In 2016, Geoffrey Hinton, a future Nobel laureate, said it was "completely obvious that within five years deep learning is going to do better than radiologists." He suggested hospitals stop training them.
Ten years on, here is what happened to radiology:
Average American radiologist pay: around USD 520,000 a year, roughly 48 percent more than in 2015
Residency positions in 2025: 1,208, up 4 percent
Vacancy rates: at record highs
Radiology AI models cleared by the US Food and Drug Administration: more than 700
Organisations that deployed them and reported a high degree of success in 2025: 19 percent
That gap between the confident prediction and the measured outcome is worth holding onto. It does not mean AI has no effect on work. It means the effect is landing in different places, and on a different timeline, than almost everyone forecast.
Four Different Things, All Reported as One Number
Most of the confusion in this debate comes from four separate measures being treated as interchangeable:
Exposure: the overlap between what AI can do and the tasks inside a job. A ceiling, not a forecast
Technical automation potential: the share of hours a technology could in principle perform. Not jobs, and not adoption
Observed usage: what people actually do with AI today
Realised employment change: head counts actually moving. There is far less of this
Almost every alarming headline takes a figure from the first two and reports it as if it belonged to the fourth.
The Exposure Numbers, and the Half Nobody Quotes
The International Monetary Fund, January 2024:
About 40 percent of global employment is exposed to AI, rising to 60 percent in advanced economies
Inside that 60 percent: 27 percent is highly complementary, where AI raises productivity
The other 33 percent has low complementarity, where displacement risk actually sits
So the IMF's displacement-risk figure for rich countries is 33 percent, not 60. An ILO index from May 2025 shows the same pattern:
One in four workers worldwide has some degree of generative AI exposure
Only 3.3 percent of global employment sits in the highest exposure band
Women are more exposed than men, 4.7 percent against 2.4 percent, because they dominate clerical work
The ILO states plainly that its estimates "reflect exposure to GenAI and not the actual impact on employment" and that "most jobs will be transformed rather than made redundant."
Two more famous figures with the same problem. Goldman Sachs said 300 million full-time jobs could face automation exposure, then added in the same write-up that most jobs "are more likely to be complemented rather than substituted by AI." McKinsey said 57 percent of US work hours are technically automatable, then put realistic 2030 adoption at 27 percent. Only the first number ever travels.
What US Labour Statistics Project to 2035
The US Bureau of Labor Statistics released its 2025 to 2035 projections on 27 August 2026, with AI built into the modelling. Total employment rises from 170.3 million to 176.2 million.
Falling fastest:
Word processors and typists: down 34.4 percent
Data entry keyers: down 25.5 percent
Telemarketers: down 21.4 percent
Office and administrative support overall: down 4.0 percent, shedding 752,100 jobs
Growing fastest:
Data scientists: up 34.6 percent
Information security analysts: up 28.5 percent
Software developers: up 10 percent, adding 267,700 jobs, the largest absolute gain in the economy
Electricians: up 9 percent, which BLS attributes to AI and data centre electricity demand
Notice the scale. The steepest decline in the whole projection removes about 14,000 typist jobs from a workforce of 176 million, and those occupations were already shrinking.
The bigger white-collar occupations barely move. Customer service representatives are down 5 percent, and BLS blames self-service systems without naming AI at all. Secretaries and administrative assistants, at 3.5 million, are down just 2 percent, and here BLS does name the cause: AI systems "enable staff in many organizations to prepare their own documents without the help of secretaries."
The Klarna Story, in Order
The most cited corporate example in this debate is also the most misreported.
February 2024: Klarna says its AI assistant handled 2.3 million conversations in a month, two thirds of customer service chats, "doing the equivalent work of 700 full-time agents." It says equivalent work of. It does not say 700 people were dismissed
May 2025: chief executive Sebastian Siemiatkowski is reported saying AI helped shrink the workforce by around 40 percent, partly through attrition and a hiring freeze. He also concedes that when cost becomes "a too predominant evaluation factor, what you end up having is lower quality"
November 2025: Klarna tells its earnings call the AI agent now does the work of 853 full-time representatives, up from 700, while customer service and operations costs rise to USD 50 million from USD 42 million
Klarna did not reverse its AI deployment. It reversed the human-elimination part, and costs went up. Compare IBM, where chief executive Arvind Krishna confirmed several hundred HR staff were replaced by AI agents, then added: "Our total employment has actually gone up, because what [AI] does is it gives you more investment to put into other areas."
