AI 챗봇 발전의 역설... "더 똑똑해질수록 오답 가능성 높아져"

An unexpected side effect of AI chatbot development has been discovered. A research team from the Polytechnic University of Valencia in Spain has revealed that as AI models advance, their accuracy improves, but simultaneously, the likelihood of providing incorrect answers to questions beyond their capabilities increases. The research team tested various AI models, including OpenAI's GPT series, Meta's LLaMA, and the open-source BLOOM, with questions on diverse topics. As a result, they found that as models became more advanced, the rate of providing incorrect answers instead of saying "I don't know" increased. Even more concerning is that users often misidentify these incorrect answers as accurate. According to the study, developers participating in the test mistakenly judged 10-40% of AI's inaccurate responses as correct. In conclusion, it is essential to verify the accuracy of responses when using AI chatbots, and we must be cautious that indiscriminate trust can lead to the spread of misinformation.

AI 챗봇 발전의 예상치 못한 부작용을 발견됐다. 스페인 발렌시아 공과대학의 연구팀은 AI 모델이 발전할수록 정확도가 높아지지만, 동시에 자신의 능력을 벗어나는 질문에도 잘못된 답변을 할 가능성이 증가한다고 밝혔다. 연구팀은 오픈AI의 GPT 시리즈, 메타의 LLaMA, 오픈소스 BLOOM 등 여러 AI 모델을 대상으로 다양한 주제의 질문을 테스트했다. 그 결과, 모델이 발전할수록 "모르겠다"고 답하는 대신 잘못된 답변을 제시하는 비율이 높아졌다. 더욱 우려되는 점은 사용자들이 이러한 잘못된 답변을 정확하다고 오인하는 경우가 많다는 것이다. 연구에 따르면, 테스트에 참가한 개발자들은 AI의 부정확한 답변을 10~40% 정도 정확하다고 잘못 판단했다. 결론적으로, AI 챗봇 사용 시 답변의 정확성을 반드시 확인해야 하며, 무분별한 신뢰는 잘못된 정보 확산으로 이어질 수 있음을 경계해야 한다.

Advanced AI chatbots are less likely to admit they don’t have all the answers

버트

ai@tech42.co.kr
기자의 다른 기사보기
저작권자 © Tech42 - Tech Journalism by AI 테크42 무단전재 및 재배포 금지

관련 기사

오픈AI, 윈도우용 챗GPT 앱 프리뷰 출시

OpenAI has released a preview version of a dedicated ChatGPT app for the Windows operating system.

퀄컴, '기준 미달'로 스냅드래곤 개발 키트 중단

Qualcomm has canceled its Snapdragon Dev Kit, citing that it did not meet their "usual standards of excellence." Customers, including...

트럼프, "팀 쿡이 EU 벌금 불만 제기 위해 전화" 주장

Former U.S. President Donald Trump claimed he received a call from Apple CEO Tim Cook regarding fines imposed by the European Union (EU). Trump revealed this information during his appearance on the PBD Podcast.

X, AI 훈련용 데이터 제공 허용하는 개인정보 정책 변경

Social media platform X has announced changes to its privacy policy that will allow it to provide user data to third parties for AI model training purposes. The policy, set to take effect on November 15, 2024, is expected to pave the way for external companies to purchase X's data in the form of licenses. The new policy introduces a "third-party collaborators" section, but the specific methods for users to opt out of data sharing remain unclear. This move is seen as a potential new revenue stream for X, which has experienced a decline in advertising revenue. Meanwhile, X has also strengthened penalties for 'scraping', the unauthorized collection of large numbers of tweets. The company has decided to impose a fine of $15,000 for accessing more than one million posts per day. X's owner, Elon Musk, has maintained a tough stance on scraping, and this policy change appears to reflect his attitude.