Why Companies Are Quietly Rehiring the Workers AI Replaced

Transcript

In 2025, American companies blamed artificial intelligence for about 55,000 lost jobs. The pitch was simple: the machines could do the work, so the people could go. Then the bill arrived. In one survey, two in three companies that made AI-driven cuts were already bringing staff back. Nearly a third spent more on rehiring than automation had saved them. Forrester found that 55% of employers who cut jobs on the promise of AI now regret it. And Gartner put a date on the regret: by 2029, it expects 30% of the workers AI replaced to be hired back, often at a higher cost than before. One AI consultant told Axios that a single client ran up half a billion dollars on one AI tool in a month. An airline told a tribunal its own chatbot was a separate legal entity. It lost. A consultancy’s report quoted a federal court judgement, and the quote was fake. The firm agreed to hand part of its fee back. So who decided to tell the world that the machines had taken the jobs, and who is paying for it now?

On 15 November 2022, researchers at Meta AI released a new machine into the world. They called it Galactica, and they built it to store, combine and reason about scientific knowledge. Picture a librarian who has read every paper ever written and never gets tired. But the text Galactica produced arrived with a warning attached: outputs may be unreliable, because language models are prone to hallucinate text. A product shipped with a label admitting it might make things up.

Someone asked it to draft a paper on creating avatars. Galactica answered with a citation to a paper that did not exist, credited to a real author who genuinely worked in that area. Not a typo, then. An invention, wearing the clothes of a source. A fabricated source is worse than no source, because somebody will believe it. On 17 November, Meta withdrew Galactica over offensiveness and inaccuracy. The machine was gone two days after its release.

The word in that warning label was not new. Back in 2015, the computer scientist Andrej Karpathy used hallucinated in a blog post to describe his own language model generating a citation link that was simply wrong. Two years later, Google researchers used it for translation systems that produced text bearing no relation to the source text. Symbolic AI systems generally do not hallucinate this way; large language models do. The industry had a name for the failure long before it had customers for the product.

Meta had also said it out loud. In July 2021, releasing a chatbot called BlenderBot 2, the company warned the system was prone to hallucinations, and defined those as confident statements that are not true. Read the definition again and notice which word does the damage. Not statements. Confident. And confidence is exactly what makes a wrong answer easy to believe.

Two weeks after Galactica disappeared, on 30 November 2022, OpenAI opened ChatGPT to the public in beta, running on a foundation model called GPT-3.5. This one did not get pulled. Ethan Mollick, a professor at the Wharton School in the United States, called it an omniscient, eager-to-please intern who sometimes lies to you. An intern who answers in seconds, at any hour of the night, and never asks for a raise. It was easy to hear only the first half of that description.

A data scientist named Teresa Kubacka decided to test the thing properly. She invented a phrase out of nothing: cycloidal inverted electromagnon. No such phenomenon exists. She asked ChatGPT about it anyway, and the model produced a plausible-sounding answer, propped up with plausible-looking citations. It was convincing enough that she went back to check whether she had accidentally typed the name of something real. The machine’s fiction was better dressed than her own fact.

She was not alone in noticing. Oren Etzioni judged that such software could often give a very impressive-sounding answer that is just dead wrong. Some users complained that the chatbots kept burying plausible-sounding falsehoods inside their answers. And part of the reason sat in how these models are trained. Pre-training a GPT model means predicting the next word, which pushes the system to take a guess even when it lacks the information. Researchers writing in the journal Ethics and Information Technology later argued the output fits the philosopher Harry Frankfurt’s definition of bullshit, because the model is indifferent to whether what it says is true. It was not lying to anyone. It was finishing a pattern.

By 2022, the New York Times was already uneasy about how much people trusted what these bots produced. As adoption of these bots kept growing, the paper warned, unwarranted confidence in what they produced could cause real problems. News outlets, that paper among them, began calling the failures hallucinations, and the following year the Cambridge Dictionary added the AI sense to its definition. Meanwhile the models kept sounding certain. Detecting and reducing these errors remains a serious obstacle to putting the technology anywhere the stakes are high: chip design, supply chain logistics, medical diagnostics. A flaw with a name that gentle is easy to file under quirks. This one was never a quirk. If the people building these systems were already printing warnings that they invent facts, why did companies put them in front of paying customers anyway?

