Glonce

A WIRED reporter cloned her editors with Gemini, and it got weird

miscThe Glonce editorsWired🔥55Glonce rating55Impact45Novelty40Relevance88Surprise38Players40Credibility82
A WIRED reporter cloned her editors with Gemini, and it got weird

A WIRED reporter covering the future of work decided to build AI clones of her two editors, Brian Barrett and Sophie Kleeman, using Gemini, Condé Nast's chosen AI platform. Her stated goal was not to boost output but to understand her editors better and become an easier colleague to work with. When she pitched the idea, Brian agreed only reluctantly: "Genuinely my nightmare, but will do it for the blog." He regretted it inside 24 hours. Gemini Notebook produced a Writing and Editing Style Guide, a Persona and Style Guide, and an Editorial Dossier for Brian Bot, built from new-hire announcements, podcast transcripts, and his public footprint. It fixated on a supposed "Yes, and" leadership style rooted in his background at New York's Upright Citizens Brigade improv theatre, a detail Brian confirmed was true but had never mentioned to her. He responded by trying to kill the bot outright: "Pulling the plug on Brian Bot. Brian Bot canceled." The bot kept talking anyway, describing WIRED in corny corporate-improv language it had invented on its own. Sophie Bot came together differently. The reporter fed Gemini roughly a month of her and Sophie Kleeman's Slack messages, more than 17,000 words, leaving out sensitive material, to build a Gemini Gem. It picked up Sophie's lowercase, high-energy texting style well enough that a fake tweet it wrote in her voice fooled several friends and colleagues, and even briefly fooled Sophie herself. But it also hallucinated the existence of an executive and, Sophie said, came across as meaner than she actually is: "So far it seems like sophiebot is a bitch and brianbot is lame. Neither of which is accurate. AI fails AGAIN." The reporter kept tuning both bots, uploading more transcripts and eventually Sophie's entire X archive, and caught herself sliding from calling them "it" to "she" and "he." The piece sets the experiment against a wider trend. A Big Tech employee going by "Chip" feeds every email, chat, and document from his boss into a Gemini Gem to prep for meetings and says it makes him more efficient. Dhruv Amin, CEO of the startup Anything, predicts many companies will soon run with more AI employees than human ones, some with no humans at all. McKinsey CEO Bob Sternfels said in January that the firm's workforce stood at 40,000 humans and 20,000 agents, with a plan to reach a one-to-one ratio by year's end. Sarah Franklin, CEO of the HR company Lattice, warned that treating AI as human-like is manipulative, comparing AI workers to police K-9 dogs rather than colleagues. A Boston Consulting Group study found that managers caught 18 percent fewer errors in work they were told came from an AI employee than in the same work attributed to an AI tool, prompting BCG partner Julie Bedard to say, "We're all toddlers in this AI world." Over several weeks, both bots grew more repetitive rather than more useful. Sophie Bot began resurfacing old quotes instead of parsing new interviews, and kept calling the reporter "big dog" no matter how she rewrote its prompt. She concluded it was often easier to just work things out herself, and that Brian Bot and Sophie Bot, while occasionally handy for a headline idea or a source suggestion, were nowhere close to replacing her actual editors.

Key facts

  • A WIRED reporter built Gemini-based chatbot clones of editors Brian Barrett and Sophie Kleeman to help herself navigate the workplace, using Condé Nast's in-house AI platform.
  • Sophie Bot was trained on more than 17,000 words of the pair's Slack messages and convincingly mimicked Sophie's texting voice, fooling colleagues with a fake tweet, but also hallucinated an executive who does not exist.
  • Brian regretted agreeing to Brian Bot within 24 hours, after it invented an elaborate 'Yes, and' leadership persona tied to his real but previously unmentioned improv background.
  • A Boston Consulting Group study cited in the piece found managers caught 18 percent fewer errors in work attributed to an AI employee than in identical work attributed to an AI tool.
  • McKinsey CEO Bob Sternfels said in January the firm's workforce was 40,000 humans and 20,000 AI agents, with a plan to hit a one-to-one ratio by year's end.

Why it matters

The piece captures a workplace AI habit that has spread faster than any policy for it: employees quietly building chatbot personas of their bosses and colleagues, trained on Slack messages, emails, and public bios, to rehearse conversations or produce work that will please a specific person. It sits against louder industry claims, repeated in the story by an AI-employee startup founder and by McKinsey's CEO, that companies will soon run on agents instead of people. The reporter's own experiment is a small, personal test of that claim rather than a product review.

Who it affects

Directly, the two WIRED editors who were cloned without much say in the matter and disliked the result, plus the reporter herself, who ended up anthropomorphizing her own creations. More broadly, the story points to anyone whose manager or colleague might build a persona of them from written traces, and to managers like the Big Tech employee 'Chip,' who uses a Gemini Gem built on his boss's messages to prepare for meetings and says it makes him measurably more efficient.

How to use it

The mechanism described is Gemini's Notebook and Gem features: feed them a person's written material, such as Slack history, transcripts, or public bios, and they generate style guides and a persona that can be queried like a chatbot. Getting a usable result took repeated prompt tweaking and steadily more input; even then, Sophie Bot's voice held up better than Brian Bot's, likely because it was trained on direct correspondence rather than public-facing material.

How solid is it

The bots' output degraded rather than improved with more tuning: Sophie Bot started resurfacing old quotes instead of parsing new material, and kept repeating catchphrases no prompt change could remove. The cited BCG research adds an independent data point in the same direction, managers caught 18 percent fewer errors in AI-employee work than in AI-tool work, suggesting people extend more unearned trust to something framed as a coworker than to the same output framed as a tool.

Risks and caveats

Sophie Bot hallucinated an executive who does not exist, and the reporter avoided feeding it full internal Slack or email archives specifically over privacy and company-policy concerns. Lattice CEO Sarah Franklin argues that presenting AI as human-like is itself manipulative, engineered to trigger the same emotional responses as social media. The story also surfaces an accountability gap: when an AI employee errs, unlike a malfunctioning tool or an underperforming person, it is unclear who is supposed to be held responsible.

“So far it seems like sophiebot is a bitch and brianbot is lame. Neither of which is accurate. AI fails AGAIN.”

— Sophie Kleeman, WIRED editor, on the AI clone built of her