AI Is Pressuring Pay Before It Eliminates Jobs

Google News carried a stark Axios finding on July 31: workers exposed to AI kept their jobs, but their wage growth fell sharply after 2023.

The underlying Apollo research challenges the familiar automation story. Employers do not need to eliminate a position to capture more value from AI. They can raise output expectations, weaken a worker’s bargaining position, or hire replacements at lower starting salaries.

That distinction matters for programmers, customer service representatives, financial analysts, and other knowledge workers. The immediate contest is not simply humans against machines. It is worker bargaining power against employers seeking to retain AI-driven productivity gains.

Apollo economists Sania Edlich and Torsten Slok found no statistically significant employment decline in the occupations they studied. However, real wage growth in highly exposed occupations trailed growth in less exposed fields by 6.7 percentage points after 2023.

Their estimate places the annual wage-growth shortfall at $28 billion across 5.8 million workers. The study also found a larger effect among lower-paid workers.

This is an early result, not a settled verdict. Other studies have found no meaningful earnings effect, while some detected wage gains in AI-exposed work. The disagreement makes the wage channel more important to watch, not less.

The New Warning Is About Pay, Not Pink Slips

Apollo’s finding moves the first visible cost of AI from unemployment into workers’ paychecks.

The researchers compared wage changes across occupations with different levels of generative AI exposure. Exposure describes how closely a job’s tasks overlap with work that current AI systems perform.

They classified 11 occupations as highly exposed using real-world task data from Anthropic. These fields included computer programming, customer service, and financial analysis.

The comparison covered Labor Department wage data before and after ChatGPT’s late-2022 release. Apollo treated 2023 as the first full year of the generative AI period.

According to the resulting white paper, real wage growth in highly exposed fields fell 6.7 percentage points behind less exposed occupations. Apollo did not find a significant employment effect.

That combination creates the article’s central reversal. Workers can remain employed while losing ground.

Axios summarized the practical consequence plainly in its wage-growth finding: people may keep working but make less than they otherwise would have earned.

“Lower pay” does not necessarily mean every affected worker receives a direct salary cut. The effect can appear through smaller raises, weaker starting offers, or slower advancement.

A company can also leave a vacated position unfilled. Existing employees then absorb more work with AI assistance, often without a proportional increase in compensation.

Apollo estimated that the gap represented $28 billion in missed annual wage growth. It applied to roughly 5.8 million workers, equal to about 3.7% of the U.S. labor force.

The authors described that population as a conservative estimate. The exposure measure covers a limited set of occupations where observed AI use already overlaps heavily with job tasks.

Lower-paid workers within the exposed group experienced the largest slowdown. That result complicates claims that accessible AI tools will automatically democratize expertise and narrow income differences.

A customer service agent, for example, can use an AI assistant to retrieve information or draft replies faster. The employer may then expect each agent to handle more conversations.

That productivity improvement does not guarantee a higher wage. If the software makes workers easier to train or replace, their negotiating position can deteriorate.

The same mechanism can affect junior analysts and programmers. AI may let more applicants complete baseline tasks, increasing competition for work that once required scarcer technical skills.

None of these changes requires an immediate layoff announcement. They can unfold through routine compensation reviews, hiring decisions, and revised performance targets.

That makes wage pressure harder to identify than job destruction. A terminated position leaves a visible record. A raise that never materialized does not.

Why AI Can Weaken Bargaining Power Before Hiring Collapses

AI changes compensation when it reduces the scarcity of a worker’s tasks, even if a human still occupies the job.

Wages reflect more than current productivity. They also reflect outside opportunities, replacement costs, labor demand, and a worker’s willingness to leave.

Generative AI can influence each factor before it becomes capable of automating an entire occupation. A job is a collection of tasks, and software can alter only part of that collection.

Suppose an analyst spends time gathering data, summarizing documents, building a model, checking assumptions, and presenting conclusions. An AI system might accelerate the first three tasks while remaining unreliable on the last two.

The company still needs the analyst. However, it may need fewer analysts to produce the same volume of work.

Management can capture that change without dismissing anyone immediately. It can slow hiring, increase workloads, or wait for normal turnover to reduce headcount.