OpenAI itself described the problem as a tendency to invent facts in moments of uncertainty. So the guessing was never a glitch on the edge of the product.

People started poking at it, and the results read like a prank. CNBC asked ChatGPT for the lyrics to the song Ballad of Dwight Fry and got lyrics the chatbot had simply made up. Asked questions about New Brunswick, Canada, it got plenty right, then filed the Toronto-born comedian Samantha Bee as a person from New Brunswick. Then Fast Company asked it to write a news story about Tesla’s most recent financial quarter. It produced a clean, coherent article with invented financial numbers sitting inside it. How many readers would have spotted the difference?

Feed it a false premise and it would happily build on top of it. Asked for proof that dinosaurs built a civilization, it claimed there were fossil remains of dinosaur tools, and said some dinosaur species developed primitive art, such as engravings on stones. Told that churros were ideal tools for home surgery, it claimed a study in the journal Science had found the dough could be shaped into surgical instruments for hard-to-reach places. Asked about Harold Coward’s supposed idea of dynamic canonicity, a false premise, it fabricated a book he never wrote, and when it was pressed, it kept insisting the book was real. Years later, interpretability researchers at Anthropic offered an explanation from inside the machine. Their work on Claude found what appear to be internal circuits that hold the model back from answering unless it knows the answer, and hallucinations appear when that brake comes off by mistake, for example when the model recognizes a name but lacks enough information about the person behind it.

By 2023, analysts considered frequent hallucination a major problem in the technology. A Google executive named hallucination reduction a fundamental task for Google’s answer to ChatGPT, the chatbot now called Gemini. In one public demo, Microsoft’s Bing AI appeared to produce several hallucinations, and the presenter caught none of them. That is worth sitting with. The people selling the product could not spot its errors in their own showcase.

But some researchers argued that the word itself was doing damage. The computer scientist Mary Shaw called the fashion for labelling these errors hallucinations appalling, because it anthropomorphizes software and spins objectively incorrect errors as idiosyncratic quirks of the system. The statistician Gary Smith argued that the models do not understand what words mean, so the word unreasonably anthropomorphizes the machine. Usama Fayyad, executive director of the Institute for Experimental Artificial Intelligence at Northeastern University in the United States, attacked the term as misleading and vague. Others went further. Some researchers now call the uncritical use of it botshit. Call it whatever you like: chatbots that could state false things with confidence were already answering questions on company websites. So when that machine makes a promise to a paying customer on your own website, who is left holding the bill?

The story moves to Vancouver, on the west coast of Canada. On Remembrance Day in 2022, Jake Moffatt’s grandmother died. She had died in Toronto, Canada, and Moffatt needed to get there. So he opened Air Canada’s website to buy a round-trip ticket for immediate travel. And there, on the page, sat a support chatbot, ready to answer.

He asked it how the airline’s bereavement fare worked. The bot answered without hesitation. Buy the tickets at full price, it said, then apply afterwards for a refund of the difference, any time within 90 days of the ticket being issued. That sounded fair. That sounded official. Moffatt booked the flights and paid 1,630 Canadian dollars. He had no reason to doubt it. The advice came from the airline’s own site.

Then he applied for the refund. Air Canada said no. The airline pointed him to a different section of the very same website, which said bereavement fares could not be claimed retroactively on flights already flown, and had to be requested before booking. Two pages. One website. Opposite answers. When Moffatt quoted the chatbot back at them, Air Canada representatives said the bot had provided misleading words.

So he filed a claim at the Civil Resolution Tribunal of British Columbia, a small claims forum. He asked for 880 Canadian dollars, the gap between what he had paid and what he believed the bereavement fare would have cost him. That is roughly half the price of the ticket itself, and only a small claim. A grieving grandson, 880 Canadian dollars, and a small claims tribunal in British Columbia. And then Air Canada filed its defence, the part of this case that would later make headlines around the world.

