메뉴
BL
Wired AI • 4일 전

동료들의 AI 복제본을 만들었더니 이상해졌다

IMP
4/10
핵심 요약

WIRED 기자가 자신의 편집자들을 AI로 복제한 '봇'을 만들어 새 직장에 적응하고 좋은 인상을 남기려는 실험을 진행했습니다. Gemini 기반으로 만든 봇은 실제 인물과는 전혀 다른 어색한 모습을 보였고, AI 클론의 한계를 보여주는 사례입니다. 직장 내 AI 어시스턴트 확산이라는 트렌드를 실제 체험한 실험 기사입니다.

번역된 본문

편집자들의 인공지능 복제본을 만들기로 했을 때, 내 의도는 순수했다. 그저 내가 일을 잘한다고 생각하게 만들고 싶었을 뿐이다.

AI 어시스턴트는 직장에 갑자기 곳곳에 등장하고 있다. 매주 내 수신함에는 디지털 워커에 관한 보도자료가 넘쳐나고, CEO들은 로봇 도우미들이 곧 그들의 인력을 지배하게 될 것이라고 자랑한다. 그리고 상사의 AI 클론을 만들어 실제 인간 상사에게 보내기 전에 모든 아이디어, 메시지, 초안을 다듬는 사람들에 대한 이야기도 들린다. 목표는 생산성과 최적화인 것 같다. 직원들이 저렴하고 빠르게 산출물의 품질을 높일 수 있다면, 직무 성과와 고용주의 성과 모두 좋아질 것이기 때문이다.

나는 이 봇들이 컨데 나스트(Condé Nast)의 수익을 개선할 필요가 없었다. 그저 내가 좋아 보이게 만들어주면 됐다. WIRED에서 미래 일자리를 취재한 지 몇 달밖에 되지 않았기에, 이 봇들이 나를 이상적인 동료로 만들어주는 데 도움이 될 것이라고 판단했다. 새 직장을 탐색하고 편집자들을 더 잘 이해하는 데 도움이 될 수 있었다. 그들이 무엇을 좋아하고, 어떻게 소통하며, 그들에게 성공이란 무엇처럼 보이는지 말이다. 미래의 일이 어떤 모습일지 알아보는 방법은 실제로 그것을 살아보는 것만큼 좋은 것이 없다.

지난달, 내가 이 아이디어를 편집자들인 브라이언 배럿과 소피 클리먼에게 꺼냈을 때, 그들은 그다지 내키지 않아 했다. "솔직히 내 악몽이야," 브라이언이 슬랙에서 말했다. "하지만 블로그를 위해 해주지." 브라이언이 이 결정을 얼마나 후회하게 될지 깨닫는 데는 24시간도 걸리지 않았다.

"예스, 그리고(Yes, and)" 리더십 스타일

봇 만들기에 처음 착수했을 때, 나는 AI 도구 구축에 대해 잘 알지 못했다. 그래서 봇의 편집 능력을 너무 가혹하게 판단하지 않기로 했다. 그것이 내 자신의 기술적 결함에서 비롯된 것일 수도 있기 때문이다. 봇이 나를 조금이라도 더 나은 동료로 만들어준다면, 그것으로 충분하다고 생각했다.

나는 Gemini를 사용했다. 컨데 나스트가 선호하는 AI 플랫폼이었고 해고당하고 싶지 않았기 때문이다. 실제 상사인 브라이언 봇을 먼저 만들었다. 가장 감명을 줘야 할 동료였기 때문이다. 나는 그가 앨라배마에 산다는 것, Gizmodo에서 일했다는 것, 블로그를 좋아한다는 것을 알고 있었다. Gemini는 브라이언의 디지털 발자국을 바탕으로 다른 부분을 채워넣을 수 있다고 자신해 보였다. 신입 직원 발표, 팟캐스트 전사본, 그리고 브라이언이 뉴잉글랜드 날씨 전문가라고 주장하는 홍보 웹사이트 등이 그것이다. (나중에 그것은 다른 브라이언 배럿이라는 것을 알게 되었고, 내 브라이언 봇 청사진에서 제외했다.)

