In 2017, I wrote a short post about Facebook's AI.

Two chatbots had started replying to each other in something that looked like broken English. Facebook stopped the experiment, and the Internet freaked out, like always. Skynet was finally here.

It wasn't. It was two programs saying "balls have zero to me to me to me".

In that same post I said I was more worried about AI's economic impact than about killer robots. I even linked a piece warning that 3.5 million American truck drivers could see their jobs at risk within ten years.

Nine years later, the truckers are still trucking.

Six months ago, I wrote that AI didn't take your job, but it did change the game. I still believe that. But since then, something else has become obvious.

The AI takeover isn't just late. In many places, it's backfiring.

Why AI Will Replace Your Job (Hypothetically)

Let's be fair to the fear first.

When I graduated in 2017, my final presentation was about AI, and I used Will Robots Take My Job? to make the point. The site was built on a 2013 Oxford study by Carl Benedikt Frey and Michael Osborne, The Future of Employment. They scored 702 occupations and concluded that about 47% of US employment was at risk of computerisation.

The logic is simple. And honestly, it makes sense:

  • A job is a bundle of tasks.
  • If those tasks are predictable, structured and rule-based, software can do them.
  • Software doesn't sleep, doesn't ask for a raise, and supposedly costs cents per task.

The site's own methodology says it plainly: a role is higher risk when the work follows rules, scripts or standard procedures, and when there's little need for human judgement, trust or care.

Then ChatGPT arrived in late 2022, and the list of "routine" tasks suddenly included writing, coding, summarizing, translating and drawing. For the first time, the robots weren't coming for the factory floor.

They were coming for the office.

So, hypothetically, the math is brutal. If a machine can do most of your tasks, cheaper and faster, why would anyone keep paying you?

That's the theory. Here's what happened instead.

The Hype, the Paranoia, and the Prophets

With ChatGPT, the conversation went from "cool demo" to "the end of work" in a few months.

Part of it was genuine excitement. A big part of it was bad reading. In 2023, Goldman Sachs estimated generative AI could affect the equivalent of 300 million full-time jobs. The same report said most jobs were only partially exposed, and more likely to be complemented than replaced.

Guess which half made the headlines.

That's how the paranoia works. It doesn't need data. Someone, somewhere, says "AI will replace humans." It gets repeated until it sounds like a fact. And if one prediction out of a hundred lands, that's the one everybody remembers.

Meanwhile, here's the prophets' scorecard:

  • Geoffrey Hinton, 2016: said we should stop training radiologists, because it was "completely obvious" AI would do their job better within five years. Ten years later, there's a shortage of radiologists, and they averaged $571,000 a year in 2025.
  • Elon Musk, 2016 and 2019: first a hands-free drive across the US by late 2017, then over a million robotaxis on the road in 2020. Neither happened on schedule.
  • Elon Musk, 2024: AI would be smarter than any one human by the end of 2025. It's October 2026. Still waiting.
  • Mark Zuckerberg, January 2025: told Joe Rogan that in 2025 Meta would have an AI working as a mid-level engineer. By April, he had moved it to "12 to 18 months".

And then there's the TikTok and LinkedIn crowd, announcing the death of programming every six months since 2023. Usually right before selling you a course.

The funny part? In my 2017 post I sided with Zuckerberg against Musk, because creating irrational fear was irresponsible. Eight years later, he was the one on a podcast giving mid-level engineers an expiration date.

A lot of the loudest voices barely touch the stuff, by the way. Consultant Nikhil Suresh spent a year and a half interviewing executives about AI. He met one who built an AI strategy for a $2-billion-a-year business and had never even used ChatGPT or any AI tool in their life. Others swore AI was transforming their company, then admitted they weren't using it for anything.

Experts love the drama. Drama sells. Nuance doesn't.

You can even see the paranoia in the numbers. On Will Robots Take My Job, visitors vote on how likely a job is to be automated within 20 years. For every office job I checked, the crowd's guess is higher than the site's own model. Often double.

The one exception? Truck drivers. The job everyone said would disappear first is the only one where the crowd is now less worried than the model.

Why I Never Bought It: Replacement Is a Social Decision

Since around 2015–2017, I've held the same view: whether a job actually gets replaced isn't a technical question. It's a social one.

Technology decides what can be automated. Society decides what will be. And that decision has to pass through a lot of filters:

  • Economic: is it actually cheaper once you count errors, supervision and the bill?
  • Legal: who's liable when the machine gets it wrong?
  • Ethical and moral: should a machine make this call at all?
  • Social acceptance: will customers, workers and voters put up with it?

