Researchers Asked ChatGPT, Gemini and Claude Which Jobs Are Most Exposed to AI. The Chatbots Wildly Diagree

Why It Matters

Inconsistent AI‑generated exposure scores undermine their use in workforce planning, education policy, and hiring decisions, prompting a need for multi‑model approaches. The findings caution governments and businesses against basing strategic moves on a single AI forecast.

Key Takeaways

• ChatGPT‑5, Gemini 2.5, Claude 4.5 rank jobs inconsistently.
• Supervisory and hybrid cognitive‑physical roles show greatest model disagreement.
• Exposure scores vary up to double depending on AI model used.
• Researchers advise against single‑model reliance for policy or career decisions.

Pulse Analysis

The working paper from Northwestern and American University reveals a fundamental flaw in the emerging practice of using large language models to generate "exposure scores" for occupations. By feeding identical labor‑market data to ChatGPT‑5, Gemini 2.5 and Claude 4.5, the researchers observed starkly different risk rankings, with the greatest variance appearing in roles that blend management, cognition, and physical effort. This divergence underscores that current AI systems lack a unified understanding of task complexity and automation potential, limiting their utility as standalone forecasting tools.

For policymakers and corporate strategists, the study serves as a warning sign. Exposure scores that swing by a factor of two across models can lead to dramatically different policy prescriptions—whether to invest in reskilling programs, adjust immigration quotas, or redesign curricula. Moreover, the research points to a feedback loop: occupations already saturated with AI tools, such as financial analysis, generate more training data, potentially inflating their perceived vulnerability in future model iterations. Relying on a single model could thus amplify biases, misinforming decisions that affect millions of workers.

The broader implication is a call for ensemble methods and transparent benchmarking in AI‑driven labor forecasts. Industry leaders like Anthropic’s Dario Amodei and Nvidia’s Jensen Huang have voiced contrasting views on the speed and scale of job displacement, reflecting the uncertainty inherent in the technology. As AI continues to evolve, combining multiple model outputs, incorporating human expertise, and continuously validating predictions against real‑world outcomes will be essential to produce reliable guidance for education, hiring, and economic policy.