Business Ethics of AI Layoffs: What Klarna, Duolingo and IBM Said, Then Unsaid

In December 2024, Klarna's chief executive said AI could already do all of the jobs that humans do. Five months later, Klarna was recruiting human support agents again. The business ethics of AI layoffs is usually argued about the cuts themselves, but the cleaner test is the messaging. Klarna, Duolingo and IBM each credited AI for a workforce decision in public, then narrowed the claim. We put 11 dated statements next to what their own filings show.

The short answer

Crediting AI for a workforce decision is not unethical in itself. It becomes an ethics problem when the claim is broader than the evidence, because employees, customers and investors all act on it. The three cases here share one fault: a headline claim with no denominator, no time window and no split between layoffs, attrition and contractors.

Why AI layoff messaging is a business ethics question

A workforce statement is not a press line. It is read by at least three audiences who make decisions on it, and each audience is owed something different.

Your employees read it as a forecast about their own jobs. Your customers read it as a statement about who will answer when something goes wrong. Your investors read it as a claim about future operating costs. A single sentence like "AI can do the work of 700 agents" lands differently in each room.

The ethical duty that covers all three is not complicated. Say what happened, in units someone could check. That is the same standard this blog applies to vendor claims in the forensic reading of AI ROI case studies, and it applies with more force when the subject is people's jobs.

Whether these reductions were driven by AI, by earlier overhiring, or by both is examined in the analysis of AI layoffs as an overhiring correction. This post is narrower. It asks what each company said, what it said next, and what its filings record.

My position, stated up front: the companies below were not lying. They were doing something more common and, at scale, more corrosive. They let the most quotable version of a claim travel, then corrected it in a quieter venue.

Klarna: from "all of the jobs" to "always a human"

Klarna is the case everyone cites, and its record is the most complete of the three.

The claims, February to December 2024

In February 2024, Klarna announced that its OpenAI-powered assistant was handling two-thirds of its customer service chats in its first month. The company equated the workload to 700 full-time agents. That "700" was a workload equivalence, not a count of people dismissed, and later coverage described the displaced capacity as contract staff.

In December 2024, chief executive Sebastian Siemiatkowski went further on Bloomberg TV. "I am of the opinion that AI can already do all of the jobs that we as humans do," he said. He added: "We stopped hiring about a year ago. We were 4,500, now we're 3,500."

He described the mechanism plainly. "People stay about 5 years, so 20% leave every year, and by not hiring, we're simply shrinking." That is an attrition story, not a layoff story, and it was the more accurate half of the interview.

The correction arrived within days, and not from him. TechCrunch found Klarna advertising more than 50 open roles. A Klarna spokesperson told the outlet the CEO was "simplifying for brevity in a broadcast interview". Klarna, he said, was "only backfilling some essential roles, predominantly engineering."

The correction, May 2025, and what the filing says

In May 2025, Siemiatkowski told Bloomberg that Klarna was hiring human support staff again. "I just think it's so critical that you are clear to your customer that there will always be a human if you want," he said. He conceded that the cheaper AI service had produced "lower quality" output. "Really, investing in the quality of human support is the way of the future for us."

Between those two interviews sat a document with legal liability attached. Klarna's F-1, filed with the SEC in March 2025, says its AI assistant "has handled 62% of customer service chats since its launch." That is a smaller figure than the launch release, and it is the honest one, because it covers a longer window.

The same filing makes the financial case directly. Average annual revenue per employee rose from approximately $344,000 in 2022 to approximately $821,000 in 2024. The filing credits "increasing use of AI and a focus on cost optimization" together, which is a more careful sentence than anything said on television.

Duolingo: the AI-first memo and the contractor line

Duolingo's case is different in kind. Its chief executive did not claim AI had replaced anyone. He announced a policy, and the policy contained one sentence that did all the damage.

The memo and the contractor line

The groundwork was laid in January 2024. Duolingo confirmed it had cut around 10% of its contractor workforce at the end of 2023. Its account of why shifted between outlets. Bloomberg was told part of the reason was AI. Semafor was told it would be an "oversimplification" to call AI the sole reason. TechCrunch was told only a "small minority" of contractors left as their projects wrapped.

In late April 2025, Luis von Ahn sent an all-hands email that Duolingo then posted publicly. It declared the company "AI-first" and compared the moment to its bet on mobile in 2012. The line that travelled was this one: "We'll gradually stop using contractors to do work that AI can handle." The memo also said, per Fortune, that new headcount would be added only if a team could not automate more of its work.

Two clarifications in four months

The backlash was immediate. In May 2025, von Ahn posted on LinkedIn: "To be clear: I do not see AI as replacing what our employees do (we are in fact continuing to hire at the same speed as before)."

In August 2025, he went further in an interview with the New York Times. "This was on me. I did not give enough context," he said. "We've never laid off any full-time employees. We don't plan to."

Read closely, the memo and the clarification are compatible. Contractors are not employees, and the memo only ever named contractors. That is precisely the problem. A distinction that matters to the company's lawyers does not register with the readers who decide whether to keep the app.

Duolingo's own filing now records the price. Its FY2025 10-K states that the April 2025 memo "may have contributed to unfavorable publicity, adverse impacts on the Company's brand and social media presence, and a deceleration in user growth." The same filing describes a team of "more than 900 passionate employees, including more than 430 engineers" at the end of 2025.

