The big gap between AI capabilities and replacing jobs

AI’s rapid improvement at specific tasks does not mean it can replace whole jobs, the central argument goes. Many roles depend on coordination, judgment in unusual cases, and personal dealings with customers or other parties - work common in service, blue-collar, and white-collar fields such as insurance, advertising, and law. The piece argues that forecasts of mass unemployment assume far more readily replaceable workers than actually exist: Anthropic research reportedly finds about 30% of workers have no measured AI exposure, with many others only lightly exposed. Even in coding and other areas where AI performs impressively, doing tasks on request is not the same as independently managing a job. In some cases, AI instead lets workers accomplish much more, acting as a set of virtual assistants while their jobs remain recognizable.

The gap between capability and adoption also comes from the difficulty of connecting AI to real-world tools and processes. In one example, a novice trying to set up a GitHub repository received outdated, confusing instructions from AI and needed repeated help; the model could not create an account because signup required email verification and a CAPTCHA. An experienced employee could have handled the task quickly. Changing interfaces, restricted access to websites, and incompatible systems create bottlenecks that AI cannot simply overcome, leaving people to bridge the gaps. These practical constraints help explain why powerful models have not yet transformed workplaces as quickly as their performance might suggest.