Every few months, a new study comes out saying AI will create more jobs than it destroys. The headlines get written. The think pieces follow. LinkedIn fills up with posts from people in the AI industry sharing the good news. And then, if you look at who funded the study, you find the same names every time.
Microsoft. Google. Amazon. The venture capital firms with eight-figure positions in AI startups. The consultancies whose entire growth strategy depends on AI adoption. The industry associations whose membership dues come from the companies building this technology.
This is not a conspiracy. It is incentive structure. And incentive structure is one of the most reliable predictors of research outcomes that social scientists have ever identified.
The Studies That Get Funded and the Studies That Do Not
In 2023, economists Daron Acemoglu and Simon Johnson published research suggesting that automation and AI would primarily benefit capital owners rather than workers, and that the net employment effects were likely to be negative in most sectors over a 10-year horizon. Acemoglu won the Nobel Prize in Economics the following year, partly for his work on how institutions and power shape economic outcomes.
That study did not get major tech company funding. It also did not get the same level of media amplification as the McKinsey report released the same year suggesting AI could add $4.4 trillion in annual value to the global economy. McKinsey’s clients include most of the major AI companies. The $4.4 trillion figure was quoted in thousands of articles. The Acemoglu research was quoted in far fewer.
This is how narrative is built. Not through lies, but through selective amplification of research that supports preferred conclusions.
The Methodology Problem
Most optimistic AI-and-jobs studies share a structural problem: they count hypothetical future jobs that AI might enable while underweighting current jobs that AI is visibly eliminating right now. They classify “prompt engineer” and “AI trainer” as new job categories without accounting for the fact that these roles employ a small fraction of the people being displaced from customer service, content creation, coding, and administrative work.
The historical precedent they reach for most often is the Industrial Revolution, which did eventually create more jobs than it destroyed. What they consistently underemphasise is that “eventually” meant roughly seventy years of disruption, poverty, and social upheaval that required massive state intervention to manage. The steam engine creating net positive employment by 1900 was cold comfort to the handloom weavers whose livelihoods were destroyed in 1830.
What We Actually Know
What we actually know is that AI is currently reducing headcount in software development teams, customer service operations, content production, legal research, financial analysis, and administrative functions. We know this because companies are reporting it in earnings calls, and because hiring data in these sectors is showing measurable declines. We know it because the CFO survey data, which does not get funded by AI companies, shows executives privately citing AI as a primary reason for reduced hiring plans.
What we do not know, because it has not happened yet, is what the compensating job creation looks like, what it pays, where it is located geographically, what skills it requires, and how long the transition takes. The optimistic studies treat these open questions as answered. They are not.
Why This Matters Beyond the Numbers
Policy follows research. If the research consensus says AI creates jobs, policymakers have less urgency to build retraining programs, strengthen unemployment systems, update labour laws, or regulate the pace of automation. If the research consensus says the effects are uncertain and potentially severe, the policy response looks different.
The companies funding optimistic research are not just managing their public image. They are shaping the policy environment in which they operate. A government that believes AI creates jobs is a government less likely to tax automation, mandate transition support, or slow the pace of deployment.
Read the studies. Then read the funding disclosures. Both pieces of information matter.