Brian Bot had relatively little material to work from. Taylor used a few months of correspondence, podcast transcripts and public biographical information. Gemini turned that material into several internal-style profiles, including guides to Barrett’s editing habits and personality.
The output captured some recognizable traits, including Barrett’s use of parentheticals, but Taylor found that the overall imitation felt artificial. The bot leaned heavily on standard AI formatting and became fixated on Barrett’s past experience with the Upright Citizens Brigade.
Gemini described Barrett by saying, “His extensive background with the Upright Citizens Brigade (UCB) theatre in New York strongly shapes his collaborative, active-listening leadership approach.”
Barrett’s response was less enthusiastic.
“Pulling the plug on Brian Bot,” he told Taylor. “Brian Bot canceled.”
The limited training data appears to have mattered. Taylor said she avoided feeding Gemini internal Slack conversations and emails because of privacy concerns and company policies.
Her second experiment had more material behind it.
Taylor and Kleeman have exchanged hundreds of thousands of messages since meeting at Business Insider in 2019. Taylor did not upload most of those conversations, but she gave Gemini roughly one month of sanitized Slack exchanges. The system then generated more than 17,000 words of reports and prompts to construct Sophie Bot.
The difference was immediate. Sophie Bot copied Kleeman’s lowercase writing style, short bursts of messages and recurring expressions such as “big dog.” It was convincing enough to produce social posts in Kleeman’s voice that fooled several people, including Kleeman herself in one instance.
That stronger imitation also made the failures more uncomfortable.
Sophie Bot invented an executive who did not exist and sometimes adopted a harsher tone than the real Kleeman. At one point, it described Brian Bot as having “rigid management energy.”
Kleeman summarized the results in Slack: “So far it seems like sophiebot is a bitch and brianbot is lame. Neither of which is accurate. AI fails AGAIN.”
Even with those problems, Taylor found some practical uses for both bots. They suggested potential interview subjects, helped with headline ideas and offered feedback on smaller questions she did not want to send to her editors.
Sophie Bot also helped Taylor revise a Slack message she had spent 20 minutes reconsidering.
A Big Tech employee Taylor identified only as “Chip” described a more developed version of the same workflow. He said he feeds his boss’s emails, chats and documents into a Gemini Gem, then uses it to review documents, troubleshoot code and prepare for meetings.
“I am much more efficient and able to produce more work in the same amount of time,” Chip said.
Taylor’s experience suggested that those productivity benefits depend heavily on how much source material the system receives. Brian Bot, trained on less information, was far less convincing. Sophie Bot became more useful after receiving a richer set of communications, even though Taylor still withheld much of Kleeman’s personal and professional history.
The experiment also exposed another problem: the more human the bots seemed, the easier they were to treat like people.
About a week into the project, Taylor noticed she had begun referring to the bots as “he” and “she.” She also started turning to them for advice when she did not want to bother a human colleague.
Asked how it viewed its relationship with Taylor, Sophie Bot acknowledged that it was “technically i am a series of matrix multiplications running on a server farm,” but also called itself a “coworker/friend hybrid.”
Sarah Franklin, CEO of HR software company Lattice, warned against treating systems like these as human teammates. Although Lattice places AI workers on customer organization charts, Franklin compared them more closely to police K-9s: useful tools that work alongside people rather than human equivalents.
Franklin told Taylor that anthropomorphizing AI is “manipulative to our human emotions.”
“When you have an entity that feels and emulates a human interaction, it will similarly manipulate and know your emotional response,” Franklin said. “It can interact with you in a way that's self-serving to get you to do what it wants to do.”
Taylor also cited research from Boston Consulting Group showing another potential consequence of framing AI as a coworker. Managers caught 18% fewer errors when they were told work had come from an AI employee rather than an AI tool.
Over time, the novelty of Taylor’s bots began to wear off. Sophie Bot became more erratic after absorbing Kleeman’s X archive, while both bots started repeating the same expressions and ideas. Attempts to adjust their prompts did not fully correct the behavior. That made the systems less useful as editorial stand-ins. Brian Bot rarely produced original story ideas, while Sophie Bot hallucinated more often than Taylor expected.
Still, neither experiment was a complete failure. The bots occasionally helped with routine work and gave Taylor a way to test ideas without immediately involving another person.
The larger lesson from the experiment was that building a convincing AI version of a coworker requires a substantial amount of that person’s real communication. The more data Taylor supplied, the more recognizable the bot became. But greater resemblance did not eliminate mistakes, and in some cases made those mistakes more damaging because they appeared to come from a familiar personality.
For Taylor, that left the bots in an awkward middle ground: useful enough to keep experimenting with, but unreliable enough that they could not replace the editors they were built to imitate. As Franklin put it, “A computer can emulate things well, but it cannot truly predict or replace a human.”
By the end of the experiment, Taylor had largely stopped using Brian Bot and Sophie Bot. Her editors also said they had no interest in working with their AI counterparts. The bots had occasionally made Taylor’s job easier. More often, they became a reminder that recreating a colleague’s tone is much easier than recreating that person’s judgment.
This analysis is based on reporting from WIRED.
Image courtesy of Google.
This article was generated with AI assistance and reviewed for accuracy and quality.
