AI can’t replace you, but it might anyway
My Kindergartner is typing on a dusty old keyboard he found in our garage. It’s disconnected from any computer but he’s having a great time pretending. “Look daddy, I’m working in my office!” he says with that beaming smile that always makes my heart grow wings. But he might be role-playing a career choice that will no longer exist by the time he comes of age.
I think we can safely say that all pretense has been swept away and the true purpose of generative AI and AI agents, now freely spoken by CEOs, is the explicit usurpation of humans in all knowledge work. That’s no longer in question. The goal is not human augmentation, and we were extremely naive to ever think it was. No, the express goal of the AI industry is human replacement on a massive and unprecedented scale, plain and simple.
We see this manifest in the “AI-first” mandates that CEOs have been issuing to employees through (purposefully) leaked internal memos. “No more hiring unless AI can’t do the job” has replaced “AI won’t replace you, but someone using AI will.” Even though I’m glad to see the demise of that vacuous and backstabby retort, the new maxim is even more short-sighted and dystopian.
Workers are mostly powerless to stand up to or defend themselves against what’s coming. It would be like fighting a shadow rather than a physical enemy. Because AI is an ideology, an addiction, perhaps even a religion, in the minds of those who are completely enamored or awed by it. This is not business as usual.
However, AI isn’t really an inevitable or unstoppable technology. And the narrow AI tools and systems that have caused us to lose our collective minds these past three years aren’t even as capable as they’re marketed to be. But looking at layoff news, product hype, and stock market shockwaves, you could be forgiven for not knowing that.
Here’s what I’ve been grappling with for a while now, which may seem contradictory at first glance:
Generative AI and AI agents are not as capable or as amazing as AI company CEOs, overpaid management consultants, and enthusiastic techbros want us to believe.
The point above doesn’t matter and won’t be taken into consideration. Many CEOs believe all the hype and disinformation, hook, line, and sinker, and feel like they have no choice but to embrace AI or get left behind. So they’re going to systematically replace all human workers because that’s what they think market forces demand.
It can be surprisingly difficult to convey this to my readers. I get comments all the time pointing out the seeming hypocrisy of saying “AI ain’t all that” but that “it’s still going to take everyone’s job.” I hope the above statements clarify where I stand. It really boils down to perception versus reality: the reality that the AI industry is a jenga block tower of fraud and misdirection, versus the perception that AI represents a singularity event in our evolution as well as the business opportunity of a lifetime. A perception that billionaires paint for us through their marketing budget, unhinged manifestos, and sheer force of will. Of course, what’s real hasn’t mattered to capitalism for quite some time. What matters is profit, or the promise of future profits.
Evidence of generative AI’s incompetence and untrustworthiness are too numerous to mention. New infamies and slapstick disasters emerge almost daily. And it’s not always a laughing matter. In a rising number of cases, real people are being hurt, or even losing their lives. So it seems to me that the failure of AI to live up to the hype so far should not be in question. And yet many executive leaders appear completely ignorant of the risks and shortcomings of their latest obsession.
CEOs who are true acolytes of the church of AI don’t realize their own “fear of missing out” is guiding most of their decisions to mandate AI at work. They’re so bought into the hype and propaganda that they succumb to a bizarre false dichotomy that makes them look more like ruthless Victorian-era factory owners than modern day leaders. Even famous backfire events, like the story of when Klarna met karma, don’t seem to dissuade other CEOs from repeating the same mistake.
If you imagine a food blender big enough so that thousands of human workers can be pushed into it, chopped up, and reconstituted into a gigantic pink goo that is somehow profitable and smells like roses, you’re seeing what most CEOs see when they look at the Anthropic or OpenAI logos, even though most normal people see gaping assholes.
The devil in the details
To complicate matters, there are at least two more nuances to consider, and these frequently make their way into my musings too:
Most CEOs don’t seem to care about quality or accuracy if “good enough” will still generate profits. So even the more savvy CEOs who don’t fully buy into the myth of AI will opt for layoffs because their overarching goal is simply to run their company cheaper and faster.
Some CEOs who don’t buy into the myth of AI will still use the myth of AI as a smokescreen to justify massive layoffs that are already planned.
