People have been concerned with new technology’s impact on the areas of labor and work for as long as the concept of labor and work have existed. Take for example the humble plow, a tool that shifted the measure of a person’s ability to extract value from the earth from the amount of work they could do with their bare hands to the amount of land they owned and harvested. Had newspapers existed in the early agricultural age, proto-reporters surely would have pulled quotes from concerned furrow-diggers about what the future held in store for them.
Last week, there seemed to be a noticeable uptick in media coverage about AI and the impact that it will have on jobs in the future. Time Magazine, Axios, and The New York Times each had articles worth reading. Time laid out potential ways to address the upheaval many predict will occur while Axios spoke at length with Dario Amodei, CEO of Anthropic, about his predictions for a future full of broadly capable AGI systems and how there could be significant impact on jobs in just the next few years. The New York Times focused a bit more narrowly on the impact AI is already having on the job market for recent college grads.
In addition to these reading these articles, I saw the following 11-minute clip that stitched together lots of predictions.
In sharing this compilation, I’m not endorsing the viewpoints or predictions therein. Nor am I compelled by the broad and sensational headline “The Great AI Job Displacement Is Closer Than You Think.” It’s important to me to showcase the viewpoints of individuals that are working to run major AI labs, have vast experience in the AI space, and have paid close attention to recent progress.
Here are a few of the most staggering quotes from the video:
“I’m actually afraid of the world where 30% of human labor becomes fully automated by AI and the other 70%…that’s going to cause this incredible class war between the groups that have been and the groups that haven’t been” – Dario Amodei, CEO of Anthropic
“That doesn’t mean the transition isn’t going to be messy, in fact, I expect it in some ways to be pretty painful” – Sam Altman, CEO of OpenAI
“Psychosocially, it’s very disturbing that you can no longer tell people what kind of world they should prepare their kids or grandkids for.” – Tyler Cowen, Marginal Revolution
Each of these utterances portends a future with serious challenges. Potential for class war, an economic transition of indeterminate length and unpredictable levels of strife, and most of all a range of potential outcomes so vast that it becomes impossible to predict, let alone prepare for. Imagine if the messages in that clip didn’t come from the stages of conferences or the interiors of well-equipped podcast studios, and instead, you encountered them walking down a city street.
If I saw this man, I’d hastily move to the other side of the street, not letting his proclamations take up space in my mind. It’s easy to ignore sweeping statements about a vastly different and uncertain future, especially in a world that has so many ongoing, more tangible challenges. It’s also understandable to chalk these statements up as hype — chum for investors looking to secure a slice of the trillion dollar labor market these predicted drop-in remote workers of the future could dominate. Realistically, the likeliest outcome lies somewhere in between automating all white-collar work by 2030 and producing systems that are economically useless and therefore inconsequential.
Impactful Assumptions
This post isn’t meant to focus on my own personal predictions, or to dissect individual predictions made in the articles or video I’ve discussed so far. There are thousands of voices out there who have done that, and they’ll continue to make predictions as we integrate AI into our lives. Something I will do though is list out a few assumptions I think the people expressing the viewpoints covered in this entry are making that hugely impact both the timeline for and magnitude of AI’s impact on the job market.
Assumption #1 – Jobs can be completely decomposed into a discrete and finite set of tasks that have clear indicators of task success and task failure.
For a job to be completely replaced by AI, all of the tasks associated with that job must be identified and either completed by that AI, delegated to another job function, or eliminated. The process of compiling an exhaustive list of everything you’re responsible for doing as part of your job responsibilities is probably going to be pretty difficult and time consuming. Now multiply that effort across every organization, every job function, and every position. That’s a lot of work – not something done at the snap of a CEO’s finger.
In addition to the herculean effort of enumerating all of these tasks, I think there’s probably a bit of a cognitive bias at play when leaders in the AI field talk about the ease of automating work. An example that’s often cited as evidence that we are clearly headed for an automated future is the position of Software Engineer – a computer programmer. Today’s most advanced models already perform extremely well in this domain, and empowering them with agentic abilities will increase their utility in a commercial important way. But there’s something particularly nifty about computer code that separates it from the work product produced in many other tasks – it’s easily verifiable. Either the program compiles, runs, and passes testing, or it doesn’t. This sort of pass/fail task lends itself remarkably well to automation because most AI systems are trained using reinforcement learning. Very simply, they see lots of good examples and lots of bad examples, and learn to do more of the things associated with the good examples than the bad. Repeat this several thousand/million times and you get a system that’s performant.
Where are there a lot of software engineers? AI companies. If you’re surrounded by people who are doing easily demarcated tasks that have verifiably good or verifiably bad results, it might be natural to over-index on this first person experience and believe that much of the work that goes on in the economy shares those characteristics. I don’t think that’s necessarily the case, and bumping up against this fact when broadening the scope of the problems that these systems tackle may result in slower timelines than anticipated.
