When the Label Changes, So Does the Judgment
A company gives its AI tool a name – Alex – assigns it a title, lists it on the org chart, and tells staff that Alex is a digital colleague. Sounds modern. Sounds efficient. But according to research from Boston University business professor Emma Wiles, that framing actively degrades the performance of the humans working alongside it. People caught 18% fewer errors when reviewing work attributed to an agentic “AI employee” compared to when the same output was described as coming from a chatbot. The only thing that changed was the name.
This is not an abstract concern about corporate branding. Microsoft, OpenAI, Anthropic, and Google have all, since April, released tools explicitly built around managing teams of AI agents – many marketed as digital colleagues with the cognitive flexibility of actual humans. Nvidia CEO Jensen Huang talked up workplaces full of “digital humans” last year. And nearly a third of the 1,261 managers in Wiles’s study said their companies already frame AI agents as employees, with 23% going so far as to list them on official org charts.
The Accountability Gap That Grows With the Hype
Wiles’s study identified something more corrosive than simple overconfidence. When participants believed they were supervising an AI “employee” rather than operating a software tool, they reported feeling less personally responsible for the output that tool produced. That shift in perceived responsibility had measurable consequences: participants were 44% more likely to escalate the AI’s questionable work upward to a manager instead of correcting it themselves, which directly cancels the time-saving rationale for deploying an AI agent in the first place.
The accountability problem does not stay inside office walls. AI agents are being embedded into health care, education, government, and military systems. As that happens, the “employee” framing creates a convenient place to deposit blame when decisions go wrong. The April bomb strike on a girls’ school in Iran was widely attributed in public discourse to Claude, Anthropic’s AI model. Evidence instead points to a cascade of human errors – in planning, oversight, and decision-making – that the AI label helped obscure. Naming a tool after a person does not transfer moral agency to it. It transfers moral confusion to everyone around it.
MIT economist Daron Acemoglu, who won the Nobel Prize in 2024 and studies AI’s effect on the labor market, is direct about the strategic problem: “AI agents right now are being marketed as things that can replace humans, and I think that’s just a losing proposition. They should instead be optimized so that they can improve human capabilities, which is not what they have [been] at the moment.” That is a meaningful distinction – not between optimism and pessimism about AI, but between two completely different design philosophies.
Agents themselves have genuinely improved. Technically, they can be understood as AI tools programmed to operate in a loop until a specific goal is achieved, and they have become measurably better at handling complicated, multi-step tasks. That real progress, however, is being packaged in marketing language that encourages users to extend human social trust to what is still, fundamentally, software. The mismatch between the tool’s actual capabilities and the expectations generated by calling it a coworker is where the performance breakdowns begin.
What Workers Actually Want AI to Do
A Stanford research effort offered a different model. Researchers presented 1,500 workers across 104 jobs with information about which tasks AI could theoretically handle in their roles, then asked those workers what would genuinely be most useful. The results cut against assumptions held by many AI developers. Law clerks, for instance, said AI could be helpful for tracking progress across a large volume of cases – a monitoring function, not a judgment function. But tasks that technical experts flagged as ideal candidates for AI automation were frequently the same ones workers said they did not want automated. Sales representatives specifically did not want an AI agent verifying customer credit ratings, a task that experts had considered a natural fit.
That gap – between what engineers decide AI should do and what workers actually need – rarely appears in product launch announcements. It is also the gap that gets papered over when companies give AI tools human names and human titles. Framing the agent as a colleague with defined responsibilities makes it easier to deploy broadly and harder for workers to push back on without sounding like they are resisting a person rather than questioning a workflow decision.
The Real Cost of Anthropomorphic Branding
There is a straightforward commercial logic to calling an AI tool “Alex.” It lowers the psychological friction of adoption, it gives the product a face in internal communications, and – critically – it distributes accountability in ways that are convenient when things go sideways. If Alex made a mistake, the question of who sanctioned Alex’s role and who was supposed to catch Alex’s errors becomes harder to answer cleanly.
What Wiles’s research makes clear is that this convenience is not free. It is paid for in reduced human vigilance, inflated deference to machine output, and a systematic weakening of the oversight instincts that are supposed to keep AI-assisted work reliable. The study participants were not careless people – they were managers operating with the same cognitive tendencies that most workers bring to their jobs. The framing shifted their behavior without their awareness.
The companies building these agent frameworks are not unaware of the psychological literature on anthropomorphism. They are also not indifferent to the liability implications of calling software an employee. That 23% of the managers in Wiles’s study work at companies already listing AI agents on org charts suggests the shift in language is moving faster than the governance structures meant to contain its consequences – and someone, at some point, will have to explain which org chart entry bears responsibility when something goes wrong.
