ChatGPT and Gemini were closest but still differed about a quarter of the time, suggesting AI-generated exposure scores used by policymakers and employers may be unreliable.
The authors said model rankings may be shaped by training data from early AI-adopting professions and urged researchers to compare multiple models and use real-world implementation surveys.
If AI models can’t agree on which jobs are at risk, why are we relying on them to plan our careers?
AI is freezing entry-level jobs. How will a generation start careers if the first step on the ladder is gone?
The 2.5% Reality: Measuring the Limited Real-World Impact of AI Automation in 2026
Overview
Despite rapid advancements in AI models, their real-world impact on job automation remains low as of May 2026. Although studies like the Remote Labour Index show steady improvements in AI capabilities, there is a significant gap between what AI could theoretically automate and what it actually achieves in practical, paying scenarios. Even the best AI systems perform well below human standards, automating only a small fraction of real-world tasks. This highlights that, while AI technology is progressing, its adoption and effectiveness in real jobs are still limited, especially for complex or creative work.