안면인식 업체 클리어뷰 AI(WIRED 단독 보도)가 'InquiryIQ'라는 실험적 AI 분석 도구를 은밀히 개발·테스트한 것으로 확인됐다. 이 도구는 수사관이 검색한 인물 정보를 바탕으로 웹을 자동 탐색해 직장, 별명, 인간관계, 신체 특징 등을 종합한 프로필을 생성하며, 일론 머스크의 xAI(현재 SpaceX와 합병) 모델도 테스트에 포함됐다. 전문가들은 수사 기간을 몇 주에서 몇 분으로 단축하는 강력한 기능인 만큼, 근거 없는 인물 감시 비용을 낮추고 편향된 출력과 결정 과정 불투명성 등 심각한 인권·프라이버시 우려를 제기한다.
번역된 본문
2020년 안면인식 업체 클리어뷰 AI(Clearview AI)는 인터넷에서 30억 장이 넘는 사진을 무단 수집해 경찰과 보안 전문가에게 얼굴을 이름으로 바꿔주는 서비스를 제공하면서 악명을 떨쳤다. 이제 이 회사는 다시 웹 검색 실험을 진행하고 있는데, 이번에는 AI를 활용해 수사 기관이 이름 뒤에 숨은 인물—그가 누구인지, 누구를 아는지, 어디에 사는지, 온라인에 어떤 흔적을 남겼는지—를 밝혀내도록 돕는 것이 목적이다.
WIRED가 확인한 바에 따르면, 클리어뷰는 자사가 '실험적 AI 분석 보조 도구'라고 설명하는 것을 조용히 구축하고 테스트해 왔다. 아직 출시되지 않은 'InquiryIQ'라는 이 도구는 수사관이 클리어뷰 검색으로 찾아낸 정보를 입력받으면 웹 전역으로 자동 확장 탐색을 벌여—웹페이지를 열고, 이미지를 분석하며—수사 대상 인물의 가능한 고용주, 별명, 관계자, 신체 특징을 담은 프로필을 조립하도록 설계되었다. WIRED가 분석한 도구 코드에 따르면 InquiryIQ 인터페이스는 나이, 성별, 인종 정보를 제공하면 시스템이 검색 중 '더 스마트한 결정'을 내리는 데 도움이 된다고 안내한다. 이러한 결정을 구동하기 위해 테스트된 모델 중 하나는 그록(Grok)을 개발한 일론 머스크의 회사인 xAI 출신이다. (SpaceX와 xAI는 2월에 합병했다.) 머스크는 그록을 이른바 '워크(woke)' AI 시스템의 대안으로 내세워 왔으며, 이 챗봇은 인종차별적·극단적·선동적인 출력으로 반복적으로 논란이 되어 왔다. SpaceXAI가 이러한 인구통계학적 입력을 어떻게 활용하고 출력에 어떤 영향을 미칠지는 불분명하며, 클리어뷰도 추측을 거부했다.
전문가들은 InquiryIQ와 유사한 도구들이 수일甚至 수주에 걸친 탐정 업무를 몇 분으로 압축함으로써 경찰 수사의 판도를 바꿀 수 있다고 말한다. 이러한 속도는 경찰이 범죄를 더 빨리 해결하는 데 도움이 될 수 있지만, 동시에 '어부지리식' 탐색(낚시성 수사)의 비용을 낮춰 경찰이 평소라면 수사하지 않았을 인물까지 감시하는 것이 실용적으로 만들 수 있다. 또한 생성형 AI는 동일한 출발점에서도 다른 답을 낼 수 있기 때문에, 시스템이 왜 한 단서를 다른 단서보다 추적했는지 사후에 재구성하기 어려울 수 있다고 그들은 지적한다.
클리어뷰는 WIRED에 InquiryIQ는 고객에게 제안하거나 전달된 적이 없으며 현재 형태로 출시할 계획도 없는 프로토타입이라고 밝혔다. 회사는 또한 이 도구가 '자동 수사관'에 해당한다는 견해에 이의를 제기하며, InquiryIQ를 탐정들이 이미 수행하는 웹 검색을 자동화하는 제한적인 방법이라고 설명한다. 인터뷰에서 CEO 아모스 카일러(Amos Kyler)는 인터페이스에 보이는 SpaceXAI 및 기타 모델 옵션은 경찰이 연구에 사용할 모델을 선택하도록 하기 위해서가 아니라 클리어뷰 엔지니어들이 테스트 중 여러 모델의 성능을 비교할 수 있도록 포함된 것이라고 말했다. 카일러는 "단언컨대 어떤 법 집행 기관 사용자도 이것을 사용한 적이 없다"고 말했다.
