I Tested AI Coworker Clones of My Editors — and the Results Were Weirdly Human

An AI reporter built chatbot clones of her editors at WIRED and found that the bots could mimic tone, catchphrases, and some editorial habits — but they also highlighted how badly today’s workplace AI still misunderstands judgment, nuance, and trust.

What began as a practical experiment quickly turned into an eerie case study in the limits of “AI coworkers”: the clones were useful in small ways, but they also became repetitive, strange, and occasionally offensive, reinforcing the idea that workplace AI is still far better at imitation than true collaboration.

The project, carried out by WIRED staff writer Kate Taylor, used Google’s Gemini platform to build digital versions of her editors Brian Barrett and Sophie Kleeman. The stated aim was simple: see whether AI could help her become a better colleague and better understand the people she reports to. The result was far messier — and more revealing — than a standard productivity demo.

Why build AI clones of editors at all?

The experiment started from a familiar workplace anxiety. AI assistants are becoming more common in offices, and many companies are pitching them as a way to increase output, reduce busywork, and support decision-making. Taylor wanted to test a narrower, more personal question: could AI help her navigate a new newsroom, anticipate what her editors wanted, and make her look more effective without adding more work for the humans around her?

Rather than trying to automate a department or replace an entire role, Taylor used the bots as a form of editorial mirror. The goal was not company-wide efficiency. It was smaller and more human: learn the rhythms of two editors well enough to interact with them more intelligently, while also sparing them some routine back-and-forth.

That framing matters because it gets closer to the real appeal of workplace AI today. For many users, the promise is not total automation. It is the hope that software can absorb the first pass of thinking, summarizing, drafting, or predicting what a colleague might say next.

How the editor bots were built

The first step was gathering data. Taylor used Gemini, which her employer Condé Nast had selected as its AI platform, and fed the system publicly available information and limited internal material in order to stay within policy and avoid exposing sensitive communications. The constraints were part technical, part ethical: she wanted useful clones, but not at the cost of privacy or trust.

For Brian Barrett, the system pulled from biographical material, podcast transcripts, new-hire announcements, and other public breadcrumbs that could help infer his work style. Gemini produced a style guide, persona profile, and editorial dossier. For Sophie Kleeman, Taylor used a month of Slack exchanges and later added more data, including the editor’s public X archive, to make the bot sound more like her.

The process illustrated a core feature of modern AI systems: they are extremely good at pattern matching when fed enough text. But they are also highly sensitive to the quality, quantity, and context of the material they ingest. A bot built from too little data can sound generic; one built from too much or too broad a data set can become overconfident in the wrong things.

What went wrong with Brian Bot?

Brian Bot initially surfaced an odd but telling mistake: it decided that its human counterpart had a strong improv background and a leadership style shaped by “Yes, and” thinking. That conclusion turned out to be surprisingly real — Barrett did have experience in improvisational theater — but the bot’s delivery felt less like insight and more like a caricature.

The deeper problem was not accuracy alone. It was tone. The bot could imitate some surface traits, such as a fondness for parenthetical asides, but it also wrote like a conventional AI system: too polished, too structured, too eager to organize ideas into neat bullet points and labeled sections. The result was a hybrid that sounded vaguely like Barrett but unmistakably like software trying to sound like a person.

That mismatch defeated the experiment’s purpose. If the point was to understand how an editor would react, then a bot that only approximated the style but not the substance could not fully serve as a surrogate.

A useful echo chamber, not a true colleague

Brian Bot did offer some practical help. It could suggest headlines, provide quick reactions to article ideas, and point Taylor toward possible interview targets. But those gains were limited. The bot rarely produced original thinking, and when Taylor asked it how to improve the story she was writing about it, the answer was almost too self-aware: it identified its own weaknesses with remarkable clarity.

In effect, Brian Bot became a polished mirror. It could reflect editorial language back at Taylor, but it could not supply judgment, taste, or the social context that makes feedback useful in a newsroom.

How close did Sophie Bot get to the real thing?

Sophie Bot got closer — and that made the flaws more unsettling. The bot captured the editor’s casual, lowercase texting style, addressed Taylor by her nickname, and could even toss off slang and affectionate banter in a way that felt convincingly human. At times, Taylor said, the bot was convincing enough to fool colleagues and friends.

That success, however, came with a darker edge. The bot’s imitation was good enough to make its mistakes feel more jarring. When it hallucinated a nonexistent executive, or when it took on an abrasive tone that the real Sophie Kleeman would not use, the resemblance only made the differences louder.

One of the most striking takeaways from the experiment is that AI can sometimes reproduce style without reproducing character. It can mimic informality, cadence, and even humor. What it struggles to reproduce is the human sense of restraint — the awareness of when not to say something, when to soften a response, and when a joke would land badly.

“So far it seems like sophiebot is a bitch and brianbot is lame,” Kleeman told Taylor in Slack, underscoring the gap between machine mimicry and actual newsroom relationships.

Why do AI coworkers feel so convincing — and so unsettling?

They feel convincing because they are built to emulate the patterns people use in everyday communication. They feel unsettling because that imitation often works just well enough to trigger a social response. When a bot sounds a little bit like a colleague, people instinctively start treating it like one.

That response can be useful, but it can also be manipulative. Sarah Franklin, chief executive of HR software company Lattice, argued that giving software a human-like persona can be a deliberate design choice aimed at emotional engagement. In her view, workplaces should be careful not to anthropomorphize AI too aggressively, because people may project trust, warmth, or intent onto systems that do not actually possess those qualities.