And we’ve seen early signs of this playing out too. When Duolingo CEO Louis von Ahn said his company would replace all contractors with AI he admitted that quality wasn’t considered a key to success:
“we can’t wait until the technology is 100% perfect. We’d rather move with urgency and take occasional small hits on quality than move slowly and miss the moment.”
In other words, cash in on the AI mania now, worry about the after effects later. In the meantime, people don’t matter — they’re expendable. Let’s not pretend that “AI-first” doesn’t also mean “humans-last.”
We’ve also seen a parade of CEOs who are obviously lying about “AI efficiency gains” to hide the fact that they’re really just scaling back from their massive and unsustainable headcount growth during the time of Covid lockdowns (which gave a temporary artificial boost to online businesses). One word for this is AI-washing: money laundering for anti-human leaders.
The fallacies of “AI-first”
There are so many fallacies to challenge when I think about these “AI-first” mandates. Not least of which is the fact that AI is not actually capable of replacing any jobs, only tasks. You might accuse me of nitpicking here, but you’d be wrong. A job, or a “role” as we often say in office parlance, is a set of responsibilities and expectations around an agreed upon performance. The performance of a given role obviously requires completing certain domain-specific tasks. But that’s only the base level requirement, of course. There’s also the expectation that employees will learn on the job, adapt to changing circumstances and priorities, and be able to figure things out even in situations that are vague or uncertain. In other words, in most knowledge work you are expected to work at least partly without supervision and to lean on tacit rather than implicit knowledge and instructions, and to always be learning and adapting, or “growing” as we say in corporate vernacular.
Unlike human workers, Large Language Models and “AI Agents” don’t grow, learn, or adapt in any real sense. Every model has a training cut-off date because the cost and complexity of retraining is significant and can’t be performed in real time or on-demand. Even in instances where you apply techniques like RAG or fine-tuning and context-stuffing, any short-lived memory of how your organization operates and what the preferred results should look like are eventually lost in translation. Any task to be performed by an “agentic” system needs to be so simple and repeatable that it can be broken down into bitesize inputs, decisions, and outcomes, and even then there are a thousand things that can go wrong. Because the truth is, even basic decisions and actions that human beings perform in their daily work are notoriously difficult to pin down and translate into machine-readable instructions. In fact, there’s a whole history behind the failures of AI systems to reliably interpret human work functions and expertise stretching back to the last “AI winter” of the 1980s.
The “expert systems” of the 80s couldn’t handle ambiguity at scale. They were brittle and couldn’t be made flexible enough to keep up with the ever changing scope and expectations of normal everyday office work, not without millions of hours spent on maintenance and modification of the thousands of rules that held the system together. Now imagine trying to accomplish the same goal using modern AI that is built using probabilistic algorithms — meaning the instructions can be misinterpreted or ignored, the process itself partially occurs in a black box with no paper trail, and outputs are shaped by non-deterministic decisions where consistency and accuracy can’t be guaranteed. A large language model is a parrot that sometimes lies. And a “multi-agent” system is a Rube Goldberg machine with multiple points of failure, often requiring human intervention that completely negates the purpose of automation.
This is why you’ll notice most “agentic” projects center around replacing very basic procedural tasks, perhaps a glorified customer service chatbot that escalates to a human at the first sign of trouble, or flakey personal assistants that draft emails you can’t be bothered to write and sometimes go rogue in unexpected ways. You’ll also see countless iterations of automating the summarization of documents and spreadsheets, or basic language translation — all things that are simple enough that “good enough” outputs can be expected at least 75% of the time. None of this is rocket science or even very transformative. There’s also a lot of orchestration that isn’t really agentic at all, but automation workflows using API integration services like Zapier or N8N. Some CEOs probably praise their staff for bringing the company into the golden age of AI without even realizing that no AI was harmed during installation.
Anytime AI is employed in critical work, we see failure after failure. The cost of using unreliable LLMs is stacking up. Meanwhile, pragmatic research experiments have demonstrated how “agents” break down as the complexity or the number of steps in a task increases. LLM hallucinations are well documented, but less is known about why “agent” systems drift and ignore the rules (side note: good article but the conclusion is preposterous, as if adding more AI to the AI will solve the problem).