WIRED는 클리어뷰 로그인 페이지가 로그인 전 모든 방문자의 브라우저로 전송하는 파일에서 InquiryIQ를 발견했다. 이는 WIRED가 지난달 플록 세이프티(Flock Safety)가 개발 중인 AI 검색 도구 'OS Investigate'를 발견하고 공개 접근 가능한 파일로부터 인터페이스 목업을 재구성했을 때 사용한 것과 동일한 기법이다. 클리어뷰의 파일에도 마찬가지로 사용자 인터페이스용 코드와 수천 줄의 텍스트—지침, 경고, 기능 설명—가 담겨 있으며, 이는 회사가 InquiryIQ를 어떻게 설계하고 설명했는지 보여준다. 다만 이 파일들은 클리어뷰 서버 내부에서 정확히 어떤 일이 벌어지는지, 도구가 얼마나 잘 작동하는지, 누가 사용했는지는 드러내지 않는다. 해당 텍스트에 따르면 InquiryIQ는 '웹 소스에서 개인 데이터를 자동으로 발견하고 보강하도록' 구축되었다. 클리어뷰는 InquiryIQ 인터페이스의 겉보기 완성도가 출시 임박 제품으로 오해되어서는 안 된다고 말한다. 회사는 현대 AI 도구 덕분에 정교한 프로토타입 구축이 훨씬 빨라졌으며, 클리어뷰 엔지니어들에게는 이제 가능하기 때문에 프로토타입을 신속히 진행하라는 지시가 내려간다고 설명한다.
2017년에 설립된 클리어뷰는 창립 초기 몇 년간 대중의 시선 밖에서 주로 활동했다. 피터 틸이…(원문 이하 생략)
Comment Loader Save Story Save this story Comment Loader Save Story Save this story In 2020, face-recognition firm Clearview AI became infamous for scraping more than 3 billion photos from the internet to turn faces into names for police and security professionals. Now the company is experimenting with trawling the web again, this time using AI to potentially help law enforcement fill in the person behind the name—who they are, who they know, where they live, and what they’ve left behind online. Clearview has quietly built and tested what it describes as an experimental AI “analyst assistant,” WIRED has learned. Called InquiryIQ, the unreleased tool is designed to take details an investigator unearths from a Clearview search and then automatically fan out across the web—opening webpages, analyzing images, and assembling what it finds into a profile containing the possible employers, aliases, associates, and physical characteristics of a person police are investigating. InquiryIQ’s interface states that supplying age, gender, and race can help the system make “smarter decisions” as it searches, according to code for the tool analyzed by WIRED. One of the models that the company tested to power those decisions comes from SpaceXAI, the Elon Musk company behind Grok . (SpaceX and xAI merged in February.) Musk has pitched Grok as an alternative to supposedly “woke” AI systems, and the chatbot has repeatedly drawn scrutiny for racist, extremist, and inflammatory outputs. It is unclear how SpaceXAI would use those demographic inputs or how they might shape its output, and Clearview declined to speculate. Experts say InquiryIQ and tools like it could reshape police investigations by compressing days or weeks of detective work into minutes. That speed could help police solve crimes faster, but it also lowers the cost of fishing expeditions, making it practical to scrutinize people police might otherwise never have investigated. And because generative AI can produce different answers from the same starting point, they say, it may be difficult to reconstruct why the system pursued one lead instead of another. Clearview tells WIRED that InquiryIQ is a prototype that has never been pitched or shipped to customers and is not currently planned for release in its present form. The company also disputes that the tool amounts to an automated investigator. Instead, it describes InquiryIQ as a limited way to automate the web searches detectives already perform. In an interview, CEO Amos Kyler says the SpaceXAI and other model options visible in the interface were included so Clearview’s engineers could compare how different models performed during testing, not so police could choose which model ran the research. “No law enforcement user has ever used it, period,” Kyler says. WIRED discovered InquiryIQ in files that Clearview’s login page sends to any visitor’s browser before they sign in. It was the same technique WIRED used last month to uncover OS Investigate , an AI-powered search tool being developed by Flock Safety , and to reconstruct a mock-up of its interface from publicly accessible files. Clearview’s files similarly contain code and thousands of lines of text for its user interface—instructions, warnings, and feature descriptions that show how the company has designed and described InquiryIQ. They do not reveal exactly what happens inside Clearview’s servers, how well the tool works, or who has used it. According to the text, InquiryIQ is built to “automatically discover and enrich personal data from web sources.” Clearview says the apparent completeness of InquiryIQ’s interface should not be mistaken for a