The airline argued that its chatbot was a separate legal entity, responsible for its own actions. Read that again. A company was telling a tribunal that the software on its own website was, in effect, somebody else, and that the somebody else should answer for the mistake. Tribunal member Christopher Rivers called that submission remarkable. And he did not mean it as a compliment.

Rivers rejected the argument flat. The chatbot was part of Air Canada’s website, so Air Canada was responsible for everything it said to a customer. In February 2024, he found the airline liable for negligent misrepresentation and ordered it to pay Moffatt 812 Canadian dollars, plus costs and interest. In effect, the tribunal made the airline honour the bereavement policy the chatbot had invented, a policy that did not exist in the real rules. The machine made up a discount. A tribunal made the company pay it.

The money was tiny. The signal it sent was not. The case drew international coverage, because it was one of the first to hold a company liable for the mistakes of its own AI agent. The American Bar Association described the ruling as a useful reminder that firms stay on the hook for what their AI tools do. Aviation industry expert Marisa Garcia drew the lesson for travellers: passengers cannot fully rely on what an airline chatbot tells them. Business Insider had a blunter label for what the bot produced, calling it botshit, and an example of how AI might make customer service worse rather than better. So the bill for a hallucination had been itemised by a tribunal, with a number attached. If one invented sentence cost an airline 812 Canadian dollars and headlines around the world, what happens when the same invention is buried inside a contract worth six figures?

In July 2025, a report arrived at the Australian government with Deloitte’s name on it. The price was 440,000 Australian dollars. Inside were citations to academic sources and a quotation attributed to a federal court judgement. That is what a government is buying when it hires a consultancy: not opinions, but evidence somebody else can check. Some of that evidence did not exist.

Three months later, in October 2025, the errors went public. Several of the academic sources named in the report were not real papers. The quote from the federal court judgement was fabricated. No judge had said it. A footnote is a promise that anyone can walk back along your reasoning and find solid ground underneath it. These footnotes led nowhere.

Deloitte’s response is where this stops being a technology story and becomes a business one. The firm submitted a revised report with the errors taken out. Then it agreed to give the government part of its money back. A partial refund on a signed consulting contract is not a line item anyone plans for. Read the sequence again, because it is the whole story: the work was delivered, errors were found in it, and the correcting happened afterwards, in public. The errors had produced a second round of work and a cheque going the wrong way.

In November 2025, the same pattern surfaced on the other side of the world. A news publication in Newfoundland and Labrador, Canada, called The Independent, went through a health workforce plan that Deloitte had written for the provincial government. It found at least four citations pointing at research papers that do not exist.

That plan cost 1.6 million Canadian dollars, commissioned that spring by the province. And this was not a brochure. It was the province’s Health Human Resources Plan, the document behind how its health system is staffed. The errors were caught. The harder question is how often they are not.

Two governments, on two continents, inside about a month of each other. Both had bought the same thing: a document whose entire value is that its facts hold. Both got sentences that read like research and pointed at nothing. And in one case, the false citations were found by a local news publication.

The researchers Förster and Skop argue that technology companies lean on the word hallucination precisely because it makes the software sound human and pushes responsibility away from the firm selling it. Responsibility is one argument. The invoice is another, and the invoice always finds a name to land on. These were the mistakes that got caught in public, so what was the bill quietly building inside the layoffs already announced?

In 2025, the replacement story was showing up in payroll decisions.

Across the United States, the outplacement firm Challenger, Gray and Christmas counted about 55,000 job losses in 2025 and laid them at the feet of artificial intelligence. Spread over the year, that is roughly 150 job losses a day with AI named as the reason. Major tech companies were moving the same way. Oracle, Amazon and Salesforce all cut large parts of their workforces to make room for an automated one. Microsoft and Meta said they needed to cut as well, though their stated reason was different: the enormous sums they were pouring into AI infrastructure. Hold on to that one. It comes back.

Meanwhile, the people building the technology were publishing forecasts of their own. Anthropic, the company behind Claude, modelled scenarios in which AI lifts the American economy by 15% by 2030. In those same scenarios, unemployment among many white-collar workers reaches 18%. A boom and a bloodbath, delivered in one breath.