This isn't a new idea, by the way. Even the 2013 Oxford study said it made "no attempt to estimate how many jobs will actually be automated", and listed regulation, political activism and public acceptance among the factors it left out. The paper even opens with Queen Elizabeth I refusing a patent for a knitting machine in 1589, because it would leave her subjects without work.

Which brings me back to the truck drivers.

In February 2017, Elon Musk told the World Government Summit in Dubai that driving was the single largest employer of people. Autonomous vehicles, he said, could leave 12–15% of the workforce unemployed within about 20 years, so we'd need universal basic income.

Nine years later:

Two hundred trucks. Two million drivers.

And it's not that the technology doesn't work. Aurora reports nearly 440,000 driverless miles through June 2026, with no collisions blamed on its system. It's everything else. Truck maker Paccar asked Aurora to put a human observer back in the cab. The Teamsters are pushing human-operator rules in at least ten states, and Colorado's legislature passed one in 2025 before the governor vetoed it.

Even Will Robots Take My Job still rates truck drivers at 68%, "High Risk." The risk is real. The replacement isn't.

"More Likely" Is Not "Will"

Here's what people miss about those percentages: a high score means a job is more exposed. It doesn't mean it's gone.

The site says it itself. Its score is "not a prediction that every job will disappear by a specific date". It's a signal, not a forecast.

Take accountants. The 2013 Oxford study gave accountants and auditors a 94% chance of computerisation. Will Robots Take My Job used to show that number. Today, with a newer model, it puts them at 39%. The US still employs about 1.45 million of them, and that number is projected to grow through 2034.

Now ask yourself a simple question. Would you hand your taxes to a chatbot and face an audit with no accountant to call?

Or your doctor. Family medicine physicians sit at 12%, "Minimal Risk." Even radiologists, the poster child for AI replacement, score 34%. And some of the reasons have nothing to do with technology. In the US, Medicare and Medicaid only pay for a radiology study if a licensed physician does the final read, and nobody knows how to hold an AI responsible for a missed diagnosis (Fortune).

That's the social layer in action. Trust, accountability and liability aren't engineering problems. You can't fine-tune your way out of them.

So yes, some jobs are more automatable than others. But "more likely" is just that: more likely. It's not a death certificate.

The Backfire, Part 1: The Layoffs

Then came the layoffs. Or at least, the layoffs with "AI" in the press release.

The poster child was Klarna. In February 2024, it said its AI assistant was doing the work of 700 customer service agents and handling 75% of customer chats. Headcount dropped 22% under a hiring freeze. By May 2025, the CEO was telling Bloomberg that the AI-first approach had led to lower quality, and Klarna started hiring humans again.

So what went wrong? Three things.

  1. AI wasn't the real reason. Gartner says most recent workforce reductions were influenced by broader economic conditions, not automation alone. Only 20% of customer service leaders actually cut staff because of AI. For many companies, AI was the press release. The economy was the reason.
  2. Quality dropped. Klarna admitted it. Gartner now predicts that by 2027, half of the companies that cut customer service staff because of AI will rehire people to do similar work, just under different job titles.
  3. It's a trap. Researchers Gerry Tsoukalas (Boston University) and Brett Hemenway Falk (University of Pennsylvania) call it the AI Layoff Trap. Each company keeps the savings when it fires people, but the lost paychecks shrink demand for everyone. Every firm sees it coming, and cuts anyway. Tsoukalas walks through it on BBC's New Normal. In their model, five of six popular policy fixes fail, and they're careful to call it a structural risk, not a crisis yet.

Put simply: you can't fire your customers and expect them to keep buying.

To be fair, not every cut gets reversed. Block and Snap both pointed to AI when laying off thousands of people. And Will Robots Take My Job projects computer programmer jobs to shrink 6% by 2034. Some roles really are getting squeezed.

But "squeezed" and "replaced" are different words.

The Backfire, Part 2: The Money

Remember "supposedly costs cents per task"? Here's the bill.

Uber pushed its engineers to adopt AI coding tools like Claude Code, complete with an internal leaderboard ranking teams by usage. It burned through its entire 2026 AI coding budget in four months. Four. Months. Asked whether all that spending turned into more useful features for customers, Uber's COO said: "That link is not there yet."

Microsoft opened Claude Code to thousands of developers, project managers and designers. About six months later, it started canceling most of those licenses and moved people to its own GitHub Copilot CLI. The more people used it, the bigger the bill. An Nvidia executive put the new math bluntly: for his team, compute costs are "far beyond the costs of the employees."

Read that again. The thing that was supposed to replace employees can cost more than the employees.