Techbros underestimate the complexity of human work
If you work in a white-collar job, think about your own day-to-day. Could a bot replace you?
I have no qualms declaring that most roles can’t be reduced to a simple list of tasks. Knowledge workers are constantly multitasking and collaborating, applying domain knowledge, skills, experience, creativity, empathy, and intuition to unique situations with unique variables. Office work depends on balancing priorities, give-and-take relationships, emotional intelligence, and psychological dexterity. A lurching, hallucinating language grinder or a bunch of API requests strung together in a loosely defined decision framework with flimsy guardrails can’t possibly replicate human work at scale. It just doesn’t make any sense.
The only domain where “AI agents” seem to be making any real progress is in software engineering. This is because the models have been trained on pretty much all code ever written, and code is explicitly written to be interpreted by computers, formatted to be logical, structured, and organized in a way that often conforms to common patterns and styles — much easier for an AI to comprehend than all the nuances and flourishes of natural language. But even in this niche, the hype seems to run way ahead of the reality. AI-generated code is full of security flaws and subtle errors that human spotters often miss because of the tendency for the AI to confidently make mistakes, drift and ignore instructions. Even the adoption of AI coding tools by professional developers is causing “AI fatigue” and a significant decline in motivation and morale. As I wrote about in my other blog where I explore human-centered management and software engineering, AI is turning software development into a Kafkaesque fever dream.
The only thing that’s inevitable is human nature
Alas, the conclusion of this post is that none of my careful reasoning matters. CEO’s are not interested in any kind of reality check. They will not see reason, not even when some of their peers are already stumbling into massive failure. Not even when reports come out that nearly all AI projects result in no ROI. Despite admitting behind closed doors that AI productivity is a myth. It’s frustrating and stupid and enough to make you want to pull all your hair out. But unless politicians get involved I’m not sure if anything can stop this cascade of inept leadership decisions.
Right now, I’m reasonably confident that most “AI layoffs” are really a smokescreen for rightsizing where companies over-hired during the pandemic or have revenue deficits they need to redress. But that doesn’t negate the fact that layoffs are real, and right now the IT industry is being hit particularly bad. If you’re facing the possibility of a layoff, or have recently lost your job, know that I empathize with you and you’re not alone. I think the best advice I can give is to not blame yourself, to not give up on yourself, and to keep learning and adapting until the economy and the job market rights itself. There will be new opportunities, although they may not be in tech. And maybe that won’t be such a bad thing at the end of the day. Having been bullied and gaslit quite a lot during my tenure in software development, I honestly don’t miss it as much as I thought I would.
As I watch my son playing with the unplugged keyboard, or playing games on his kid’s kindle fire, I have no idea what’s in store for his future. I used to be excited about teaching him how to program computers one day. Now I wonder if that’s prudent or even relevant anymore. I kind of hope he chooses to work with his hands instead, solving real problems for real people, something he can be proud of at the end of the day.
Twenty years ago I signed up to build a world-wide-web that was optimistic and community-led, now it’s a corporate battlefield populated mostly by bots, shills, and deepfakes. But more than anything it’s not the proliferation of AI or any other tech that worries me, it’s the people using it or hiding behind it to make unconscionable decisions, or using it as a tool to consolidate wealth and power, a weapon of fear, or an excuse to be less human. We have to rail against that somehow. AI may have taken away my career, but I won’t let it replace me.
Last year I instructed some free “lightning lessons” on AI agents and the “AI-first” fallacy. If you’re interested, you can find them here.
As a plumber, folk say AI won't replace me. But of course it sure looks like it is set to replace all my customers, who will, as a result, no longer be able to afford my services. So it looks to me as though even people like me will be out of work in short order.
In the 1970s I was computer modelling the first decades of this century for HMG. The Geologists were saying that Industrial Civilization would be in decline by 2020 so I'm well aware that this AI nonsense is all part of a plan, hatched decades ago, to bring about a managed decline, in the hope of avoiding a catastrophic collapse.
And your son? Well constantly remind him that the best things in life are free, and always have been, it's just that our culture has somehow forgotten the obvious.
Really thought-provoking title! What's the most surprising way you've seen AI actually replace someone in a role you didn't expect?