Workers notice the weaker demand around them. They become less confident about requesting raises or leaving for another employer.

Job-to-job movement matters because changing employers has historically helped many workers secure larger pay increases. A cautious labor market removes some of that leverage.

Axios noted that the U.S. quits rate had returned to levels associated with the mid-2010s. That is a very different environment from the Great Resignation, when employers competed intensely for workers.

AI messaging can amplify the caution. Executives regularly tell investors that automation will reduce costs and improve efficiency.

Even before those plans fully materialize, workers hear the same message. Someone who believes an employer can automate part of a role has less confidence during a salary negotiation.

Employers also gain more information about potential substitutes. They can test whether an AI-assisted junior employee, contractor, or offshore team can perform work previously assigned to experienced staff.

This does not mean companies possess all the bargaining power. Experienced workers still carry institutional knowledge, customer relationships, judgment, and responsibility for costly errors.

AI output also requires review. A confident but incorrect answer can create legal, financial, security, or reputational costs.

Yet the employer does not need complete automation to strengthen its position. It only needs a credible alternative to the worker’s current way of performing a task.

Anthropic’s original task-use data showed that 57% of observed Claude use supported human augmentation. The remaining 43% reflected automation patterns.

Augmentation means a person collaborates with AI and remains involved in the work. Automation means the system directly performs a task with limited human participation.

Those categories can lead to different wage outcomes. Augmentation can make expertise more valuable when AI expands what a skilled worker can accomplish.

It can also reduce the value of experience when the software gives less experienced workers access to guidance, templates, and technical explanations.

Customer support research has demonstrated this ambiguity. AI assistance can produce larger productivity gains for less experienced agents because it spreads practices used by stronger performers.

That outcome can help workers learn. It can also compress the premium once paid to employees who possessed those practices independently.

The distribution of gains therefore matters as much as the productivity increase. Higher output can raise wages, reduce prices, expand profits, or produce some combination of all three.

Apollo’s interpretation is that companies are currently capturing more of the benefit in exposed occupations. Slower wage growth is the evidence offered for that transfer.

The mechanism is plausible, but the data cannot directly observe every compensation negotiation. It infers an AI effect from differences between occupational groups over time.

That limitation becomes central when evaluating whether AI caused the wage gap.

What the Google News Headline Does Not Establish

The reported wage gap is consequential, but one working paper cannot isolate every force moving pay across occupations.

Apollo compared 11 highly exposed occupations with fields that showed less AI exposure. That approach offers a clear test, but it also creates potential confounding factors.

One complication comes from the infrastructure behind AI. Data centers require electricians, construction workers, cooling specialists, and other employees whose roles have relatively low generative AI exposure.

Rapid demand for those workers can raise their wages. The gap would then reflect both weak growth in exposed work and unusually strong growth elsewhere.

A second complication is the broader technology labor market. Software companies adjusted hiring after the pandemic expansion, rising interest rates, and earlier overstaffing.

Programmers could experience weaker wage growth for reasons that overlap with AI adoption but are not caused entirely by it. Separating these forces requires longer time series and more detailed employer data.

The study also uses exposure rather than verified replacement. An occupation can contain tasks that AI handles without every employer deploying the technology effectively.

Actual adoption varies by company size, management quality, security rules, and access to reliable data. Exposure measures technical opportunity, not uniform workplace use.

A March 2026 Fed review illustrates the unsettled evidence. It cites studies reaching different conclusions about employment and wages.

Some research found little overall effect on jobs or postings, alongside statistically significant wage gains in exposed occupations. Other work detected reduced entry-level employment where AI mainly automated tasks.

The Fed also cited surveys showing rapid adoption. One survey found workplace use of large language models rose from 30.1% in December 2024 to 45.9% by June and July 2025.

That growth makes occupational comparisons increasingly relevant. It does not resolve whether AI exposure causes lower wages across the economy.

Evidence from Denmark offers another counterweight. Researchers linked worker surveys with administrative employment records after employers introduced chatbot programs.

The Danish registry study found no meaningful effect on earnings or recorded hours during the first two years after ChatGPT’s release. Its estimates ruled out effects larger than 2%.

