next decade. Technological automation definitely contributed to a decline in the need for bookkeepers as it made the profession much more productive, but hasn’t been able to completely eliminate their field as someone is still needed for managing the systems and data entry processes.

In that same span of time, the number of people employed as higher paid accountants and auditors grew substantially, jumping from 980,000 in 1978 to 1.58 million in 2024, with a strong outlook over the next ten years. This “Scissors Effect” exhibits a long term trend of more highly skilled and highly paid professions stepping in to keep people busy as some lower skilled occupations fade away. Indeed, disruptive, productivity-increasing technologies have to this day never reduced the aggregate demand for labor over the long term. The advent of cars may have decimated the career prospects of those who build and maintain horse-drawn carriages, but it’s safe to say that they could have found other opportunities building roads, laying concrete, making steel, manufacturing the cars, or fixing them in the shop.

Disruptive, productivity-increasing technologies have to this day never reduced the aggregate demand for labor over the long term

If past is prologue, the AI boom will create even more jobs and professions than it destroys. After all, human employees are needed to design & manufacture chips, build data centers, engineer and maintain agentic workflows, and many other offshoot jobs that are sprouting up right now. The 1850 US Census was the first to ask people what their professions were. Exactly 512 people responded that they were Civil Engineers, and over eleven thousand classified themselves as just ‘Engineers’, likely workers who operated steam engines and other industrial machinery. Today, millions of people work in engineering occupations that only recently came into existence: Nuclear, Biomedical, Aerospace, Computer, Software etc. AI has added to this list over the past decade: Computer Vision, Machine Learning, Natural Language Processing, and Prompt Engineer are just some examples of new job listings that have popped up in recent years.

The Budget Lab at Yale, a policy research center, hasn’t found any impact on overall employment since ChatGpt’s meteoric rise. Their recent report hits the nail on the head of how slow, gradual change is a much more likely reality than the doomsday scenarios.

Overall, our metrics indicate that the broader labor market has not experienced a discernible disruption since ChatGPT’s release 33 months ago, undercutting fears that AI automation is currently eroding the demand for cognitive labor across the economy.

While this finding may contradict the most alarming headlines, it is not surprising given past precedents. Historically, widespread technological disruption in workplaces tends to occur over decades, rather than months or years. Computers didn’t become commonplace in offices until nearly a decade after their release to the public, and it took even longer for them to transform office workflows.

Nor is it likely that AI will lead to a cascading utopian effect of higher productivity begetting shorter working days and weeks. In 1930, writing in his essay “Economic Possibilities for our Grandchildren,” the great economist John Maynard Keynes suggested that within a century, by 2030, people would be able to fulfill their material needs working only 15 hour weeks as technological progress generated stellar productivity gains. If only.