Will AI Make Inequality Worse? Facts, Myths and Practical Responses

Updated September 16, 2026. Every figure below was read that day from the primary sources listed at the end: IMF and OECD publications, US Census Bureau survey releases, Bureau of Labor Statistics wage and productivity data, and the original experiment papers. Wages are US national medians. Nothing here is a forecast of your own job.

"AI will widen inequality" and "AI will narrow it" are both being argued from the same small pile of studies. The reason they can both be argued is that almost nobody converts the headline percentages into anything you can check. A 14 percent productivity gain sounds decisive until you work out what 14 percent is in hours, what those hours cost, and who ends up holding them. So this piece does the conversion first and the argument second.

One occupation, costed out

Take customer service representatives, occupation code 43-4051. The Bureau of Labor Statistics puts the May 2025 median wage at $21.53 an hour, or $44,770 a year, across roughly 2,666,000 US workers. That is a large, specific, well-measured population, and it happens to be the exact setting of the best field experiment we have.

Brynjolfsson, Li and Raymond studied a real contact centre and found that access to a generative AI assistant raised issues resolved per hour by about 14 percent. Convert that. If output per hour rises 14 percent and the volume of customer issues stays flat, the labour hours needed fall by 14/114, which is 12.3 percent. Applied to a 2,080-hour work year that is 255 hours per worker per year. At $21.53 an hour, that is $5,499 of wage cost per worker per year, and across the occupation, about $14.7 billion a year.

Step	Value	Where it comes from
Measured productivity gain	+14% issues per hour	NBER field experiment
Hours removed per worker-year	255 hours	My calculation, 12.3% of 2,080
Wage cost of those hours	$5,499 per worker per year	BLS May 2025 median hourly wage
Across the occupation	about $14.7 billion a year	BLS 2025 employment base
BLS projection for the occupation	minus 5% by 2035, about 141,800 jobs	BLS Employment Projections

Now the distributional part, which is the finding people skip. In that same experiment the gain was not evenly spread and it did not favour the strongest workers. Novice and low-tenure agents improved by about 34 percent; experienced, high-performing agents saw close to no gain at all. Two-month agents with the tool performed on a par with six-month agents without it. Within that occupation, the technology compressed the skill gap rather than widening it.

That single result should make anyone cautious about the confident version of either argument. The same intervention that removes $14.7 billion of paid hours from an occupation projected to shrink 5 percent also, inside the occupation, narrows the gap between the best and worst paid performers. Both effects are real and they point in opposite directions.

What the four experiments actually measured

The studies quoted in this debate are not measuring the same thing, which is why their numbers look so different.

Study	Task	Result	Note
Noy and Zhang, Science 2023	Short professional writing	17 minutes vs 27 minutes, about 37% faster; grades up 0.45 SD (4.54 vs 3.79 on a 1 to 7 scale)	Faster and better, no speed-quality trade
Brynjolfsson, Li and Raymond, NBER 2023	Live customer support	+14% overall; +34% for novices; near zero for experts	Real workplace, not a lab
Peng and colleagues, arXiv 2023	Write an HTTP server in JavaScript	55.8% faster with Copilot	A single, well-specified, self-contained task
METR, 2025	Real issues in large open-source repositories	19% SLOWER with AI tools	Experienced developers, familiar codebases

Look at rows three and four together. Both involve professional software developers. One reports 55.8 percent faster, the other 19 percent slower. The difference is not that one study is wrong. It is that the first measured a clean greenfield task with a known shape and the second measured maintenance work inside a codebase the developer already knew intimately. The tools are excellent at the first and, on that evidence, actively costly at the second.

The same conversion, run on the task where the tools lose

Do the arithmetic in the other direction, because nobody ever does. BLS puts the May 2025 median wage for software developers at $135,980 a year, which is about $65 an hour. On the METR finding, a four-hour maintenance task takes 19 percent longer, so it takes about 46 minutes more, which is roughly $50 of developer time burned per task. Ten such tasks a week across a team of eight is on the order of $200,000 a year, spent to be slower.