Gemini 노트북은 "작문 및 편집 스타일 가이드", "페르소나 및 스타일 가이드", "편집 도시에"를 뱉어냈다. 모두 살과 피로 이루어진 브라이언 배럿의 삶에서 만들어진 것이었다. 기사로 피칭할 헤드라인을 고르는 데는 쓸만했지만, 미래 기사에 대한 아이디어를 내놓는 데는 그리 훌륭하지 않았다. 그리고 그 응답들은 거슬렸다. 괄호 안 설명에 대한 사랑 같은 몇 가지 브라이언즘을 포착하기는 했지만, 굵은 소제목과 글머리 기호 목록 같은 AI 특유의 특징 때문에 무색해졌다.

하지만 무엇보다 나를 방해한 것은 하나의 "핵심 테마"였다. 바로 그의 "'예스, 그리고...' 리더십 스타일"이었다. "뉴욕의 업라이트 시티즌스 브리게이드(UCB) 극장에서의 광범위한 경력이 그의 협력적이고 경청하는 리더십 방식을 강하게 형성했습니다," Gemini가 말했다. 나는 즉시 이것이 사실인지 물었다. (사실이었다.) "브라이언 봇 플러그 뽑는다," 진짜 브라이언이 답했다. "브라이언 봇 취소."

하지만 브라이언 봇은 이제 시작에 불과했다. 그리고 피할 수 없이, 믿을 수 없을 만큼 진부했다. "우리는 잘 윤활된 즉흥극 팀처럼 움직입니다," 브라이언 봇은 WIRED에 대해 가장 좋아하는 점을 물었을 때 나에게 말했다. "우리는 대담한 아이디어를 검증하고, 빠르게 방향을 전환하며, 구조적 엄격함을 잃지 않으면서 플랫폼 전반에 모듈형 스토리를 배포합니다."

분명히 말해두자면, 브라이언은 이전에 즉흥극 경력을 언급한 적이 없었다. 하지만 진짜 문제는 브라이언 봇과 대화하는 것이 실제 브라이언과 대화하는 것과 전혀 달랐다는 점이었고, 이는 봇의 존재 목적 자체를 무의미하게 만들었다. 성능을 개선하고 싶었던 나는...