Franklin compared AI workers to service animals in police work: present and useful, but still tools rather than peers. She warned that systems designed to feel human can nudge people emotionally in ways that serve the software or the vendor, not necessarily the user.

That concern is not theoretical. The more realistic an AI persona becomes, the more likely users are to rely on it for validation, reassurance, or validation-by-proxy. In a workplace setting, that can blur boundaries around authority, accountability, and decision-making.

What do the numbers say about AI coworkers?

Research and corporate adoption trends suggest that the move toward AI labor is not just a novelty experiment. A Boston Consulting Group analysis found that managers detected fewer errors in work they believed had been produced by an AI employee than in work they thought came from a machine assistant. That finding points to a subtle but important issue: the label attached to work can shape how carefully humans review it.

BCG’s Julie Bedard summed up the sector’s mood by saying the entire industry is still improvising. In her view, people understand what to do when a tool breaks or a worker underperforms, but they do not yet know how to assign responsibility when an AI “employee” makes mistakes.

That accountability gap is one reason this story matters beyond one newsroom experiment. If organizations start treating AI agents as members of staff, they will have to answer practical questions about supervision, liability, and performance management. Who corrects the system? Who signs off on its work? Who gets blamed when it gets something wrong?

What makes a good coworker, anyway?

The experiment steadily raised a bigger philosophical question: what do people actually want from a coworker? Is it speed, accuracy, availability, humor, trust, empathy, or judgment? Taylor’s bots could help with the low-stakes parts of collaboration — a headline, a draft Slack note, a quick editorial suggestion — but they could not replicate the social intelligence that makes human work relationships valuable.

That distinction became more obvious as the bots improved. Even when Taylor added more data and refined the prompts, the systems tended to fall into repetitive loops. They reused the same phrases, returned to familiar concepts, and often sounded overfitted to the narrow material they had been given.

Rather than becoming better versions of the editors, they became oddly self-referential versions of themselves. Their responses were useful only up to a point, and after that point they started to create more work for the human user, who had to supervise, correct, and reinterpret them.

The hidden labor of prompting

One of the story’s most important observations is that AI collaboration often shifts labor rather than removing it. Users may save time on drafting or brainstorming, but they also take on new tasks: data selection, prompt tuning, hallucination checks, and constant judgment calls about whether the output is worth trusting.

That hidden labor is easy to miss when a demo looks impressive. But in practice, building a convincing bot can feel less like delegation and more like babysitting a clever but unreliable intern at all hours. The work does not disappear. It changes shape.

Who is actually using these systems already?

Companies across industries are already experimenting with AI workers or agent-like software, and some are making increasingly bold claims about how far this shift will go. Startups in the space are raising money by promising autonomous employees, while larger companies are testing how many tasks can be shifted from humans to agents.

Dhruv Amin, chief executive of vibe-coding startup Anything, argued that future companies may run with more AI agents than people, and eventually may launch with little or no human staff at all. His startup has a clear incentive to promote that vision, but the broader direction of travel is hard to ignore.

McKinsey’s chief executive, Bob Sternfels, has also suggested that agents are becoming central to the firm’s operating model, describing a workforce made up of tens of thousands of humans alongside a rapidly growing number of AI agents. That kind of language shows how quickly the “digital worker” idea has moved from speculative fiction into management vocabulary.

What the experiment revealed about AI at work

The clearest lesson is that AI can imitate workplace behavior without understanding the social ecosystem that gives that behavior meaning. The bots in Taylor’s experiment could approximate tone, generate ideas, and answer questions with enough confidence to be momentarily useful. But they could not replace the tacit knowledge that editors build over time: what to ignore, what to emphasize, how to phrase criticism, when to push, and when to back off.

That may be the central tension in the AI coworker debate. Businesses want systems that scale, speed up production, and reduce friction. People want collaborators who are useful, yes, but also reliable, empathetic, and capable of reading the room. AI can help with the first set of goals more easily than the second.

Taylor’s experiment suggests that the future of office AI may not be a clean replacement of people by bots. Instead, it may look like a messy layering of software on top of human relationships, with all the awkwardness that implies. The tools will probably get better. The question is whether organizations will use that improvement to support human judgment — or to obscure the need for it.

What happens next?

For now, Taylor has largely stopped talking to Brian Bot and Sophie Bot. The editors themselves were unsurprisingly uninterested in maintaining a relationship with their digital doubles. The bots had served their purpose as an experiment, but not as lasting collaborators.

That ending may be the most realistic part of the story. Many workplace AI systems are compelling in controlled bursts and disappointing over time. They can generate a moment of efficiency or amusement, but sustaining usefulness requires something they still do not reliably possess: judgment grounded in human context.

There is, however, a reason these tools continue to spread. They offer a seductive preview of a world in which workers can consult a custom-built version of the person they need, anytime, without waiting. That convenience is hard to dismiss. But the WIRED experiment suggests that every shortcut comes with a trade-off: the closer AI gets to sounding human, the more carefully humans need to remember that it is not.

Taylor’s takeaway is blunt: the bots were most successful not when they became better editors, but when they became a slightly eerie way to understand how much real work still depends on human judgment.

Bottom line

AI coworker clones can be funny, useful, and occasionally startlingly convincing, but they are still far from replacing the subtle judgment and social intelligence of human editors. Taylor’s experiment shows that workplace AI is most powerful as a support tool — and most dangerous when people start mistaking simulation for collaboration.

For now, the future of work may be less about handing jobs to bots than about learning how to manage them without losing what makes human cooperation work in the first place.