Every previous wave of automation followed the same general script: the least skilled, lowest-paid work got hit first, then mid-skill clerical and administrative roles, then -- eventually, slowly -- higher-skilled work at the edges. The political and policy response was built around that script too: help displaced workers retrain for better, higher-skill jobs, and trust that automation would eventually lift wages at the bottom even as it displaced people from specific tasks. A new analysis from Tufts University suggests AI may not follow that script. It may be rewriting it in reverse.
The Tufts AI Jobs Risk Index, developed by researchers at the Fletcher School in collaboration with Digital Planet, applied a task-decomposition methodology to 797 occupations using Bureau of Labor Statistics data and AI capability benchmarks. The result is a ranked exposure score for each occupation: what share of its tasks can be performed by current AI systems, and how quickly is that share growing. In a mid-range adoption scenario, the index identifies 9.3 million US jobs as high-risk, with projected income losses between $200 billion and $1.5 trillion. Those numbers are within the range of similar analyses. What is different is where those jobs sit.
Among the 33 occupations classified as "tipping point" roles -- where more than 50% of tasks are automatable under current AI -- the index places computer programmers, financial planners, web designers, and data analysts. These are not low-wage jobs. A software engineer earning $120,000 a year is more exposed, in task terms, than a plumber or a nurse. The Tufts researchers are explicit about this inversion: "AI is not just automating routine tasks -- we see it moving up, targeting the cognitive and analytical work that defines high-skill, high-wage careers."
The reason is structural. AI systems, particularly large language models and generative design tools, are very good at modular, well-specified tasks -- writing code to a specification, generating financial projections from structured data, producing first drafts from a brief. These are the core competencies of many high-skill knowledge roles. By contrast, AI struggles with work that is unstructured, high-context, and physical: reading a situation in real time, adjusting to an unpredictable environment, building trust with a specific person who has a history and an agenda. That kind of work tends to be lower-paid and harder to credentialize, but it is also harder to automate.
A separate analysis published this week by Computerworld offers a counterweight. Researchers at Gartner and analysts at staffing firms argue that jobs displaced by AI will reappear elsewhere, pointing to LinkedIn data that estimated AI created 1.3 million new jobs globally in recent years, concentrated in data annotation, AI-adjacent engineering, and model training roles. The argument is not that displacement does not happen, but that the savings companies realize from cutting certain roles tend to get redeployed: more quality assurance, more AI trainers, more people managing the systems that replaced earlier workers.
Both things can be true simultaneously. Jobs can morph and new roles can emerge, and the transition can still be deeply painful for specific workers with specific skills in specific industries. The problem with the displacement-then-retraining story, as the Tufts data makes uncomfortably clear, is that it assumes workers can move up. If the high-skill, high-wage roles are under as much pressure as the entry-level ones -- and the evidence suggests they are -- then "up" becomes a less useful direction. The question of where displaced knowledge workers go, when the knowledge work itself is being automated, does not yet have a convincing answer.