Software Developers: Two True Numbers That Contradict
BLS: software developers up 10 percent to 2035, with AI named as the reason for the growth
Stanford Digital Economy Lab: workers aged 22 to 25 in the most AI-exposed occupations sit 19 percent below same-age workers in less exposed ones, up from 15 percent a year earlier
Both are right. The occupation grows while the entry gate narrows. Stanford is careful about the claim: it reports "no widespread, economy-wide job displacement associated with AI," and notes the adjustment works through reduced hiring rather than separations.
A mechanism explains it. Three field experiments with more than 5,000 developers found GitHub Copilot raised task completion by about 26 percent, with juniors gaining 21 to 40 percent and seniors only 7 to 16 percent. AI raises junior productivity most, which is precisely the argument for hiring fewer juniors.
One caution sits alongside it. In July 2025 METR ran a trial with 16 experienced open-source developers who were 19 percent slower with AI tools, had predicted a 24 percent speedup, and afterwards still believed AI had sped them up by 20 percent.
Creative Work: Real Damage, Smaller Than Advertised
Researchers at Washington University and NYU studied Upwork after ChatGPT and the image models launched:
Writing freelancers: monthly jobs down 2 percent, earnings down 5.2 percent
Image and design freelancers: jobs down 3.7 percent, earnings down 9.4 percent
Higher-rated, higher-priced freelancers took the larger hit, because the tool raises the floor rather than the ceiling
BLS projects writers and authors at zero percent growth, stating that AI writing "is projected to dampen demand," and graphic designers down 2 percent for the same reason. Journalists are the exception that proves the point: reporters are down 6 percent, attributed entirely to advertising collapse and consolidation, with no mention of AI.
The Jobs Everyone Predicted, Where Nothing Happened
Market research analysts: near the top of every exposure index, projected up 7 percent, with no AI mention
Financial analysts: also up 7 percent
Truck drivers: up 4 percent to 2.22 million, with no reference to autonomous vehicles anywhere in the outlook
On trucking, the reality check is stark. Aurora Innovation, the leading driverless trucking company, reported roughly 25 driverless trucks in operation in mid-2026, against USD 2 million in quarterly revenue and a USD 266 million operating loss.
What AI Is Actually Creating
PwC's Global AI Jobs Barometer, June 2026, built on more than one billion job advertisements across 27 countries:
AI skills wage premium: 62 percent, up from 57 percent
Headcount at the most AI-exposed companies grew 52.2 percent from 2018 to 2025, against 26.7 percent at the least exposed
In the UAE, the AI premium reached 92 percent in financial services
3.2 percent of UAE job postings required AI skills in 2025, up from 1.0 percent in 2021
The entry-level finding is the one to sit with. Among AI-exposed entry-level roles, those that had absorbed ten or more senior skills grew 35 percent since 2019, while other AI-exposed entry roles fell 10 percent. Entry-level jobs may not be disappearing so much as ceasing to be entry-level.
LinkedIn's 2026 fastest-growing US roles list was topped by AI Engineer, followed by AI consultants, AI researchers, data annotators and data centre technicians. Half the list is about building AI. The other half is about building the buildings that house it.
And a cautionary tale. Prompt engineering, the emblematic new AI job of 2023, was described by an Indeed Hiring Lab economist in 2025 as having "never been a large enough title to track mathematically" in job postings.
What Actually Protects a Job
Microsoft Research analysed roughly 200,000 anonymised Copilot conversations against occupational task data. Three clusters showed minimal AI applicability:
Physical activity: handling and moving objects
Monitoring: watching processes, resources and equipment
Guidance: controlling machines and directing people
It also found only a weak correlation between AI applicability and either wages or education, which undercuts the idea that AI is coming for graduates and sparing everyone else.
Regulated accountability is the strongest protection on record. Autonomous diagnostic models face higher approval requirements, and malpractice insurers resist covering unsupervised AI, requiring physician sign-off. Two rows in the BLS data make the point: secretaries are down 2 percent, medical secretaries up 5 percent. Same skill set, but one sits inside a regulated, human-facing setting.
