AI Washing: How Companies Are Using “AI Efficiencies” to Disguise Offshoring and Displace American Workers
A decomposition model showing why the labor market data cannot tell the two apart, and why that is precisely the point.
1. The Tell
In the first half of 2025, Amazon posted 22,700 job openings in Seattle. By the first quarter of 2026, that number was 4,540.
The accompanying corporate narrative was about artificial intelligence. Efficiency, streamlining, the reallocation of capital from headcount toward infrastructure. It was a plausible story. Amazon cut roughly 16,000 corporate roles in the first quarter of 2026 while reporting its fastest AWS growth in thirteen quarters, and the company has committed to an enormous AI capital expenditure program alongside its peers.
But there is a detail that does not appear in the efficiency narrative. In that same quarter, Milan and Pisa emerged as Amazon’s second and third largest hiring cities.
Not data centers. Hiring cities. Locations with fully loaded labor costs a fraction of Seattle’s, in a country that features nowhere in the AI story being told to investors.
This is the pattern that sits underneath a meaningful share of what is currently reported as AI driven job loss. It is not, in these cases, software replacing a worker. It is a worker in a high cost market being replaced by a worker in a low cost market, with AI named as the mechanism because that framing carries a better narrative and a better market signal than the alternative, which is ordinary labor arbitrage of the kind that has been running since the 1990s.
There is now a term for this. The Society for Human Resource Management calls it AI washing: overstating the role of artificial intelligence in a business decision.
This article makes three claims. First, that AI washing is happening at scale. Second, and this is the part that has not been argued carefully, that the labor market data we have cannot distinguish an automated job from an offshored one, which is exactly why the label is safe to use. Third, that once you build a model separating the two mechanisms, you find something more uncomfortable than either side of the debate expects.
2. The Scale of the Label
The label has spread remarkably fast.
MetricValueSource2026 layoff events explicitly citing AI or automation54 to 56 percentSkillSyncer trackerU.S. workers in AI cited job cuts, 2026 year to dateAbout 205,000ResumePulse, August 2026Cumulative AI cited cuts since tracking began in 2023About 150,000 announcedChallenger, Gray and ChristmasShare of that cumulative total occurring in the first half of 2026About two thirdsChallenger, Gray and ChristmasConsecutive months in 2026 with AI as the single most cited reasonFourChallenger, Gray and ChristmasU.S. tech jobs cut since January 2026About 140,000Financial Times analysis
Challenger, Gray and Christmas reported that AI became the leading stated reason for corporate job cuts in March and April 2026, the first time a technology factor has topped its monthly rationale rankings in the firm’s history.
A reason that goes from marginal to dominant in eighteen months is either a genuine technological discontinuity or a change in what companies have decided to say. Distinguishing between those two possibilities is the entire problem.
3. What the Label Covers Up
Here is the first piece of evidence that something is being relabeled rather than newly caused.
If AI were the cause of the decline in a given occupation, you would expect the decline to begin around the time the capability arrived. Instead, posting level data shows the declines running smoothly through periods when nobody was citing AI at all.
Customer service representative postings fell 24.9 percent between the first half of 2024 and the second half of 2025, before a single press release named AI as the reason. Data entry clerk postings declined in a straight line across six consecutive reporting periods.
These are not cliff edges. They are linear, multi year structural trends that acquired a new and convenient label in 2025 and 2026. The shape of the curve did not change when the explanation did.
That is a signature of relabeling, not of causation.
4. Why This Particular Label
The obvious question is why a company would prefer “AI replaced them” to “we moved the work.”
The answer is that the two statements have very different audiences and very different costs.
Offshoring is a political liability. It invites WARN Act scrutiny, congressional attention, union response, and press coverage with an identifiable villain and an identifiable victim. It carries thirty years of accumulated negative valence in American politics.
Automation is a technological inevitability. It frames the same headcount reduction as evidence of forward capability rather than an admission of cost pressure, overhiring, or soft demand. It converts a decision into a condition. Nobody organizes a protest against a phase transition.
There is also a straightforward accounting dimension. As one industry practitioner put it bluntly on an anonymous forum, every domestic role eliminated funds several offshore roles, and routing those through contracting firms takes the cost off the employee line entirely.
