AI Is Not Taking These 5.8 Million Jobs, It Is Shrinking Their Raises

Real wage growth ran 6.7 percentage points lower for workers in high artificial intelligence exposure occupations than for comparable less exposed workers after 2023, according to a white paper by economists Torsten Slok and Sania Edlich at Apollo Global Management. That covers roughly 5.8 million US employees, about 3.7 percent of the labour force. At Metaintro, we track roughly 50 million job postings, and this reframes the whole AI and jobs debate. The measurable damage so far is not mass unemployment, it is suppressed pay.

What Did the Research Actually Measure?

The gap between wage growth in exposed and unexposed work, rather than headcount. The analysis used a difference-in-differences model across 321 occupations matched to US Bureau of Labor Statistics data covering 2015 to 2025, comparing 11 high exposure occupations, including computer programmers, customer service representatives and financial analysts, against fields with less exposure.

The exposure measure is worth understanding because it is unusually direct. Rather than guessing which jobs an AI could theoretically do, the authors used Anthropic's Economic Index, which is built on actual usage data from Claude. That grounds exposure in what people are really delegating to these systems today rather than in speculation about what might be automated eventually.

The estimated cost is 28 billion dollars in forgone annual earnings, which the authors describe as conservative. They also note the 5.8 million worker estimate may itself be conservative, meaning the affected group is likely larger than the headline figure suggests.

Why Does Wage Compression Show Up Before Job Losses Do?

Because it is the path of least resistance for an employer. Cutting roles is visible, legally involved and expensive in severance and rehiring risk. Holding pay flat while a tool absorbs part of the work is none of those things. The employee keeps the job and quietly stops gaining ground, and there is no announcement, no filing and no headline.

That is why the public conversation has been mismatched with the evidence. Coverage has concentrated on displacement while the measurable effect has been on compensation, and the two require completely different responses from a worker. If your risk is redundancy, the answer is contingency planning. If your risk is a flattening pay curve, the answer is repricing yourself, which is a very different activity.

It also explains why the effect can persist without ever feeling like a crisis. A raise that comes in two points below what it would have been is not an event anybody notices in a given year, but compounded across several years it produces a large and permanent gap. We traced a related pattern in the widening gap between AI driven productivity and worker pay.

Who Is Actually Losing the Most?

The people with the least room to absorb it, which is the finding that matters most. Wage growth for the bottom quartile of earners fell 10.7 percent, the second quartile 5.4 percent and the third quartile 4.0 percent, while the top quartile showed no statistically significant effect at all.

By occupation group the pattern is just as uneven. Service occupations fell 24.3 percent, although the authors flag that this rests on a small subsample and should be read with caution. Management and professional roles fell 4.1 percent, and blue collar workers showed no significant effect.

That distribution overturns the assumption that has dominated this debate for two years. The expectation was that AI would squeeze the middle and spare both ends, or that it would come for senior knowledge work first. What the data shows instead is an effect concentrated at the bottom of the income ladder and largely absent at the top, which makes this a story about inequality as much as about technology. Related work from the Federal Reserve Bank of St. Louis on AI and young adult unemployment points in a similar direction.

Is Your Own Role Exposed?

The practical test is not whether AI could replace your job but whether it can already do a meaningful slice of your weekly tasks. The occupations named in the analysis are instructive, since computer programmers, customer service representatives and financial analysts are all roles where a large share of the work is producing text, code or analysis from a defined input. That is precisely the shape of task these systems handle well.

Three questions separate exposure from safety in practice. First, what proportion of your week produces an output that follows from an input by rules rather than judgment. Second, how much of your value depends on accountability, relationships or physical presence that a model cannot supply. Third, whether your employer has already deployed a tool into your workflow, because the compression effect tends to follow adoption rather than precede it.

There is one important caveat the authors themselves raise. Some of the measured gap may reflect the AI infrastructure buildout pushing pay up in low exposure trades such as electricians and construction work, rather than pay being pushed down in exposed roles. The relative gap is real either way, but the mechanism matters, and it is a reminder that the trades are currently benefiting from the same wave that is squeezing office work. Our analysis of how the labour market is splitting between AI exposed office jobs and hands on hiring covers that divide.

What Should You Do if Your Pay Is Being Compressed?

Change what you are being paid for, because arguing about the old work will not win. If a tool now does a share of your previous output, your value has moved to what you do around it, which is judgment about when the output is wrong, accountability for the decision and the ability to handle the cases the system cannot. Compensation conversations go better when they are framed around that shift than around effort or tenure.

Bring evidence rather than a feeling. The most persuasive version is specific, describing the volume you now handle, the quality problems you caught that a tool missed, and the work you absorbed that used to belong to someone else. If your productivity rose because of a tool your employer deployed, that gain is real and it is reasonable to ask to share in it. Our guides to salary negotiation tactics that work and asking for a raise cover how to structure that conversation.

Be aware that internal raises may be the harder route in an exposed role. Compression happens inside existing employment relationships, so an external offer is often the faster correction, and moving is frequently the only way to reset a pay level that has drifted below market. Our guide to negotiating salary in an interview matters more than usual for anyone in this position.