Assumption #2 – User, customer, and market preferences are multifaceted, and include a component of acceptance for inclusion of AI technology. Eventually the benefits of including AI outweigh factors that may make someone less likely to want to engage with an AI offering.
If I were to pitch you a new software solution that promises to replace your entire accounting department with an army of AI agents at a tenth of the cost, that deal may sound pretty good. There’s a catch though, it can’t present at your quarterly board meetings like your current head of accounting does. Oh, and it’s not very persuasive chasing down past due accounts. Those tradeoffs might be worth it, so you let your team go and install your agents. For a 90% discount, you’ll stomach some discomfort and shift those tasks elsewhere.
These tradeoffs might not always be worth it, and sometimes will downright backfire. In a recent example, Swedish fintech firm Klarna recently hired back some human employees after AI customer service bots frustrated customers and reduced service quality.
Just because an AI replacement is capable of performing some economically valuable activity doesn’t mean that whoever is on the receiving end is going to want AI to do that work. Sometimes, they will swallow the tradeoff. Sometimes, they’ll vote with their feet and change their consumer behavior. And if a manager cannot reliably understand how a given conclusion was arrived at, or cannot wholly replace an entire job function, it may be easier to stick with the status quo of human labor.
Assumption #3 – The problems that face current day AI systems, like hallucinations narrowly, and general inscrutability broadly, will be solved well enough to make their functional deployment tenable.
Hallucinations, when a model fabricates information, answers, hyperlinks, or court cases still plague large language model-based AI systems today. While hallucination rates have trended down over time, and savvy users of these tools know to look out for these errors, there’s not yet a silver bullet to address them. One mistake in one response to one prompt is bad – but what happens if that hallucination occurs in the initial step of a many-step workflow undertaken by an autonomous agent? That butterfly wing flap might substantially throw off the end work product, making the entire endeavor worthless. Add to this the fact that despite advances in mechanistic interpretability research — the study of how AI systems “think” — we still don’t have much of an idea of how exactly these systems work, so it’s not like you can reliably interrogate the process used to arrive at the final result. Some systems have started to show users the reasoning used to arrive at an output, but this feature currently lacks the depth and completeness necessary to provide an exact and thorough accounting in the way that a human employee could.
Compounding and Additive Effects
The extent to which these assumptions are accurate will have a huge impact on AI’s ability to have real impact in the economy. IF jobs can be easily and completely broken down into discrete task lists AND IF consumers of AI work output accept any tradeoffs or shortcomings AND IF problems like hallucinations and interpretability get sufficiently addressed – we get one version of the future. This version of the future results in drop-in remote workers, a country of geniuses in a data center, and the fundamental restructuring of white-collar work.
If however, it’s hard to fully account for and replace every action completed in a job role, consumers and organizations resist AI integration due to lack of efficacy or completeness, and the challenges of current day AI systems persist, that’s a very different future. This future looks more like a patchwork implementation of agents with varying levels of autonomy that work in more narrowly and well-defined domains. Still economically impactful, but not as capable of swiftly wiping out entire job categories.
I think the next 12-30 months (from mid-2025 to the end of 2027) will give us a good indication of the true level of job displacement we might expect to see before the end of the decade. As AI model providers turn more of their attention (and compute) to reinforcement learning paradigms, it will become clear whether or not we have/can collect sufficient data to train models in the tasks that make up these white-collar jobs or if the models become smart enough to generalize to many domains of computer-focused work without explicitly training on them. AI agents that can use computers like people do – i.e. navigating a browser, opening a webpage, filling out fields – are still in their infancy. Currently, they are not capable in the ways that would be necessary to replace white-collar workers, working too slowly or too unreliably to make for an effective substitute.
Not Everything, Not Everywhere, Not All At Once
I expect AI to have a huge impact on the way we work. A recent survey indicated that 42% of workers are using Generative AI tools in a professional capacity, and that number shows no sign of slowing down. I don’t think we will have a drop-in, generally capable remote worker by 2027. But by 2030? It’s more likely by then. And the likelihood of this scenario, and penetration into the real economy, will increase year after year. The scenario Dario Amodei is worried about, where 30% of jobs are capable of being replaced by AI and 70% aren’t, seems a lot more likely to me than 0% replaceability or 100% replaceability. To his credit, he has made the media rounds himself trying to raise the alarm about these economic possibilities.
How long does it take to get to 30% of jobs being replaced? Then from 30-31%, and 31-50%? If it’s over the course of decades, I think we’re better equipped to absorb that societally than if it takes three years.
I don’t want you to brush off the proclamations in that video clip as all hype and bluster. I also don’t want you to leave this post with an overwhelming sense of dread about your job being replaced in the next several years. My hope is that you stay tuned, here and anywhere else you see fit, to understand the rapidly changing AI landscape and seek out nuance in a sea of embellishment, bravado, and naysaying.
Leave a Reply