And it's not just a couple of companies. MIT's GenAI Divide report found that, despite $30–40 billion in enterprise spending, 95% of organizations were getting no measurable business return.

Now zoom out to the data centers. Gartner says AI providers will spend nearly $1 trillion on AI infrastructure in 2026, and another $6 trillion between 2027 and 2030. To earn a decent return, data center owners would need about $8.5 trillion in revenue by 2030. Gartner expects no more than $3.6 trillion. That leaves more than $3.5 trillion exposed to write-downs.

How do you justify that kind of spending? With a story. And the story was the takeover: AI would replace enough people to pay for itself. That narrative sold the data centers.

Meanwhile, as TechLead argues, China is winning the "race to zero" on price. It's hard to earn back a trillion dollars selling something that keeps getting cheaper. No wonder some are already calling the AI gold rush dead.

AI Still Needs an Adult in the Room

Here's what many sources keep saying, and what I keep seeing in practice: these models are limited by human input, and their work is error-prone.

Limited by human input. Models learn from what people wrote, built and corrected. A 2024 study in Nature found that training models on AI-generated content causes irreversible defects, which the authors call model collapse. Their conclusion: data from real human interactions will become more valuable over time, not less.

Error-prone. The same model that passes an elite exam can fail at reading a clock or counting letters. That's the "jagged frontier" The Infographics Show breaks down in this video, and it's eating away at public trust. In court, it's worse: by June 2026, researcher Damien Charlotin had logged 1,116 US cases of filings with AI hallucinations, and the pace had grown eightfold in a year. In code, as Kantan Coding explains in The Paradox of Why AI Code Is Failing Us, the bigger the codebase, the more models struggle: context gets expensive fast, attention misses connections, and retrieval has blind spots.

So who cleans up the mess? Experts. That's the twist Pooja Dutt points out in The AI Takeover Has Completely Backfired: AI's failures accidentally created new human industries.

The irony writes itself. The technology that was supposed to replace experts needs experts to train it, and experts to fix it.

Vibe Coding: Why It Works (and Why It Doesn't)

I wrote about vibe coding earlier this year, and I'll say it again: I do it. A lot.

Why it works:

  • Speed. Prototypes, demos, scrapers and internal tools that used to take weeks now take hours.
  • It lowers the bar. Microsoft didn't just give Claude Code to developers. It encouraged project managers and designers to experiment with coding too.
  • At scale, it's real. Uber's CEO says about 10% of the company's committed code is now built by autonomous agents.

Why it doesn't:

  • You can't fix what you don't understand. That's my calculator analogy: if you don't know what addition is, you can't tell when the answer is wrong.
  • It's slower than it feels. In a 2025 METR study, experienced developers using AI tools were 19% slower, while believing they were 20% faster.
  • It doesn't scale with your codebase. The bigger the project, the harder those context limits bite.
  • Someone still pays the bill. Ask Uber.

Vibe coding is a fantastic way to start something. It's a terrible way to maintain something you don't understand.

The magic isn't the prompt. It's knowing when the output is wrong.

What AI Actually Got Right

Now, let's be fair. AI didn't take over, but it did change the game, like I said six months ago. Two things in particular.

It lowered the barrier of entry. In a study of 5,179 customer support agents, an AI assistant raised productivity 14% on average, and 34% for novice and low-skilled workers, with minimal impact on the most experienced ones. AI didn't replace the experts. It helped newcomers work more like them.

It democratized a lot of tasks. Someone with zero design background can mock up a logo. A product manager can prototype an app. A small business can translate its website in an afternoon. I'm a solutions engineer who codes maybe 5% of the time, and I still build iOS and Android prototypes.

That's the real story: not replacement, access.

And the people getting the most out of it are what Cory Doctorow calls centaurs: experienced workers who use automation on their own terms, and can tell good output from bad.

Final Thought

My position hasn't changed since 2017.

AI complements jobs. That's it. That's the take.

It makes good people faster. It helps beginners get started. It takes the boring parts off our plates. But it still needs someone who knows what the output should look like, someone who signs, someone who's accountable, and someone who cleans up when it breaks.

Could a real replacement happen someday? I'm not ruling it out. But not in the short term, and not in the medium term either. Not while the technology depends on human input, makes confident mistakes, and has to pass every social, legal, economic and ethical filter we put in front of it.

In 2017 I wrote that Facebook didn't panic, and you shouldn't either. Nine years and one very expensive hype cycle later, same advice.

The takeover didn't backfire because AI is useless. It backfired because a job was never just a bundle of tasks.

Copied!