The market, it turns out, is not entirely credulous. Financial Times analysis found that companies citing AI as a factor in job cuts subsequently underperformed the Nasdaq by nearly 10 percent over the following 30 trading days.
This is the strange part. The label is chosen for investor signaling, and the signal does not work. Which suggests the audience that actually matters is not investors at all. It is regulators, journalists, and the political process.
5. The Measurement Problem, Including in My Own Model
I need to turn this on my own prior work.
In an earlier piece I modeled labor displacement using a hazard function in which the displacement rate rises exponentially with AI capability, calibrated against BLS data from 2015 to 2025. That calibration produced a baseline displacement rate of 1.2 percent annually and an amplification factor of 0.41. I then ran it forward and reported a high probability of demand collapse.
That calibration has a defect I did not flag at the time. It was fit to an aggregate displacement series that cannot distinguish an automated job from an offshored one. Both appear in the data as domestic headcount going to zero in an exposed occupation. Both show up in BLS separations. Both look identical in the labor share series.
So 1.2 percent was never an automation rate. It was a total domestic displacement rate being read as an automation rate. Every forecast built on it inherits that ambiguity.
The Capital Capture Coefficient I introduced in the same piece has the same problem, and in a way it is worse. The coefficient measures how much income shifts from labor to capital, and it is entirely agnostic about whether capital captured that income by buying a model or by buying cheaper labor. That agnosticism was presented as generality. It was actually a blind spot.
The rest of this article is an attempt to fix it.
6. The Decomposed Displacement Model
6.1 Structure
Split the observed displacement rate into two components that are equal at the point of observation but governed by different dynamics.
d_total(t) = d_ai(t) + d_off(t)
Let psi = the share of observed displacement that is
actually labor arbitrage rather than automation.
At t = 0:
d_ai(0) = (1 - psi) * d_total(0)
d_off(0) = psi * d_total(0)The automation channel is a hazard that rises with capability and saturates only at the technical automatability ceiling.
d_ai(t) = min( h_ai * exp(k * (C(t) - 1)), cap )
* (1 - S_ai(t) / ceiling_ai)
where C(t) = AI capability index
k = capability sensitivity (0.41 in the prior model)
S_ai = cumulative share already displaced by automationThe arbitrage channel is a hazard proportional to the surviving wage gap, which closes as offshore compensation converges toward domestic.
A(t) = A0 * exp(-g * t)
d_off(t) = h_off * (A(t) / A0)
* (1 - S_off(t) / ceiling_off)
where A(t) = surviving arbitrage gap at time t
g = net convergence rate of offshore toward domestic cost
ceiling_off = offshorable (tradable task) share of employmentThis is the crux of the whole model. Automation displacement compounds. Arbitrage displacement extinguishes itself. The gap that makes offshoring profitable is the same gap that offshoring closes.
6.2 The assumption I had to remove
My earlier model let capability grow exponentially without bound. Over a 24 year horizon that forces the automation channel into saturation regardless of every other parameter, which means it pre decides the answer. I replaced it with a logistic path toward a ceiling drawn separately for each run across a wide range, so the simulation spans both the world where capability keeps compounding and the world where it plateaus, rather than assuming one of them.
6.3 Employment, wages, demand
L(t+1) = L(t) * (1 - d_ai(t) - d_off(t))
+ rho * (1 - L(t))
w(t) = w0 * (1 + g_w)^t * L(t)^eta
W(t) = w(t) * L(t)
C(t) = mpc_wage * W(t) + mpc_capital * K(t)
where rho = annual reabsorption rate of the displaced pool
eta = wage scarring elasticity from labor market slackPoisson distributed policy shocks, meaning tariffs, visa rules and disclosure requirements, perturb the arbitrage channel only, because that is the channel policy can actually reach.
The model runs 10,000 Monte Carlo simulations from 2026 to 2050.
7. Result One: The Two Stories Are Indistinguishable
INSERT IMAGE 1 HERE: chart1_unidentifiable.png
Two incompatible stories, one indistinguishable data series. Median paths with 10th to 90th percentile bands, 10,000 runs.
Run the model twice on identical parameter draws. In the first, take corporate attribution at face value, so the arbitrage share is zero and all displacement is genuine automation. In the second, decompose it, with the arbitrage share centered on the roughly half figure Bloomberg’s analysis suggests.
