AI Layoffs: What’s Actually True, From Someone Who Trains the Models

I got talking to an HR manager at a coffee shop the other day, and within five minutes we were on AI layoffs. She said people are getting cut constantly, because big tech companies have put AI on work that humans used to do. I think she’s half right, and the half she’s missing is the part I see every day.

The short answer

I’m an AI trainer. My work is reviewing and scoring what models produce, checking it against what a competent person would have done. So when people ask “will AI take my job?” or claim AI can do almost everything now, I don’t hear a headline. I hear a claim I can test against my own screen.

AI Layoffs: What the Numbers and the Companies Actually Say

The cuts are real. Amazon cut 14,000 corporate jobs in October 2025 and about 16,000 more in January, roughly 30,000 in total. Microsoft cut over 15,000 in 2025. Meta cut about 8,000 (roughly 10%) in May, per TechCrunch’s running list.

What changed is the wording. Layoffs.fyi’s founder told Crunchbase that AI was cited in 33% of tech layoff events this year, up from 1% in 2024.

Here’s what a few companies have actually said or done:

• Apple: cut about 200 roles in August 2026 across Vision Pro, Siri and AI teams, which 9to5Mac’s report ties to a reorganization around a new Siri AI platform, not AI doing those jobs. Its last big cut was 614 people in 2024, after it cancelled its car project.
• Salesforce: took its support team from 9,000 to about 5,000 in 2025. Benioff said he needed “less heads”, and also that hundreds of people were moved into other roles.
• Klarna (fintech, but the famous case): said its chatbot did the work of 700 support agents, then started hiring humans again after service quality dropped.
• Epic’s CEO said its layoffs “aren’t related to AI”.

My read: only a few say AI directly took the jobs. Most describe reorganizing, or freeing up money for AI spending. That’s my interpretation, not a company statement.

The people who never show up in the numbers

Layoff trackers count employees. But a lot of platform work, content moderation especially, is done by outside vendors, so those cuts never land on the company’s own headcount. In March 2026, Meta said it would cut back on third-party vendors and contractors in favor of AI for content enforcement, over a few years. It has used contractors from firms like Accenture, Concentrix and Teleperformance. It also said it will still use experts to design, train and oversee those systems, and keep humans on the most complex decisions.

Then there’s the money. Meta raised its 2026 capital spending guidance to $125 billion to $145 billion and its stock dropped. By June it was down 13% over twelve months. Investors weren’t picking AI over people. They were nervous about the size of the AI bill, while revenue grew about a third in the first quarter.

My read: when the bill gets that big, other costs get squeezed, and headcount and vendor contracts are on the list. Meta didn’t say that.

Who is less exposed

Not “safe.” Less exposed. A recent paper on career choice found newer models show lower exposure for physical occupations, and well paid, low exposure jobs show up most in hands-on work, including trades. Healthcare practice showed the strongest combination of good pay and low exposure of any field in the paper.

Experience matters too. Stanford tracked workers aged 22 to 25: between November 2022 and June 2026, employment in the two most exposed groups of occupations fell about 11%, while the three least exposed grew about 10%. Employment also rose among experienced workers in jobs built on knowledge you pick up by doing the work. The authors call these early signals, not proof of cause, and say there’s no widespread displacement.

And the ILO estimates 1 in 4 jobs worldwide is exposed, but says transformation, not replacement, is the most likely outcome. My read: the pattern isn’t a job title. It’s whether your work depends on things you can only learn by being there.

What I see, and what I’d do

I’m one person in one corner of this industry (I’ve already written about how to tell a real AI trainer job from a scam), so take this as a view, not a forecast.

I train chat AI to understand people better, and it’s still not fully there. The model answers well as long as the question matches the data it was trained on. The moment it doesn’t, things fall apart. Airline chatbots are a good example: ask about a standard booking issue and it handles it fine, but ask something outside the airline’s written policy, the kind of edge case a human agent would use judgment on, and the model doesn’t just get the answer wrong, it visibly stops understanding the question. That’s usually when people get frustrated with AI. Not because it’s slow, but because it clearly isn’t following what they’re actually asking. Phone support is even further behind than chat. Unless a company is running one of the frontier models under the hood, live voice calls still need a human.

If you’re asking will AI take my job, I don’t know if any of this will still hold in a year. But two things feel safe to say. Find the most repetitive, rule-based part of your week, because that’s where the pressure lands first. And build some income that doesn’t depend on one employer’s headcount decisions, before you need it. If you’re weighing a side project, run the 5-minute math check first, and see your runway with the emergency fund math for Canadians.

Frequently Asked Questions

KEEP READING

• How Much Emergency Fund Do You Actually Need in Canada?
• Before You Start a Side Hustle, Run This 5-Minute Math Check
• The Side Hustle Numbers Nobody Agrees On
• Free Side Hustle Income Tracker (Spreadsheet You Can Copy Right Now)