A study by Emma Wiles, a business professor at Boston University, finds that people perform worse when an AI tool is presented as a “coworker” or “employee.” Participants caught 18% fewer errors when the output was said to come from an “AI employee” compared with when it was described as a chatbot.
How the research was done
The study involved 1,261 managers; nearly a third reported that their companies already frame AI agents as employees, and 23% said such agents are listed on organizational charts. Wiles’s results show that labeling a tool as an employee reduces human accountability: participants felt less responsible for the AI’s outputs and were 44% more likely to escalate questionable outputs to a manager for review rather than correcting them themselves.
Why this matters
The label used for an AI system influences how people interact with it. Presenting AI as a “coworker” or “employee” increases the risk of misplaced blame and abdicated human responsibility: failures that stem from human choices, incentives, or poor oversight can be attributed to the AI instead. The article cites the example of a bombing of a girls’ school in Iran, which many attributed to Claude, despite evidence pointing to a cascade of human errors.
Technical progress versus marketing
Agentic AI — systems programmed to iterate toward a goal until it is achieved — have become measurably better at complex tasks. Yet describing these systems as colleagues or employees is a large conceptual leap: the label creates unrealistic expectations about capabilities and, according to Wiles’s research, harms the humans expected to supervise them.
Alternative approaches and empirical findings
Daron Acemoglu, an economist at MIT and the 2024 Nobel Prize winner, argues that AI should be optimized to augment human capabilities rather than replace humans. A Stanford effort asked 1,500 workers across 104 jobs what tasks AI could do and what would actually be helpful. While workers welcomed automation in some areas (for instance, law clerks saw potential for AI to track case progress), in many cases the tasks tech experts thought suited to automation — such as verifying customer credit ratings for sales representatives — were exactly the tasks workers said they did not want or need an agent to perform.
Consequences
Branding an AI tool as an “employee” is easy and convenient, especially when something goes wrong, but it is largely a marketing move. It does not make the tool more capable, and Wiles’s study indicates it degrades human oversight and accountability. Because human actors retain actual agency and responsibility, systems should be framed and designed to enhance human performance rather than serve as scapegoats.
Summary
The research highlights how terminology and framing shape workplace interactions with AI: labeling tools as employees reduces people’s error detection and correction, increases escalation and blame-shifting, and obscures human responsibility. Developers, employers, and policymakers should avoid casual personification of AI and prioritize designs that support and extend human decision-making.