I think that risk factor is the most important sentence in this entire record. It is a company telling the SEC, in writing, that the way it talked about AI and jobs may have cost it growth.

IBM: 7,800 jobs as a sum, a few hundred in practice

IBM's record runs longest, from May 2023 to November 2025. It is also the case where the gap between the headline and the filing is easiest to measure, because IBM publishes a headcount table every year.

The 2023 estimate

In May 2023, Arvind Krishna told Bloomberg that hiring in back-office functions such as human resources would be suspended or slowed. Those non-customer-facing roles, he said, amounted to roughly 26,000 workers. "I could easily see 30% of that getting replaced by AI and automation over a five-year period."

The number that went around the world was 7,800. Krishna never said it. It was the reporter's multiplication of his two figures. An IBM spokesperson added that part of any reduction would come from not replacing roles vacated through attrition.

This is a case where the company's statement was more careful than its coverage. Even so, a chief executive who offers a percentage and a base in the same sentence should expect the product to be printed.

The 2025 revision and the 10-K

Two years later, Krishna told the Wall Street Journal that AI had replaced a few hundred HR roles, not thousands. He said that despite "a huge amount of work inside IBM" on AI and automation, "our total employment has actually gone up." The savings, he explained, gave IBM more to invest in other areas.

IBM's own annual reports do not show that rise in its core workforce. The Human Capital table counts 282,200 employees at IBM and wholly owned subsidiaries for 2023, 270,300 for 2024 and 264,300 for 2025. The table also lists less-than-wholly owned subsidiaries and a "complementary" workforce of temporary and limited-term staff, which move separately.

Then, in November 2025, IBM told Bloomberg: "In the fourth quarter we are executing an action that will impact a low single-digit percentage of our global workforce." This time AI was not the stated reason. The company described a routine rebalancing toward software.

Each of these statements can be defended on its own. Laid side by side against the filing, they do not add up to a story a reader could check.

The pattern across all three AI layoff timelines

Put the 11 statements on one axis and a shape appears. The broad claim comes first, in the venue with the largest audience. The narrowing comes later, in a spokesperson email, a LinkedIn post or a regulatory filing.

Three things repeat across the table. First, the broad claim and the correction are rarely made by the same person in the same venue. Second, every original claim omits at least one of the three facts a reader needs: the base, the window, and the mechanism. Third, the filing is always the most precise document and the least read.

There is also an asymmetry in how the claims served the companies. The broad AI claim told investors a cost story at the moment it was most useful, around an IPO for Klarna and a strategy pivot for Duolingo. The correction told employees and customers a reassurance story once the backlash arrived.

That is the business ethics finding of this exercise. The problem is not that AI was credited. It is that the credit was sized for one audience and then resized for another.

Where the ethics case is weakest

A post arguing for precision has to apply it to itself. Here is where the argument above is soft.

Forecasts and opinions are not false statements

Krishna offered a five-year estimate in 2023. Siemiatkowski offered an opinion about what AI can do. Von Ahn announced a direction. None of those is a factual claim that later evidence can falsify, and treating a revised forecast as a retraction is unfair to the speaker.

Executives also have to communicate strategy before it is proven. Some of what reads as backtracking is ordinary learning in public, and Klarna's quality admission is arguably the most candid sentence in the record.

The data cannot settle cause

The filings show headcount and revenue per employee. They cannot show how much of a change came from AI, from attrition, from earlier overhiring, or from a weak market. My IBM comparison also excludes the complementary workforce, and Krishna may have been measuring a different base or period when he said employment had risen. I cannot rule that out, and neither can anyone outside the company.

The economy-wide picture is similarly unhelpful to a strong claim. AI-cited cuts remain small against total layoffs, as the comparison of AI job loss announcements with payroll data shows. These three companies are cases, not a sample, and they were chosen because their statements were prominent, which biases the record toward loud claims.

So the defensible conclusion is narrow. These companies communicated imprecisely about AI and jobs, and the imprecision had costs. The record does not support the stronger charge that AI was used as a deliberate cover story.

A disclosure standard for AI workforce statements

If the failure is precision, the fix is a standard. Below is the one I would hold any AI workforce statement to, including one from a company I worked for.

The last row is the one that would have prevented most of this record. If your chief executive would not put a sentence in the annual report, it should not go on television.

There is a commercial argument here as well as an ethical one. Customers increasingly ask who, or what, is serving them, and the evidence on how disclosure affects trust is laid out in the piece on AI disclosure and customer trust. Klarna's own correction was framed around the customer's right to reach a person, which is a trust argument, not a cost argument.

What I would do before the next announcement

I would require one internal document before any public AI workforce claim. It states the headcount at two dates, splits the change into layoffs, attrition and contracts, and names the AI system and the workload it absorbed. If the statement cannot be traced to that page, it does not go out.

I would also put the most cautious version of the claim in the highest-reach venue. That runs against every instinct in a communications team. It is also the only way to avoid writing the correction four months later.

Frequently asked questions

How to write the next workforce statement

Pull your last public statement that mentioned AI and headcount, whether a post, an earnings remark or an all-hands email. Check it against your own HR system for the same dates. If the two disagree on direction, you already have a correction waiting to happen, and it is cheaper to issue it on your own schedule.

Then write the one-page internal record described above before anyone drafts the next announcement. The companies in this post each spent months correcting a sentence that one page would have made unnecessary. For what AI-native headcount looks like when it is measured, read the analysis of revenue per employee at AI-native companies.

Related reading