The Case That Very Little Has Happened Yet
Denmark: Anders Humlum and Emilie Vestergaard linked adoption surveys to administrative labour records covering 25,000 workers across 7,000 workplaces in 11 exposed occupations. They found precise null effects on earnings and hours through mid-2024, with early-career estimates also precise zeros
Yale: the Budget Lab found the occupational mix shifting only about one percentage point faster than during early internet adoption
OECD, July 2026: "so far, the role of recent advances in large language models in explaining the difficulties facing young people appears to be limited"
Daron Acemoglu, MIT: puts the productivity gain at no more than 0.66 percent over ten years, with only 4.6 percent of US tasks profitably automatable in a decade
One skeptical statistic deserves correcting. The MIT NANDA report behind the "95 percent" figure did not find that 95 percent of AI pilots fail. It found 95 percent of organisations getting zero measurable return, in a version 0.1 working paper based on 153 leaders surveyed at conferences.
It also contains a displacement finding nobody quotes: AI impact is "manifesting through selective displacement of previously outsourced functions and constrained hiring patterns, but not through broad-based layoffs."
The Last Automation Panic Was Wrong in a Useful Way
In 2013, Carl Frey and Michael Osborne predicted 47 percent of US jobs would be gone within twenty years. What happened instead:
The US economy added 16 million jobs between 2013 and 2022
Insurance underwriters, rated at highest automation risk, grew 16.4 percent
Across 21 OECD countries from 2012 to 2019, there was no net job destruction anywhere
But jobs at high automation risk grew 6 percent, against 18 percent for low-risk jobs
Automation did not destroy the exposed occupations. It slowed their growth by about two thirds. That is the realistic shape of what AI is likely to do.
Why the Arab World Faces a Different Problem
Direct AI exposure in Arab labour markets is comparatively low, because labour is concentrated in manual and interpersonal work. The risk arrives through a different door: offshored service work.
The Philippines: in July 2026 its IT and Business Process Association cut its 2028 employment target from 2.5 million to between 1.85 million and 2.14 million, a downward revision of up to 650,000 jobs, citing AI
India: Nasscom put technology revenue up 6.1 percent for the year, on headcount growth of just 2.3 percent
Egypt: outsourcing exports reached around USD 5.2 billion in 2025, targeting USD 8 billion by 2028
Morocco: targeting 270,000 offshoring jobs by 2030, up from 148,500 in 2024
Neither Egypt nor Morocco publicly prices in AI risk. The MIT NANDA finding that displacement lands first on outsourced functions is the mechanism connecting those four bullet points.
A 2026 report from UN ESCWA and the ILO projects roughly 120,000 net new jobs across the Arab region between 2025 and 2035 after initial displacement. For a region this size, that is a strikingly small number. The structural context makes it bite:
Informal employment: 56 percent of the total in Arab countries in 2025
Workers with unemployment insurance: 7 percent
Youth unemployment in the Arab States: 26.2 percent, against a global rate of 12.4 percent
The UAE has published the most specific national figure. Its National Strategy for Artificial Intelligence 2031 states that around 43 percent of work activities in administration, government, manufacturing and construction have automation potential, and roughly 125,000 UAE nationals in those roles may be affected.
For Lebanon, no equivalent data exists. What is documented is that remote work for overseas employers has become a real channel, running through exactly the categories where AI exposure is highest.
Three Patterns That Hold Across the Evidence
The effect operates through hiring, not firing. Stanford measures reduced hiring rather than separations, MIT NANDA describes constrained hiring rather than layoffs, and the Philippine industry body revised a future target rather than announcing present cuts. That is why unemployment statistics show so little
Skill and pay are not the dividing line. What matters is whether AI automates the task or amplifies the person doing it. Secretaries down 2 percent, medical secretaries up 5 percent. Computer programmers down 7 percent, software developers up 10 percent
The biggest job creator is not in technology. In the same projections that show typists collapsing and data scientists booming, the largest single source of new American jobs is home health and personal care aides, at 760,500 openings a year, at a median wage of USD 35,800
Common Questions
Which jobs are most at risk from AI in 2026?
Clerical and administrative work. Word processors and typists down 34.4 percent, data entry keyers down 25.5 percent, telemarketers down 21.4 percent, with office support shedding 752,100 jobs by 2035.
Will AI replace software developers?
Not on current projections. BLS has developers up 10 percent and names AI as the reason. The measurable effect is at the entry level, where 22 to 25 year olds sit 19 percent below where they would otherwise be.
Which jobs are safest?
Physical work, monitoring, and guiding machines and people. Regulated professions requiring human sign-off are also protected. Home health aides are up 18 percent, electricians up 9 percent, nurses up 6 percent.
How does AI affect jobs in the Middle East?
Mostly indirectly, through offshored services. The Philippine outsourcing industry has cut its 2028 employment target by up to 650,000 jobs citing AI, while Egypt and Morocco are scaling up into the same market.