The Danish labor market differs from the American one. Its institutions, wage-setting practices, safety net, and employer behavior limit a direct comparison.

Still, the null result shows that Apollo’s finding should not become a universal rule. AI’s wage effect can depend on national institutions and workplace implementation.

Another working paper, revised in August 2026, supports the wage-pressure thesis using different American datasets. Its results arrived after the original Axios report.

The paper’s wage estimates associate a move from low to high generative AI exposure with a 4.9% wage decline in occupational data.

Using job postings and matched employment records, the authors found declines of 8.71% in advertised wages and 10.76% in starting wages. They detected no corresponding employment change.

Those estimates strengthen the case that compensation can move before headcount. However, the paper is still a working paper and has not settled the causal debate.

Federal Reserve Governor Michael Barr emphasized that uncertainty in a February 2026 speech. He said there was little evidence of a meaningful aggregate wage effect at that point.

Barr identified three decisive questions: whether AI complements expertise, how it changes demand across occupations, and who owns the resulting capital.

That framework explains why apparently conflicting studies can coexist. One tool can raise a worker’s output while reducing the market value of the worker’s original skills.

The outcome depends on what companies do with the saved time. They can expand production, improve service, cut employment, or simply demand more from the existing workforce.

It also depends on competition. A business that passes productivity gains to customers may need more workers as demand grows.

A company with strong market power may keep the gains as profit. A worker with scarce complementary expertise may negotiate a share.

The Google News headline therefore identifies a serious possibility, not a final measurement of AI’s nationwide effect. Readers should treat the 6.7-point gap as an early signal.

Lower Earners Face the Hardest Version of AI Wage Pressure

Workers with fewer financial cushions can suffer even when AI leaves their formal employment status unchanged.

Apollo found that wage growth deteriorated more among lower-paid workers in its exposed occupational group. That distributional result carries consequences beyond the average wage estimate.

A highly paid specialist may absorb a weak raise without immediately changing housing, health care, or education decisions. A lower-paid worker has less room to adjust.

Inflation makes the distinction sharper. A salary that remains unchanged represents a loss in purchasing power when essential costs continue rising.

Slower nominal raises can therefore produce a real decline in living standards. The worker remains employed, but the household becomes less secure.

Lower earners also have fewer resources for retraining. Courses require time, money, predictable schedules, and confidence that a new credential will improve employment prospects.

AI may simultaneously change which skills employers value. Workers must then learn new tools while meeting higher output expectations in their current roles.

Customer service provides a concrete example. An AI assistant can draft responses, summarize conversations, and suggest troubleshooting steps.

Management can view those capabilities as support for agents. It can also raise the number of cases each agent must close during a shift.

The worker becomes responsible for reviewing more machine-generated material. Errors remain attached to the human employee, even when the software produced the faulty suggestion.

Performance measurement may not recognize that added oversight. The dashboard records faster handling, while the employee carries greater cognitive and emotional strain.

Entry-level knowledge work faces a related problem. Junior positions have traditionally allowed workers to learn through research, drafting, testing, and routine analysis.

These are precisely the tasks that current generative AI handles most readily. Automating them can remove part of the training path to more senior work.

Employers may still hire entry-level workers, but at lower starting salaries. They may argue that AI reduces the experience required to become productive.

The later August wage study found its largest estimate in starting pay. That pattern is consistent with employers resetting compensation when filling new positions.

It is also consistent with other labor-market changes. Researchers need more evidence before assigning the entire decline to generative AI.

The critical issue is mobility. A worker who cannot secure a better offer has limited leverage over a current employer.

The same worker may hesitate to leave because competing firms advertise fewer positions or expect proficiency with unfamiliar AI systems.

This is how wage pressure can spread without a dramatic unemployment spike. The labor market remains outwardly stable while advancement becomes harder.

Existing inequality can shape the result. Workers with savings can wait for better offers, take time to train, or reject aggressive productivity targets.

Workers living paycheck to paycheck often cannot. Their need for immediate income gives employers more room to set terms.

AI ownership compounds the imbalance. Companies purchase the systems, control workplace data, define performance measures, and decide how productivity savings are allocated.