product nearing release. Modern AI tools have made it much faster to build sophisticated prototypes, the company says, and the directive to Clearview’s engineers is to move prototypes along rapidly because it’s now possible to do so. Founded in 2017, Clearview spent its first years operating largely out of public view. Peter Thiel invested $200,000 that year, and the company soon began signing up police departments and offering free trials around the country. A 2020 HuffPost investigation later detailed founder Hoan Ton-That’s ties to the far right, including a 2016 Republican National Convention dinner with white nationalist Richard Spencer and participation in a private online community populated by far-right activists and extremists. Ton-That later apologized for his past writings. In 2020, The New York Times revealed that Clearview had scraped more than 3 billion images from Facebook, YouTube, Venmo, and millions of other websites. Overnight, the report transformed the company into one of the country’s most controversial surveillance firms. Tech companies demanded that it stop harvesting their users’ photos; lawsuits and regulatory investigations followed. But Clearview kept scraping. Its database grew from more than 3 billion images in 2020 to what the company now says is well over 70 billion, and Clearview says its technology is used by more than 2,000 law enforcement agencies nationwide. Ton-That stepped down as CEO in December 2024 and left the board the following spring. Kyler, who became CEO in October last year, joined Clearview as an engineer in 2019 and presents a more deliberate, technical face of the company. Where Clearview’s early years were defined by boundary-pushing, Kyler talks instead about controls, auditing, and oversight. “The mission today is the same,” Kyler says. But he describes the company’s recent focus as “refinement” and “ensuring that the product hits the kind of expectation of integrity that our customers expect.” Signals and Noise For years, Clearview said its job stopped at surfacing possible leads for law enforcement using face recognition. In a 2022 post, Ton-That wrote that it was “up to the investigator to follow those links and do more research to find additional information.” InquiryIQ appears designed to take on some of that work. As Clearview describes it, InquiryIQ would begin after an investigator has already run a face-recognition search and identified details they consider relevant. Those details can be added to a profile alongside information such as age, gender, race, hair color, and eye color, which the interface says can help the AI make “smarter decisions when running its searches.” Kyler says InquiryIQ was conceived as a way to test pieces of information an investigator had already identified as potentially relevant. “We looked at it in the form of trying to put together a proposition and then invalidating a proposition,” he says. The system would take a “factoid,” run searches around it, and ask, as Kyler puts it, “Is this related or not?” The code WIRED reviewed describes InquiryIQ as able to run web and image searches, browse web pages, and use face recognition on photographs it encounters. As the system searches from information supplied by the investigator, it is designed to build what Clearview calls a “Candidate Graph” of possible identities and associates, while filling out the subject’s profile with possible addresses, phone numbers, employers, social media accounts, arrest history, and aliases. Clearview is not the first company to automate the process of trawling publicly available information. Other intelligence platforms sold to law enforcement, including ShadowDragon’s SocialNet, Penlink’s Tangles, and Fivecast, help investigators uncover aliases and associates and map a person’s digital footprint. Andrew Guthrie Ferguson, a George Washington University law professor who studies AI and policing, describes this kind of automated investigation as “digital rummaging.” “They're basically going to create a profile of you based on all of the random digital clues you left on the internet,” he explains. In practice, that means gathering scattered bits of information that once lived in separate places but were difficult to connect. Woodrow Hartzog, a Boston University privacy scholar, argues