At 2030, the two readings produce median domestic labor income of 0.9631 and 0.9673 respectively.
A difference of 0.4 percentage points, with 10th to 90th percentile intervals of 0.926 to 0.996 and 0.931 to 1.000. Almost entirely overlapping.
These are two completely incompatible descriptions of the American economy. One says a technological transition is irreversibly dissolving domestic employment. The other says firms are executing a decades old cost strategy under a new name. Through 2030, they generate the same data.
This is the finding that matters most, and it is not a forecast. It is an identification result. It means the AI attribution is, for practical purposes, unfalsifiable on the timescale over which the political and regulatory response has to be decided.
A claim that cannot be checked is a claim that can be made freely. That is not incidental to why it is being made. It is the reason it is safe.
8. Result Two: The Mechanisms Diverge, Eventually
INSERT IMAGE 2 HERE: chart2_hazards.png
Same starting rate, opposite trajectories. Median annual displacement hazard by channel.
Calibrated to the same rate in 2026, about 0.57 percent annually in each channel, the two hazards separate immediately and never reconverge. By 2050 the automation hazard has roughly doubled to 1.10 percent while the arbitrage hazard has decayed to 0.14 percent.
Over the full horizon this produces:
MetricAttribution readingDecomposed readingMedian labor income, 20320.9530.961Median labor income, 20400.9140.943Median labor income, 20500.9020.925Probability labor income falls below 80 percent of 2026, by 204010.2 percent5.4 percentMedian trough level0.8850.918Median trough year20462047
The decomposed reading is meaningfully less severe, with roughly half the probability of serious impairment by 2040. Splitting by capability regime sharpens it:
Capability regimeDivergence in 2050 attributable to the splitPlateau (low capability ceiling)plus 0.031Moderateplus 0.034Compounding (high capability ceiling)minus 0.001
So the decomposition matters only if AI capability plateaus. If capability keeps compounding, the automation channel eventually swamps everything and it makes no difference whatsoever whether the 2026 layoffs were honestly labeled.
That is a genuinely awkward result for the thesis I set out to argue, and I am reporting it rather than burying it.
9. Result Three: The Argument Is About the Wrong Parameter
INSERT IMAGE 3 HERE: chart3_reabsorption.png
The parameter nobody argues about moves the outcome five times as much.
Here is the finding I did not expect.
Vary the reabsorption rate, meaning how quickly displaced workers find replacement domestic roles, and hold everything else fixed:
Reabsorption rateMedian labor income, 2040Median labor income, 20501.0 percent0.9120.8532.0 percent0.9260.8883.5 percent0.9410.9275.0 percent0.9590.9658.0 percent0.9811.012
Reabsorption moves the 2050 outcome by nearly 16 percentage points. The entire argument about AI versus offshoring moves it by about 3.3.
The parameter that dominates the outcome is the one nobody is arguing about.
Whether an American worker is displaced by a language model or by a colleague in Milan turns out to matter roughly five times less than whether the domestic economy generates a replacement role for them. And the empirical basis for optimism there is thin. The trade shock literature, particularly Autor, Dorn and Hanson on the China shock, found local labor markets that had not recovered a decade after exposure, implying reabsorption rates at the low, punishing end of that table.
10. What Follows
The three results point somewhere other than where the AI washing argument usually lands.
The disclosure case is strong, and it does not depend on the model. If AI attribution is unfalsifiable on a five year horizon, then it functions as a free option. A firm can claim it, receive whatever narrative benefit accrues, and face no risk of contradiction. The remedy is not to litigate causation, which cannot be done, but to make the facts observable. Requiring large employers to report headcount by country and by function, at the same cadence as financial results, would resolve within one reporting cycle what no amount of modeling can resolve from aggregate data. The Amazon pattern was visible only because somebody assembled posting level data by city. That should not require detective work.
