Here’s a finding worth sitting with: the moment you call an AI agent a “coworker” instead of a “tool,” you start doing a worse job overseeing it. According to MIT Technology Review, new research from Boston University business professor Emma Wiles found that people caught 18% fewer errors when work was described as coming from an agentic “AI employee” rather than a plain chatbot. Same output. Different label. Measurably worse human judgment.
What stands out here is how small the trigger is. Nobody changed the technology. They changed a word. And that word reshaped how seriously people checked the work.
What the study looked at
Wiles surveyed 1,261 managers and tested how framing changed their behavior toward AI-generated work, MIT Tech Review reports. The setup was simple: show people output, attribute it either to a software tool or to an “AI employee,” and watch what they do with it.
The results point in one clear direction. When the AI was framed as an employee, participants:
- Caught 18% fewer errors in its work
- Saw themselves as less responsible for the output
- Were 44% more likely to escalate questionable work up to a manager instead of fixing it themselves
That last number is the quiet killer. The whole point of using an AI agent is to save time. If people stop trusting their own corrections and start kicking decisions upstairs, the time savings evaporate. You’ve added a layer, not removed one.
Why the framing flips who’s in charge
Wiles’s read, per MIT Tech Review, is that calling an agent an “employee” inverts our sense of authority. We supervise tools. We defer to colleagues. When a spreadsheet macro produces a number, you check it. When a “colleague” hands you the same number, you assume they own it. So you relax. You escalate the weird stuff instead of owning the call.
This isn’t a fringe quirk. Nearly a third of the managers said their companies already frame AI agents as employees, and 23% say they list them on org charts. The vocabulary is being set by the people selling the tools. Last year Nvidia CEO Jensen Huang talked up workplaces full of “digital humans.” Since April, Microsoft, OpenAI, Anthropic, and Google have all shipped tools for managing teams of AI agents, many pitched explicitly as digital colleagues.
Why this matters beyond the office
The stakes climb fast once agents move out of the back office. As these tools get embedded into health care, education, government, and warfare, MIT Tech Review warns they risk becoming a convenient place to dump blame for failures that were really caused by bad human decisions and weak oversight. The piece points to how a bomb strike on a girls’ school in Iran got popularly pinned on Claude, when the evidence points to a cascade of human errors. Call something a coworker and you’ve built yourself a scapegoat.
Worth noting the research isn’t anti-AI. Agents, basically AI tools programmed to loop until they hit a goal, have gotten genuinely better at complex tasks. The problem isn’t the capability. It’s the marketing language wrapped around it.
What you can actually do with this
The practical takeaway is unusually concrete for a research finding:
- Watch your language. Call them tools, not teammates. The word sets the expectation, and the expectation sets how carefully your people check the work.
- Keep humans named as owners. If an agent’s output ships, a specific person is responsible for it. Don’t let “the AI did it” become an acceptable sentence.
- Audit your escalation paths. If reviewers are bumping AI work upstairs instead of correcting it, you’re paying for AI and getting more meetings.
- Push back on org-chart theater. Listing agents as employees may feel modern, but the data says it makes your actual employees less careful.
One limitation to keep in mind: this measures managers’ framing and behavior in a study setting, not long-term outcomes inside live companies. The direction is clear, the long-run magnitude less so.
The trend isn’t slowing down. Every major lab is racing to sell you a “digital colleague.” The smarter move is to take the capability and leave the costume. You can read the full study writeup at MIT Technology Review.