The longer play is to move toward where the wage effect is absent. The top quartile showed no statistically significant effect and blue collar work showed none either, so both seniority within knowledge work and skilled hands on trades currently sit outside the squeeze. That points to two very different but equally valid strategies, and the wrong move is to stay in a compressed middle and hope it passes. Technology has continued to lead wage growth despite the layoff headlines, which shows the sector is not uniformly compressed.

Why Did This Effect Go Unnoticed for So Long?

Because nothing about it generates a news event. A layoff produces a filing, a press release and a wave of coverage. A raise that lands two points lower than it would have produces a slightly disappointing conversation with a manager, repeated separately across millions of households with no shared moment that makes it visible.

The measurement problem compounds the visibility problem. To see this effect at all, you have to compare wage growth in exposed occupations against a matched control group of unexposed ones, which is exactly what the 321 occupation model was built to do. Looking at exposed occupations alone would show wages still rising, just more slowly, and slower growth is easy to attribute to the economy rather than to the technology.

There is also a reporting incentive at work. Companies that want to signal efficiency to investors have every reason to announce automation driven job cuts loudly, and no reason at all to announce that they have quietly stopped raising pay in a function where a tool now does part of the work. The visible half of the story is the half someone benefits from telling.

What Happens to the Compressed Middle From Here?

Two plausible paths, and the difference matters for how you plan. In the first, compression is temporary and pay recovers once employers work out how to redeploy the capacity a tool has freed up, meaning workers who sit tight eventually see their curve resume. In the second, the compressed level becomes the new baseline, because a role that has been repriced downward rarely reprices back up without the worker changing employer.

The evidence available so far does not settle which path applies, but the asymmetry of the risk is clear enough to act on. If you assume recovery and are wrong, you lose several years of compounding at a lower base. If you assume permanence and are wrong, you have merely repriced yourself earlier than you needed to. The second error is much cheaper than the first.

That argues for treating a flattening pay curve as a prompt rather than a phase. The workers most likely to come out of this period whole are the ones who moved while the labour market still had reasonable mobility, not the ones who waited to see whether their employer would eventually correct it.

Does This Mean AI Is Not Costing Jobs at All?

No, and the research does not claim that. It says the dominant measurable effect so far, across these 321 occupations over this period, has shown up in wages rather than in employment levels. Announced job cuts attributed to AI are real and we track them separately in our 2026 tracker of companies citing AI in layoffs.

The more accurate framing is that these are two stages of the same adjustment. Wage compression is what happens while a technology is absorbing tasks inside existing roles, and structural job change is what happens when employers eventually redesign those roles around the tool. Seeing the first stage clearly is useful precisely because it gives you time to act before the second arrives.

For a worker, the practical consequence is that a stable job title is not evidence that nothing is happening. Your pay trajectory is the earlier signal, and if your raises have quietly flattened while your output has not, that is worth treating as information rather than as bad luck.

What Should Employers Take From This?

That the cost of compression is not zero, even though it does not appear on a severance line. A workforce whose pay quietly stops tracking its output notices, and the resulting disengagement is expensive in ways that are harder to measure than headcount savings. Sharing productivity gains is not only an equity argument, it is a retention argument.

There is also a straightforward reputational point. The distribution of this effect, falling hardest on the lowest paid and not at all on the highest, is exactly the pattern that turns a technology story into a fairness story. Employers deploying these tools without adjusting pay bands for the people absorbing the change are accumulating a problem rather than avoiding one.

Related Articles

• The AI Productivity and Pay Gap
• Technology Leads Wage Growth Despite the Layoff Headlines
• St. Louis Fed on AI and Young Adult Unemployment
• The US Labour Market Is Splitting Between AI Office Jobs and Hands On Hiring
• Salary Negotiation Tactics That Work
• How to Ask for a Raise
• How to Negotiate Salary in an Interview
• Companies That Cited AI in Their 2026 Layoffs
• Companies Spending Most on AI and Hiring Fastest
• AI Certifications That Boost Salary and Career Growth
• Adaptability Beats Raw Experience for Career Growth
• How Much Is Six Figures, Really

People Also Asked

Q: Is AI actually reducing wages or just slowing raises?

A: It is slowing growth rather than cutting pay outright. The research measured real wage growth running 6.7 percentage points lower for exposed workers than for comparable unexposed workers after 2023. Your salary does not fall, it stops climbing at the rate it otherwise would have, which compounds into a large gap over several years.

Q: Which workers are most affected by AI wage compression?

A: The lowest paid by a wide margin. The bottom quartile of earners saw wage growth fall 10.7 percent while the top quartile showed no statistically significant effect. Service occupations recorded the steepest fall at 24.3 percent, though that figure rests on a small subsample.

Q: How do I know if my job is AI exposed?

A: Look at your actual weekly tasks rather than your job title. The occupations studied included computer programmers, customer service representatives and financial analysts, all roles where much of the work turns a defined input into text, code or analysis. The larger that share of your week, the more exposed you are.

Future-proof your career before the effect compounds. Metaintro tracks roughly 50 million job postings so you can see what employers are paying right now for roles that value judgment alongside AI fluency, and you can create a free account to get matched with openings that reprice what you already do.