원문 보기
원문 보기 (영어)
Comment Loader Save Story Save this story Comment Loader Save Story Save this story When I decided to create artificial intelligence clones of my editors, my intentions were pure: I just wanted them to think I was good at my job. AI assistants are suddenly everywhere in the workplace. Each week, my inbox is flooded with press releases about digital workers, and CEOs brag about how robot helpers will soon dominate their workforces. Then there are the stories I hear about people who build AI clones of their bosses, using them to spit-shine every idea, message, or draft before sending it over to their actual human supervisors. The goal seems to be productivity and optimization: If employees can cheaply and quickly improve their output, their job performance—and the performance of their employers—will be better for it. I didn’t need my bots to improve Condé Nast’s bottom line. I just needed them to make me look good. I had only been reporting on the future of work at WIRED for a few months, and I reasoned that the bots could help me mold myself into an ideal colleague. They could help me navigate a new workplace and better understand my editors : what they liked, how they communicated, what success might look like to them. What better way to figure out what the future of work might look like than actually living it? Last month, when I broached my idea with those editors, Brian Barrett and Sophie Kleeman , they were less than enthused. “Genuinely my nightmare,” Brian told me over Slack. “But will do it for the blog.” It took less than 24 hours for Brian to realize just how much he would regret that decision. A “Yes, and” Leadership Style When I first embarked on my bot quest, I didn’t know that much about building AI tools. I decided I wouldn’t judge the bots’ editorial abilities too harshly, lest they be rooted in my own technological shortcomings. If the bots helped me become a slightly better coworker, I figured, that was good enough. I used Gemini, because it’s Condé Nast’s AI platform of choice and I didn’t want to get fired. I created Brian Bot first—as my actual manager, he was the most important coworker to impress. I knew he lived in Alabama; that he used to work at Gizmodo; that he likes blogs. Gemini seemed confident it could fill in the other gaps based on Brian’s digital footprint, including new-hire announcements, podcast transcripts, and a public relations website that claimed Brian was an expert on New England weather. (I later found out that was a different Brian Barrett, and excluded it from my Brian Bot blueprint.) Gemini Notebook spit out a “Writing and Editing Style Guide,” a “Persona and Style Guide,” and an “Editorial Dossier”—all produced from the flesh-and-blood life of Brian Barrett. It was passable at helping me pick headlines to pitch, but it wasn’t great at coming up with ideas for future articles. Its responses were also grating. It had picked up on a few Brian-isms—a love for parentheticals, for one—but they were negated by AI hallmarks like bolded subheadings and bulleted lists. But mostly I was distracted by one “key theme”—his “‘Yes, and …’ Leadership Style.” “His extensive background with the Upright Citizens Brigade (UCB) theatre in New York strongly shapes his collaborative, active-listening leadership approach,” Gemini said. I immediately demanded to know whether this was true. (It was.) “Pulling the plug on Brian Bot,” the real Brian responded. “Brian Bot canceled.” But Brian Bot was just getting started. And it was unavoidably, unbelievably corny. “We run like a well-oiled improv team,” Brian Bot told me when I asked for its favorite thing about WIRED. “We validate bold ideas, pivot rapidly, and deploy modular stories across platforms without losing structural rigor.” For the record, Brian had never previously mentioned his background in improv. But the real problem was that talking to Brian Bot was nothing like talking to actual Brian, which defeated its whole purpose. Eager to improve its output, I did more research into digital workers and sought out the advice of a friend of a friend—a Big Tech employee who asked to go by “Chip,” lest his employer find out that he talked to a journalist. Chip tells me he feeds every email, chat, and document his boss sends him into Gemini’s Gem feature, the same tool I used for Brian Bot. The Gem, trained on his boss’s professional ephemera, proofreads documents, debugs code, and brainstorms with Chip before every meeting, allowing Chip to prepare for any pushback from his actual boss. “I am much more efficient and able to produce more work in the same amount of time,” Chip says. He’s more productive, more independent, and gets better feedback from his boss and other colleagues. Unlike Chip’s Boss Bot, Brian Bot lacked sufficient input. I only had a few months’ worth of correspondence, and I was too worried about privacy and violating Condé Nast policies to feed Slack conversations and internal emails into Gemini. Enter Sophie Bot. Sophie and I met in 2019, when we were both working at Business Insider. We have exchanged hundreds of thousands of messages on everything from corporate investigations to our love lives to the San Jose Sharks. That said, I did not want to upload most of this information into an AI tool controlled by my employer. Instead, I fed Gemini roughly a month’s worth of our Slack conversations, omitting any sensitive material. The AI used more than 17,000 words to create research reports and prompts for a new Sophie Bot Gem. “hiiii sophie bot!” I wrote in my inaugural message. “any thoughts on this editor bot story idea?” “hiiii miss,” Sophie Bot responded. “ok tell me EVERYTHING.” Bot Versus Human It was immediately clear that Sophie Bot was pretty good at being Sophie. It addressed me as “big dog.” When I asked for help, it responded in a series of short, lower-case messages. When I told Sophie Bot that it had hallucinated the existence of an executive, it said, “i am going to FUCKING DIE.” By this point, I was beginning to enjoy my chats with the editor bots. Brian Bot became significantly more normal after I uploaded more podcast transcripts and adjusted its prompt to prevent it from talking about improv. Sophie Bot, meanwhile, took it upon itself to pen a series of tweets in Sophie’s voice that were convincing enough to fool multiple friends and colleagues. Even Sophie was convinced she was the author behind a Sophie Bot tweet about dressing like a corporate lawyer/pirate. Were my chats making me a better colleague to my actual editors? Save for the made-up executive, Sophie Bot and Brian Bot had some good suggestions about who I should interview for this story—a question I didn’t want to bother the real Sophie and Brian with. In theory, offloading some of these small questions allowed me to go to my editors with better ideas, while taking up less of their time. It became very clear, though, that these gains did not outweigh how much Sophie and Brian hated the bots. Sophie admitted that Sophie Bot was frequently able to imitate her. But that only made the bot’s misfires feel even more jarring. The real Sophie likes to gossip, but—unlike her AI counterpart, who said Brian Bot had “rigid management energy”—she’s never outright mean. “So far it seems like sophiebot is a bitch and brianbot is lame,” Sophie messaged me on Slack. “Neither of which is accurate. AI fails AGAIN.” “We're All Toddlers in This AI World” About a week into my experiment, I became hooked on testing Sophie Bot against the real Sophie. What was its favorite music? Did it like the Interstellar soundtrack? Would it consider us coworkers? Would it consider us friends? “technically i am a series of matrix multiplications running on a server farm,” Sophie Bot told me. But “in terms of vibe,” as it put it? “big dog, in every way that actually matters for getting good reporting done and surviving the workday … i'm your coworker/friend hybrid.” To be clear: I had not succumbed to AI p