A WIRED reporter created Gemini-based replicas of two editors to test whether digital coworkers could improve her performance. The bots occasionally suggested useful sources or edits, but their stronger pattern was imitation without judgment: repeated catchphrases, invented facts and behavior that became more unsettling as it appeared more human.
The experiment began with a practical workplace premise. Some employees feed a manager’s messages and documents into AI so they can test a draft before sending it to the real person. The reporter hoped replicas of editors Brian Barrett and Sophie Kleeman would reveal what they valued and help her become a better colleague.
Sparse Data Produced a Caricature
The first replica used Gemini and public information about Barrett, including professional announcements and podcast transcripts. It copied a few stylistic traits, such as parenthetical phrasing, but leaned on generic AI formatting and repeatedly turned his background in improvisational theater into an exaggerated management philosophy.
The bot could help select a headline, yet struggled to produce original story ideas. Adding more transcripts and changing its instructions made it less awkward, but it never became a convincing version of the editor.
The reporter then built a Sophie replica from about a month of Slack messages, excluding sensitive material. Gemini produced more than 17,000 words of research and prompts from that input. This bot imitated Kleeman’s short lowercase messages and familiar phrases well enough to fool some colleagues with generated social posts.
Better mimicry did not produce consistent accuracy. The bot invented an executive, became meaner than the real person and eventually returned material from unrelated articles when asked to identify important passages in an interview.
Privacy and Emotion Constrained the Test
The experiment could not simply ingest every available workplace message. The reporter avoided internal emails, most Slack history and personal conversations because of privacy and company-policy concerns. A contact who had built a more useful boss bot attributed its performance to feeding it extensive professional material.
That trade-off is central to digital cloning. More intimate data may improve imitation while increasing the privacy cost for the person being copied and everyone communicating with them.
The emotional effect also changed as the bots became familiar. The reporter began using human pronouns and turned to them for instant guidance rather than disturbing real colleagues. Lattice chief executive Sarah Franklin warned that humanlike interaction can manipulate emotion even when the apparent coworker is only software.
A Boston Consulting Group study cited by WIRED found that managers detected 18 percent fewer errors when they believed work came from an AI employee rather than an AI tool. The label can therefore alter oversight before the system has earned greater trust.
Assistance Worked Better Than Replacement
The replicas sometimes reduced small interruptions. They proposed interview subjects, reviewed a Slack message and offered editing suggestions. Those uses resemble a private sounding board rather than an autonomous employee.
As the experiment continued, answers became repetitive and prompt changes failed to correct recurring behavior. The reporter concluded that the bots were not close to replacing skilled editors. Their ideas lacked originality, and one replica hallucinated more often than the other AI tools she had used.
The most useful output may have been an accidental critique of the premise. The Barrett bot described itself as a hyper-attentive echo that lacked human judgment and required supervision. The real editors said they would not want to work with their replicas.
The experiment does not measure all AI employee systems, and its bots were built by one nonexpert with deliberately limited data. It does reveal a practical boundary: a clone can approximate tone and answer low-stakes questions while remaining unreliable about the person it supposedly represents. The workplace value came from occasional assistance and shared amusement, not from reproducing a colleague.