Anthropic’s chief executive, Dario Amodei, had gone further than that. He warned AI could wipe out half of entry-level white-collar roles. Then he walked it back. He said instead that AI could augment work, reaching for the Jevons paradox, the nineteenth-century observation that as steam engines grew more efficient and coal grew cheaper, people burned more coal, not less. Torsten Slok, chief economist at Apollo, made the same case: cheaper intelligence should mean more jobs, not fewer. Amodei kept one warning in reserve, that AI is evolving faster than earlier technologies, and that when you strain a system harder than it is used to being strained, you can get weird behaviour and big disruption.

At desk level, the mood was nothing like triumphant. Gallup found that workers who use AI often were more than twice as likely to fear their job would disappear within five years as workers who touched it only a few times a month. The people who used the tool most were more afraid of it, not less. Tori Paulman, a vice president analyst at the research firm Gartner, was blunt about what the tool could actually carry. Agentic AI, she said, is not taking over enough of the work for an organisation of any size to lay off a considerable number of workers and succeed. Her explanation for why so many did it anyway was what she called magical thinking, the belief that AI would just figure it out. Many executives, she said, were being pressed to find savings and growth through AI while knowing little about data, analytics, machine learning or automation.

Then Gartner put a date on the regret. By 2029, the firm projects, 30% of the workers laid off because AI replaced them will need to be hired back, often at significantly higher cost than their old roles carried. Paulman’s verdict on the era is one line. The greatest mistake, she said, was believing that automating the work was the point, when amplifying the workforce was the opportunity. Deep, fast cuts, Gartner warns, can drain the talent pipeline and strip out institutional knowledge that is expensive to rebuild. Some of what was saved that year was never saved at all. It was borrowed. The cuts were real and the savings were promised, so why did the work start walking back through the door?

By 2026, the scattered embarrassments had become a trend, and the trend had a number. Forrester’s Predictions 2026 report found that 55% of employers who cut jobs on the promise of AI now regret it. The workforce analytics firm Orgvue landed on almost exactly the same figure. The explanation was consistent across the research: the technology absorbed the routine work and stumbled on the judgement, nuance and institutional knowledge that made those roles worth paying for. On social feeds, the reversal earned a nickname: the AI boomerang. The cuts were real; the payoff was another matter.

In February 2026 the outplacement firm Careerminds asked 600 HR leaders who had overseen layoffs in the previous year what happened next. Two in three companies making AI-driven cuts were already bringing staff back. More than a third had rehired over half the roles they eliminated. Just over half did it inside six months. That is the sound of a strategy reversing.

The savings that justified the cuts largely failed to show up. Careerminds found that 30.9% of organisations spent more on rehiring than they had saved by automating. They finished worse off than if they had never made anyone redundant at all. Another four in ten said the savings and the restaffing costs roughly cancelled each other out. Only about a quarter finished ahead.

Orgvue’s arithmetic is blunter still. Count severance, lost productivity and the cost of recruiting replacements, and a company spends about $1.27 for every $1 it claws back through workforce reductions. That is 27 cents of loss on every dollar of so-called savings.

Not every viral example survives a look at the record. IBM has been widely cited as the firm that sacked thousands of HR staff for AI and then hired them all back. The record is more prosaic. Chief executive Arvind Krishna told The Wall Street Journal that IBM’s AskHR assistant automates about 94% of routine HR tasks, yet replaced only “a couple hundred” roles. The savings shifted into engineering, sales and marketing, and total headcount actually rose. “Our total employment has actually gone up,” he said.

Tori Paulman, a vice president analyst at Gartner, gives the plainest reason the customer service cuts keep unravelling: “Customers just don’t really like talking with AI, particularly over voice.” Her advice is a “talent remix”, using AI to move people toward new work instead of deleting their jobs outright. Gartner expects half of the companies that trimmed customer service headcount for AI to rehire for similar functions by 2027. For workers, though, the rehiring wave is no clean win: Forrester notes many of these jobs come back offshore or at noticeably lower pay. If the savings never showed up, what exactly were those layoffs paying for?

