Daniil Orel noticed anything odd in the kids’ code while he was judging a national AI Olympiad in Kazakhstan. Some remarks appeared to be instructions or text produced by a huge language model rather than notes written by a contestant. After a while, he was unsure of whose work he was evaluating. “As a human judge, I realised at one point that I could no longer be absolutely certain whether I was evaluating a student’s own work or code generated with the assistance of a language model,” Orel remarked.
This ambiguity turned into a research issue. It also highlights a more significant issue that AI researchers are increasingly facing: once people start depending on a system, it may deliver the correct response, write useful code, or score highly on a benchmark, but it may still fail in significant ways.
Different approaches to the topic are being taken by three incoming PhD candidates at Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi. Orel is researching the security and dependability of code produced by AI. Ali Aljaberi, an Emirati researcher, is investigating vulnerabilities brought about by AI-assisted software development. Amna Alhammadi, another Emirati, is transitioning from machine learning to human-computer interaction in order to investigate if technically sound AI systems are truly beneficial, intelligible, and suitable for the people they serve.
Also Read:
Saudi Arabia Closes a Vital Oil Pipeline Following an Iraqi Drone Strike
Investigators Start Looking into why a Cargo Plane Overran the Runway in Miami