The detail from that study that should worry any manager rolling out a tool: after finishing, the developers still believed AI had sped them up by about 20 percent. The measurement said minus 19; the felt experience said plus 20. A 39-point gap between perception and outcome is not the sort of thing that self-corrects through user feedback.

The consultant experiment by Dell'Acqua and colleagues found the same shape from the other side. On tasks inside what they called the jagged technological frontier, consultants with AI completed 12.2 percent more tasks, 25.1 percent faster, at over 40 percent higher rated quality. On a task deliberately chosen to sit outside that frontier, consultants using AI were 19 percentage points less likely to reach a correct answer than consultants without it. Same people, same tool, opposite sign, depending entirely on task selection.

Why 14 percent in a contact centre is 1.4 percent in the national accounts

Here is the gap that should anchor the whole debate. BLS reported that nonfarm business labour productivity rose 1.4 percent at an annual rate in the second quarter of 2026, and 2.2 percent against the same quarter of 2025. That is the economy-wide measurement, taken in the same period that task-level experiments were reporting gains of 14 to 56 percent.

Both are correct. Task-level gains do not become economy-level gains unless the freed hours are redeployed into something that produces measurable output, and that redeployment is slow, uneven, and often does not happen at all. A support agent who resolves 14 percent more tickets in a queue that is not 14 percent longer has produced no additional output; the firm has produced a staffing decision, not growth. This is the ordinary shape of technology diffusion, not evidence that the experiments were faked, but it does mean that anyone quoting a 55.8 percent figure as a description of what AI is doing to the economy is quoting a laboratory task as if it were a national statistic.

Exposure and automation risk are two different measurements

The two headline figures in this debate get treated as rival estimates of the same quantity. They are not.

The IMF's January 2024 staff discussion note found about 40 percent of workers worldwide, and about 60 percent in advanced economies, in occupations with high AI exposure. Crucially, it splits that 60 percent: roughly 27 percentage points are high exposure with high complementarity, where the technology is likely to augment the worker, and roughly 33 percentage points are high exposure with low complementarity, where substitution is the likelier outcome. So the IMF number is not "60 percent of jobs are at risk." It is closer to "60 percent of jobs are touched, and slightly more than half of those in a way that is not obviously good for the worker."

The OECD's 2023 Employment Outlook put 27 percent of jobs in occupations at the highest risk of automation. That is a different question with a different method, and it is a mistake to read 27 against 60 as though one of them is losing an argument. More interesting is what the OECD found later: a November 2024 paper by Georgieff, using 2014 to 2018 data, concluded that AI had not so far altered the gap between high-wage and low-wage occupations, and that higher AI exposure may be associated with lower inequality between high- and low-wage workers within occupations. Compression, not widening, in the actual wage data available at the time.

Two caveats on that, stated plainly. The data window ends in 2018, which predates generative AI entirely, so it describes an earlier generation of the technology. And the paper reports a direction rather than an effect size I can quote. It is the best wage-level evidence available from a primary source and it is also thin. Anyone who tells you the wage data settles this question has not read how narrow the wage data is.

Who keeps the saved hour

This is where the inequality question actually lives, and the honest answer is that the measurement is incomplete. The labour share of income in the US nonfarm business sector stood at an index value of 93.4 in the second quarter of 2026, against a 2017 base of 100, meaning labour's share of what the sector produces remains below where it was nine years ago. That is a pre-existing trend with many causes, and attributing it to AI would be exactly the sort of unsupported claim this article is trying to avoid. What it does establish is the baseline condition: the freed hour enters an economy that has, for a decade, been routing a shrinking share of output to wages.

The occupational projections give a sharper picture than the aggregate does. Set them side by side.

Occupation	Median annual wage, May 2025	Employment	Projected change to 2035
Customer service representatives	$44,770	about 2,666,000	minus 5%, about 141,800 jobs
Paralegals and legal assistants	$62,890	about 404,900	0%, about 1,100 fewer
Writers and authors	$76,910	about 140,300	0%, about 500 fewer
Software developers	$135,980	about 1,700,000	plus 10%, about 174,700 jobs

The pattern is not subtle. The highest-paid of the four occupations is the one projected to grow, and the lowest-paid is the one projected to shrink by the largest absolute number. BLS's own commentary makes the same point, noting that generative-AI adoption is expected to fuel growth in computer and mathematical occupations, with data scientists projected up 33.5 percent, while dampening demand in administrative and clerical roles. That is the mechanism by which this technology could widen inequality, and note that it operates between occupations, which is precisely the dimension where the OECD wage study found no effect through 2018. The between-occupation story is a projection; the within-occupation compression is a measurement. Do not confuse their epistemic status.