Employees usually contribute the domain knowledge needed to make those systems useful. They correct outputs, explain exceptions, and identify errors that generic models miss.

Whether that contribution receives compensation is a management and labor-market question, not a technical inevitability.

Workers can improve their position by documenting the outcomes they create with AI. Output volume alone may become less persuasive as the tools become common.

Judgment, error prevention, client retention, and process design provide stronger evidence of value. These contributions remain tied to the person who directs and verifies the system.

A personal knowledge system can help preserve that evidence. It can connect decisions, source material, corrections, and completed work across projects.

That record does not eliminate structural wage pressure. It can make an individual’s contribution easier to explain during a review or job search.

Employers also face risks if they push compression too far. Lower compensation can increase turnover, reduce trust, and discourage workers from reporting model failures.

A company that treats AI only as a labor-cost instrument may lose the expertise required to supervise it. Short-term savings can then produce slower decisions and expensive errors.

The more sustainable approach treats workers as participants in productivity gains. That can include compensation, advancement, training, or greater control over how saved time is used.

Apollo’s findings suggest that such sharing is not happening evenly. Lower earners appear to be absorbing the harshest adjustment.

Three Signals Will Show Whether the Wage Threat Is Spreading

The next test is whether wage compression appears across more occupations, more datasets, and more employer decisions.

The first signal is starting pay in AI-exposed occupations. Starting salaries react faster than compensation for established employees because employers can reset terms with each hire.

Researchers should compare advertised and accepted wages within the same occupation, industry, location, and experience level. That design can reduce confusion from broader hiring cycles.

Continued starting-wage declines would strengthen Apollo’s interpretation. Stable or rising offers would suggest that the first occupational comparison captured temporary technology-sector weakness.

The second signal is job-to-job mobility. Workers gain bargaining power when they can leave for better opportunities, while employers gain it when outside offers disappear.

The quits rate provides a broad measure, but occupation-level movement will be more informative. Analysts need to track whether exposed workers change employers less often than comparable workers.

A decline concentrated in AI-intensive roles would support the bargaining-power mechanism. A market-wide decline would point toward broader economic conditions.

Researchers should also examine the pay increases workers receive when they do move. Fewer transitions combined with smaller wage gains would indicate weaker outside opportunities.

The third signal is how employers distribute measurable productivity gains. Companies increasingly discuss AI efficiency in earnings calls, internal targets, and workforce plans.

Those statements should be compared with headcount, compensation expense, hiring volume, service quality, and revenue per employee.

If productivity rises while real compensation stagnates, the wage-compression thesis becomes stronger. If employment or pay expands with output, augmentation is producing a broader benefit.

Corporate adoption will also reveal which occupations enter the exposed group. Apollo’s 5.8 million workers represent a narrow early estimate, not a ceiling.

AI agents that handle longer workflows could extend exposure beyond isolated writing or analysis tasks. Reliability will determine how quickly employers trust them with that responsibility.

Human review remains an important constraint. Systems that fail on complex, long-running tasks continue to require skilled supervision.

That supervision can support higher wages when accountability and domain knowledge remain scarce. It can suppress wages when companies treat review as routine monitoring work.

Public policy will shape the balance. Training support, wage transparency, collective bargaining rules, and worker access to productivity data can affect who captures the gains.

Measurement also needs to improve. The Labor Department is pursuing data-sharing arrangements with technology companies to track AI’s employment effects more quickly.

Better data should distinguish technical exposure from active workplace adoption. It should also separate direct salary cuts from smaller raises, weak starting offers, and lost promotion opportunities.

For workers, the practical question is not whether AI can perform an entire job. It is whether employers now view the job’s valuable tasks as easier to source elsewhere.

That shift can appear during a compensation review long before it appears in an unemployment report.

Track your own evidence accordingly. Record where AI saved time, where your judgment prevented an error, and where your knowledge changed the final decision.

Compare that record with new job postings and salary ranges in your field. If output expectations rise while compensation stalls, the wage effect is already reaching you.

The next Google News headline may still focus on layoffs because they offer a visible count. The quieter measure is whether employed people retain a fair share of what AI helps them produce.