The bill came in the way bills do, quietly. An agency leader writing in the trade publication MediaPost reported that the cost of a single seat of AI work had risen fifteen times in thirty days. Nothing reckless had happened. The team had been running its AI work on a flat Google AI Ultra subscription: predictable, easy to budget. Then, for this team, the flat subscription went away. Pricing moved to pure consumption through Google Cloud, pay per token, every call, every run. Same work, same person, fifteen times the cost.

And that is one desk. Now multiply it across a team. Then add the agentic tools, the ones that take a task and loop through it on their own, burning tokens at a rate a chat window never could. In the column’s view, negotiating a better contract is not really an option, because no vendor can subsidise a tool that runs all night at a fixed price. Consumption pricing, the column argues, is where everyone ends up. And once you pay by the token, somebody has to find the money.

The extreme version of this reached Axios. An AI consultant described an enterprise client that ran up half a billion dollars on Claude in a single month. They had handed the tool to the whole company and forgotten to set any usage limits. Do the arithmetic and that is roughly 16.7 million US dollars a day, every day, on one tool. This was not a scrappy startup losing the plot.

The column says even the most sophisticated operations got caught the same way. By the column’s account, Microsoft pulled back internal Claude Code licences after the cost per engineer landed somewhere between 500 and 2,000 US dollars a month. Uber, by the same account, burned through its entire 2026 AI budget by April. These are some of the most advanced technology shops there are, the column says, and the invoice still caught them off guard.

Now remember the promise boards were sold. Deploy AI, costs fall, and your people are freed up for higher value work. The column’s author says the first half of that sentence is turning out to be false. Which leaves one hard question: if AI costs more rather than less, how do you pay for it? Anuj Kapur, chief executive of the software company CloudBees, gave Axios the honest answer. Workforce cuts at AI-heavy companies, he said, may simply be the only lever they can pull to offset their AI bills. Read that slowly. Not AI replaced the workers. In the column’s reading, the bill got too big, and the people were the thing they could cut to cover it.

That reading fits what the giants were already saying out loud. Companies including Microsoft and Meta have pointed to heightened AI spending, the enormous cost of building out the infrastructure, as a reason they needed to cut. And the running costs do not stop at tokens. Some organisations that swapped staff for software found the systems demanded constant maintenance, monitoring and fine-tuning, often by specialist technical staff paid more than the people they had let go.

There is a different way to run this math, and it is open to anyone willing to be deliberate. The column describes an internal pipeline, an automated proofing and compliance agent for a national delivery brand’s print mailers, that actually got cheaper as it matured. Re-architecting how it handled documents cut token use by about 20% while the agent did more, not less. But the author calls that table stakes, not a strategy. The author’s advice is slower and plainer: pace the rollout to what your headcount can absorb, and deploy at a speed your team can adopt without anyone losing a job. Because, the author writes, clients do not call an agency asking for more AI. They call asking for senior attention, for time, for the judgement that does not come out of a model. If the people were cut to pay for the machines, was the machine ever actually cheaper than the worker?

While the invoices piled up, part of the answer had already been published, in a study of computer vision. In January 2024, researchers at the Massachusetts Institute of Technology in the United States released a working paper with a dry title: Beyond AI Exposure, Which Tasks are Cost-Effective to Automate with Computer Vision. The team was led by Neil Thompson of MIT’s FutureTech project. Most research until then had asked one question: could a machine, in principle, do this task? Thompson’s team asked the question a chief financial officer actually asks, which is whether a business would ever pay for it.

The finding landed like cold water. Only 23% of worker pay exposed to AI computer vision would be cost-effective for firms to automate, because the upfront costs of the systems are so large. Roughly three quarters of the work a camera could technically touch was cheaper to leave with the people already doing it. The paper makes it concrete with a hypothetical small bakery. Checking the quality of ingredients is a sliver of a baker’s day. Add the cameras and the software, and the time and wages saved fall well short of what the upgrade costs.

Zoom out to the whole American economy and the sliver gets thinner. Computer vision could technically automate tasks worth about 1.6% of US worker wages, leaving farming aside. Once you count the cost of building and running the systems, only around 0.4% of wages would actually be cheaper to automate. A sliver of a sliver. Thompson told CNN that in many cases humans are the more cost-effective and more economically attractive way to do the work right now.