Adoption is what sets the clock

None of the above happens on the timetable the coverage implies, because most firms are not using these tools yet. The Census Bureau's Business Trends and Outlook Survey put current AI use at roughly 19.8 percent of US businesses for the period ending May 3, 2026. In November 2023 the same survey put it at 3.8 percent. So adoption has gone up more than fivefold in two and a half years, and four out of five firms still are not doing it.

The size split matters more than the headline. About 37 percent of firms with 250 or more employees reported AI use, and about 32 percent of firms with 100 to 249, against under 20 percent for firms with four or fewer. If the productivity gain is real and it is concentrated in large firms, the near-term inequality effect may show up between firms before it shows up between workers, as scale advantages compound. That is a hypothesis, not a finding, and I am labelling it as one.

Three things I would actually do with this

First, if I managed a team, I would stop rolling tools out uniformly and start sorting tasks by whether they look like the HTTP-server task or the maintenance task. The evidence says the same tool produces plus 55.8 percent on one and minus 19 percent on the other. Task selection is the entire ball game and almost no deployment plan treats it that way.

Second, I would measure rather than survey. The METR finding that developers felt 20 percent faster while being 19 percent slower means that asking your team whether the tool helps will produce a confidently wrong answer. If you cannot measure cycle time, you do not know.

Third, if I were early in a career in one of the shrinking occupations, I would read the novice-gain result as the actionable one. A 34 percent improvement concentrated among low-tenure workers is, in the short run, the strongest individual case for learning these tools well: it is the one documented mechanism by which someone without tenure closes a gap quickly. It is also, in the long run, the mechanism that makes the tenure itself worth less. Both of those are true, and I would rather be on the near side of that trade than the far side of it.

What this article does not claim

It does not claim AI is causing the decline in labour share; that trend long predates it. It does not claim the BLS projections will be right; they are projections. It does not claim the four experiments generalise beyond their tasks, which is the central point of the METR and jagged-frontier results. And it does not claim the wage evidence is settled, because a paper using data through 2018 cannot settle a question about a technology that arrived in 2022.

Tip: If the practical takeaway you draw from this is to measure your own cycle time on a few real tasks, the setup that makes long comparison sessions bearable matters more than the software, so a laptop stand and a set of bluetooth earbuds are the two cheapest changes to a working day spent in that kind of comparison. (These are Amazon Associate links - we may earn a small commission on qualifying purchases.)

Continue with the Everyday AI Tools series

• Everyday AI tools guide, part 3: 15 practical prompts and safe automation ideas
• How to build reliable AI workflow prompts

This is general economic and technology reporting, not personalised career, legal or investment advice. Wage and employment figures are US national medians published by the Bureau of Labor Statistics and do not describe any individual job or region. No retailer prices or product ratings are quoted anywhere in this article.

Where these figures came from

• IMF, Gen-AI: Artificial Intelligence and the Future of Work, Staff Discussion Note, January 2024
• OECD, What impact has AI had on wage inequality? (Georgieff), November 2024
• OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market
• US Census Bureau, Business Trends and Outlook Survey AI use, May 2026
• Noy and Zhang, Experimental evidence on the productivity effects of generative artificial intelligence, Science, 2023
• Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161
• Peng and colleagues, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot
• METR, Measuring the impact of early-2025 AI on experienced open-source developer productivity
• Dell'Acqua and colleagues, Navigating the Jagged Technological Frontier, Organization Science
• BLS, Customer Service Representatives, Occupational Outlook Handbook
• BLS, Software Developers, Occupational Outlook Handbook
• BLS, Productivity and Costs, second quarter 2026 revised, released September 3, 2026
• FRED, Nonfarm Business Sector labour share for all workers
• BLS, Artificial intelligence, information technology, and employment