AI News HubLIVE
站內改寫2 分鐘閱讀

待翻譯:An AI-supervised remote exam went so badly that 58,000 students must retake it

AI 服務暫時不可用,以下為來源摘要,待恢復後補全翻譯:No one yet knows what kinds of cheating may have taken place. The exam was multiple choice rather than essay-based, so it is difficult to look for typical indications of AI cheating (complete answers being pasted immedi…

來源Hacker News AI作者: sbulaev

AI 服務暫時不可用,以下為來源正文,待恢復後補全翻譯。

No one yet knows what kinds of cheating may have taken place. The exam was multiple choice rather than essay-based, so it is difficult to look for typical indications of AI cheating (complete answers being pasted immediately into text boxes, etc.). It is also possible that students relied on cheat sheets, leaked questions, and other traditional cheating methods. A New York Times report last week on the cheating scandal noted that “cheating tips were circulating widely before the tests went live this spring. Online accounts advised students to position monitors outside the web-camera frame where they could access ChatGPT or other AI models. Others suggested hiding earphones under their hair, or hiring someone else to take the exam, off camera.” Because this year’s exam was administered remotely, applicants had to install and use security products that tried to protect testing integrity. A report in El País said that the test required the LockDown Browser from Respondus. According to the company, once an exam begins in this browser, students are “unable to print, copy, go to other web addresses, or access other applications on the computer. Web searches, instant messaging, minimizing the browser, and hundreds of other functions are also prevented. Once the exam is completed, students are permitted to exit the browser and their computer will return to its normal state.” UNAM also used a proctoring system from Territorium that monitored each test-taker’s webcam with AI algorithms. The system watched for signs that another person had taken the applicant’s place in front of the computer, or for the use of cell phones or earphones, or even for the applicant leaving the frame. The test also had one human supervisor for every 150 applicants; supervisors received alerts from the system and could notify UNAM officials about irregularities. These measures do not seem to have worked well. (UNAM canceled nearly 2 percent of total exams for unspecified conduct issues, so it apparently caught a few people.) Even UNAM has acknowledged “the probability that a significant number of applicants may have received help” on the test. AI expert Raul Rojas, speaking to NPR, said statistical modeling suggested “that almost half of the students were cheating when taking the exam online.” Now, almost 58,000 applicants will have to try again in an official exam room, watched by a human proctor.