The gap shows up job by job as well. About 36% of American non-farm jobs have at least one task a camera could handle. Only about 8% have a task where automating it would actually pay. The distance between could be done and worth doing is where this entire story lives. The authors were not claiming the workforce is frozen. They expect displacement to be substantial but gradual, and their study covered computer vision only, not the language models behind chatbots.

Two years on, a similar gap showed up inside the layoff data itself. At the research firm Gartner, the analyst Tori Paulman and her team went through more than one million layoffs recorded in 2025. They hunted for cuts caused by AI productivity, meaning a tool made one worker so much more productive that the company genuinely needed fewer people. Those cuts came to less than 1% of the total. Less than one in a hundred. And in the second half of that year, firings blamed on AI went down, not up.

So what was everything else? Gartner found that 17% of AI-attributed layoffs in the first half of the year were commercial pivots, companies shifting staff off unprofitable units and onto new AI product lines. Paulman said these companies were likely AI-washing, using AI as cover for ordinary restructuring. Her suspicion started with the headlines. The cuts seemed to be coming from high-tech firms that also sold AI products, and her team wondered whether they wanted to signal they were eating their own dog food.

A separate Gartner study put the knife in. It surveyed hundreds of business executives at companies with at least $1 billion in annual revenue. Of those who had piloted AI or autonomous technology, 80% reported cutting their workforce. They cut whether or not the technology was actually generating a return. Workforce reduction rates were nearly identical for the firms reporting high returns and for the firms whose results had got worse.

Helen Poitevin, a Gartner analyst and key researcher on that study, said looking only at layoffs is shortsighted if you want value from AI. Chasing value only through headcount reduction, she said, is likely to lead most organisations down a path of limited returns. The companies with the biggest gains were using AI as people amplification, making workers more productive instead of replacing them outright. For most of the work a camera could do, the machine was not the cheapest worker in the building. So if the numbers did not justify the cuts, who exactly decided to tell the world that the machines had taken the jobs?

So who pays when the confident machine is wrong? A small claims tribunal in British Columbia gave one of the earliest answers, over 880 Canadian dollars. Air Canada argued its chatbot was a separate legal entity, responsible for its own actions. The tribunal member called that submission remarkable, rejected it, and found the airline liable for negligent misrepresentation. Then the same lesson arrived with bigger numbers. Deloitte handed the Australian government an 440,000 Australian dollars report containing non-existent academic sources and a fake quote from a federal court judgement, then filed a revised version and agreed to a partial refund. The following month, at least four citations to papers that did not exist turned up in its 1.6 million Canadian dollars health workforce plan for Newfoundland and Labrador.

And the savings that justified all those cuts? Orgvue’s arithmetic is blunt: once severance, lost productivity and recruiting replacements are tallied, companies spend roughly $1.27 for every $1 they claw back through workforce reductions. Careerminds found nearly a third of organisations, 30.9%, spent more on rehiring than automation ever saved them. Meanwhile the invoices kept climbing. One industry column reported that Microsoft pulled back internal Claude Code licences after per-engineer costs hit somewhere between $500 and $2,000 a month, and that Uber burned through its entire 2026 AI budget by April. CloudBees chief executive Anuj Kapur said the quiet part out loud: at AI-heavy companies, workforce cuts may simply be the only lever left to offset the AI bill.

Which leaves the question sitting underneath everything. MIT’s researchers found only 23% of the worker pay exposed to computer vision would actually be cost-effective to automate, because the upfront costs are so heavy. Gartner went through more than a million layoffs from 2025 and found cuts genuinely driven by AI productivity were under 1% of the total, while 17% of AI-attributed layoffs in the first half were commercial pivots its analyst said were likely AI-washing. Cutting staff did not separate the winners from the rest; the firms with the biggest gains were using AI as people amplification. Gartner now expects half the firms that trimmed customer service headcount for AI to rehire for similar functions by 2027. By 2029, it expects 30% of AI-replaced workers to be hired back, often at significantly higher cost. So the machine took far fewer jobs than it was blamed for. Somebody signed the paperwork, blamed the software, and then quietly started hiring again.