The policy target is different from what either camp proposes. If displacement is automation, the standard answer is redistribution through capital taxation, transfers, some version of universal income. If it is arbitrage, the standard answer is trade and tax policy. But the model says the binding constraint is reabsorption, and reabsorption responds to neither. It responds to job creation, active labor market policy, portable benefits, wage insurance, and the unglamorous machinery of getting a 47 year old claims processor into a different job at comparable pay. Denmark spends on this. The United States spends roughly 0.1 percent of GDP on active labor market programs, near the bottom of the OECD. That number, not the AI narrative, is the one that determines the 2050 outcome.
And a caution for people who agree with me. The finding that offshoring is masquerading as automation is not the reassurance it initially appears to be. Arbitrage driven displacement self limits as wages converge. But converge here means converge downward, toward the domestic worker, not upward toward parity. The mechanism that shuts off the offshoring channel is the erosion of the American wage premium. That is not a recovery. It is an equilibrium nobody voted for.
11. Limits
Stated plainly, because the argument in section five loses its force if I do not apply it to myself.
The arbitrage share carries most of the weight and is the weakest input. The estimate that roughly half of AI attributed layoffs reappear offshore or at lower pay rests on a single Bloomberg analysis. Nothing in my model validates it independently.
This is a model, not evidence. It demonstrates that two hypotheses are observationally equivalent given the available data. It does not establish which is true. No forecast here should be read as a prediction.
The exposed occupation index is an abstraction. Real displacement is concentrated in specific functions such as support, back office, entry level engineering and data operations, with very different offshorability and automatability profiles. A one sector model averages over that.
Reabsorption is the load bearing parameter and the least constrained. I calibrated it from the trade shock literature, which may not generalize to a services sector transition.
The AI displacement is real too. Klarna cut roughly 700 customer service roles and partially reversed after service quality fell, which is evidence that some substitution is genuine and some of it is worse than firms expected. Nothing here should be read as claiming the automation channel is fictional. The claim is narrower. We cannot currently measure its size, and that measurement failure is being exploited.
12. What Would Settle It
The decisive evidence is not econometric. It is administrative, and it already exists inside these companies.
First, country level headcount disclosure, quarterly, by function.
Second, WARN Act filings cross referenced against subsequent foreign entity registrations and capability center expansions. India alone hosts roughly 1,700 global capability centers, about 53 percent of the world total, with Nasscom and Zinnov projecting 2.5 to 2.8 million employees by 2030.
Third, contractor and outsourcing spend broken out from employee cost, since the accounting shift from payroll to vendor line is part of the mechanism.
Fourth, task level productivity audits at firms claiming AI substitution, which is the thing Klarna’s reversal actually demonstrated.
Any one of these would do more to resolve the question than another decade of modeling. Their absence is not an oversight.
Sources
- JobsPikr, AI Layoffs 2026: The ROI Reality Check, March 2026. Posting level city data and pre trend analysis.
- Financial Times analysis, via TechCrunch, running list of major tech layoffs citing AI, July 2026. Cut totals and post announcement equity performance.
- Challenger, Gray and Christmas. Monthly layoff rationale tracking, 2023 to 2026.
- Bloomberg. Analysis of offshore and lower salary rehiring in AI attributed layoffs.
- Society for Human Resource Management. AI washing definition.
- ManageEngine Insights, What’s actually driving 2026’s AI layoffs, July 2026.
- Nasscom and Zinnov. India global capability center market projections to 2030.
- SkillSyncer and ResumePulse. 2026 layoff trackers.
- Autor, Dorn and Hanson. China shock local labor market adjustment.
- BLS Productivity and Costs Program. BEA National Income Accounts.
Model code and simulation output available on request. All figures generated from a 10,000 run Monte Carlo.
About the Author
Brian Curry is a Kansas City–based Machine Learning and AI researcher, data scientist, and the founder of Vector1 Research, where he works on the intersection of cognitive AI architecture, causal inference, and economic systems. His research introduced Memory-Node Encapsulation (MNE), an original data structure for artificial episodic memory, and the NeoCortex-M architecture for memory-driven AI. He is the creator of Papilon, an open-source framework for complex system optimization with agentic orchestration, PyCausalSim, a Python framework for causal discovery through simulation, and Daedalus, a platform for causal and economic analysis.
His practitioner work focuses on building production intelligence systems — causal marketing analytics, agentic systems, and the serving infrastructure that makes them deployable inside enterprise environments. He writes about that work at vector1.ai and on Medium.
