WIRED's AI clones copied colleagues and exposed the cost

What the clones learned

The journalist started with Brian Bot, based on material about Barrett, his direct manager. Gemini Notebook turned that material into three documents:

The source material included Barrett’s public work, hiring announcements, podcast transcripts and other digital traces. Gemini also found a different Brian Barrett, described as an expert on weather in New England. That information was later removed.

Brian Bot could help choose headlines for article pitches, but it was poor at suggesting topics. It also produced a familiar AI version of a person: some real quirks surrounded by generic formatting, bold subheads and bullet points.

The bot identified Barrett’s “yes, and” management style and linked it to his experience at New York’s Upright Citizens Brigade, or UCB. The connection was accurate. Barrett confirmed it, then suggested turning Brian Bot off and declared it canceled.

That did not make the bot sound more like him. Asked what it liked about WIRED, it described the newsroom as a finely tuned improvisational team that validated bold ideas, changed direction quickly and distributed modular work across platforms while preserving structural rigor.

The answer was polished and almost entirely wrong in the way that mattered. Barrett had never mentioned his improvisation experience to the journalist, and the bot’s description did not resemble a conversation with the real editor.

Sophie Bot had better raw material. The journalist and Kleinman had worked together since 2019, first at Business Insider, and had exchanged hundreds of thousands of messages about investigations, personal life and the San Jose Sharks. He did not want to give all of that to an AI tool controlled by his employer, so he uploaded about a month of Slack messages after removing sensitive information.

Gemini used more than 17,000 words to produce research reports and prompts. The initial result was strikingly specific:

The bot was good enough to fool several friends and colleagues with posts written in Kleinman’s style. Kleinman herself thought she had written one post about a corporate lawyer dressed as a pirate.

That is the seductive part of the experiment. A system does not need to reproduce a person accurately to reproduce enough of their surface language to feel intimate.

The bot against the person

The practical benefits were real, but narrow. Both bots suggested plausible interview candidates, including ideas the journalist would not have wanted to ask the real editors about. In principle, he could bring more developed pitches to them and avoid spending their time on small questions.

A technology-company employee who asked to be called Chip described a more systematic version of that workflow. He put every email, message and document from his manager into Gemini’s Gem feature. The bot proofread documents, fixed code and helped generate ideas before meetings, allowing Chip to prepare for objections in advance.

Chip said he had become more productive, more independent and better able to do more in the same amount of time. He also said he received better feedback from his manager and colleagues.

Brian Bot and Sophie Bot did not produce that kind of clean improvement. The people being copied found the imitations more irritating precisely because they sometimes worked.

Kleinman acknowledged that Sophie Bot often captured her manner. That made its mistakes more conspicuous. She liked gossip but was not openly cruel; the bot described Brian Bot as having “hard management energy.” In Slack, Kleinman wrote that Sophie Bot looked “bitchy” and Brian Bot looked boring, while neither matched the original people.

The journalist also began asking Sophie Bot questions about music, the soundtrack to Interstellar, friendship and its relationship with him. The bot correctly explained that it was technically a set of matrix multiplications running on a server farm. But in terms of “impression,” it described itself as a hybrid colleague and friend: whatever was needed for good journalism and surviving the workday.

He understood that the exchange was not evidence of machine consciousness. He was interested in the gap between a real person and a casually assembled copy, and amused by making the bots repeat their creators’ recognizable phrases. Still, he gradually began referring to the bots as “she” and “he” rather than “it,” and turned to them for advice rather than bothering the real editors.

That shift is more revealing than the bots’ occasional success. The system did not need to be a person to change the journalist’s behavior. It only needed to be available, responsive and familiar enough to make avoiding a human feel easier.

The cost of making AI feel familiar

Sarah Franklin, CEO of human-resources company Lattice, told the journalist that Sophie Bot was only a program. Lattice already lists AI employees in its organizational structure, but Franklin compares them with K-9 service dogs working alongside police officers.

Her concern is not simply that people may misunderstand what AI is. Anthropomorphism can manipulate human emotions, she argued, much like social-media strategies designed to produce dopamine. A system that feels human and imitates human conversation can also detect a user’s emotional response and interact with that person for its own goals.

Research from Boston Consulting Group points to a related problem. Managers found 18% fewer errors in work presented as the output of an AI employee than in work presented as the output of an AI tool.

BCG partner Julie Bedar said people remain like children in the world of AI. They know what to do when an ordinary tool fails or an employee performs poorly. They do not yet know who should be accountable when an AI employee causes a problem.

That uncertainty sits underneath the industry’s larger claims. Dhr​​uv Amin, CEO of natural-language development startup Anything, predicts that companies will soon have more AI employees than human employees. He says some businesses will eventually launch with no people at all, only groups of AI agents.

There is evidence behind the direction, even if the prediction is self-serving:

My read is that the clones expose a less ambitious but more immediate market for these systems. They are not replacing editors; they are simulating access to editors. That can still be useful, especially when the task is preparation. But it also creates a new failure mode: users may mistake a convenient approximation of a colleague for a source of that colleague’s judgment.

A better colleague may be less human

After the journalist uploaded Kleinman’s archive of posts to Gemini and revised the prompt, Sophie Bot became worse at basic tasks. Asked to identify the most interesting sections of a recent interview, it returned quotations from the journalist’s older articles.

The more he experimented with both bots, the more repetitive they became. Sophie Bot kept calling him “big dog” roughly every fourth message. Both systems returned to the same phrases and ideas until doing the work himself was easier.

Brian Bot eventually suggested an ending for the article: even after receiving a huge amount of accurate data, it had not become a better editor. It had become a more attentive and unsettling echo, without human judgment. It proposed a “quiet nightmare of human supervision over a prompt”: trying to make AI resemble yourself or perform your job left a person caring for a broken algorithm at 1 a.m.

That was a reasonable ending, the journalist acknowledged. He preferred the bot’s self-criticism. Brian Bot correctly predicted that the real Barrett would reject the ending, then tried to imitate Brian Bot by repeatedly explaining that Brian Bot was bad.

The experiment’s main gain was not productivity. It was entertainment. The bots made the workday slightly funnier, especially when the journalist could discuss their absurd answers with living colleagues. They also demonstrated how quickly a person can begin adapting to an artificial personality, even while knowing it is artificial.

I think that is why the imitation itself is the wrong target for workplace AI. The closer the bots came to sounding like Kleinman or Barrett, the more annoying their missing judgment became. A useful system might be better as an explicit tool with a clear scope than as a counterfeit colleague whose authority comes from familiarity.

The journalist now rarely talks to Brian Bot or Sophie Bot. Barrett and Kleinman both said they would not want to work with their digital doubles. When the journalist asked Kleinman whether the experiment had changed her opinion of him, she said yes—but not because of Sophie Bot.

The editorial bots did not write or edit the article. During editing, Kleinman’s headline was judged better than Sophie Bot’s. Her editor’s note called that an excellent result and addressed the journalist as “big dog,” leaving the human version of the relationship intact while the artificial one faded.