While judging a national AI Olympiad in Kazakhstan, Daniil Orel began noticing something strange in the students’ code. Some comments looked less like notes written by a contestant and more like instructions or text generated by a large language model. Eventually, he was no longer certain whose work he was judging.
“At one point, I realised that as a human judge I could no longer be completely certain whether I was evaluating a student’s own work or code generated with the assistance of a language model,” Orel said.
That uncertainty became a research problem. It also points to a larger one increasingly confronting AI researchers: a system can produce the right answer, generate functional code or score highly on a benchmark and still fail in ways that matter once people begin relying on it.
Stay up to date with the latest news. Follow KT on WhatsApp Channels.
Three incoming PhD students at Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi are approaching different versions of that problem. Orel is studying the reliability and security of AI-generated code. Emirati researcher Ali Aljaberi is examining vulnerabilities introduced through AI-assisted software development. Amna Alhammadi, also Emirati, is moving from machine learning into human-computer interaction to ask whether technically capable AI systems are actually useful, understandable and appropriate for the people they serve.
All three completed master’s degrees at MBZUAI before returning for doctoral study. More than 40 incoming PhD students this academic year previously completed their master’s at the university, while MBZUAI has also admitted its first Human-Computer Interaction PhD cohort.
When working code is still unsafe
Orel’s Olympiad observation eventually developed into research on distinguishing human-written, AI-generated and hybrid code.
One of his projects, AICD Bench, now evaluates AI-code detection across two million examples, 77 models and nine programming languages. The research found that detection systems struggled particularly when the language or domain changed, and when code combined human and AI contributions or had been deliberately modified to obscure its origin.
His PhD moves deeper into what happens after AI-generated code is accepted.
“A major limitation is that Code-LMs are still judged largely on whether they can solve a task or produce functional code,” Orel said. “That is important, but functionality is not the same as reliability or security.”
“A model may generate a working website from a single prompt, for example, while the resulting code still contains vulnerabilities that matter enormously in a real deployment.”
Independent research published at ACL this year has reached a similar warning. Repository-level evaluations found that leading language models could perform well on smaller coding tasks yet continue to struggle with secure coding once placed in more realistic software environments. Another benchmark built from real-world repositories found that improving functional correctness did not necessarily improve security.
For Aljaberi, that gap between “works” and “safe” is also the centre of his research.
His master’s examined vulnerabilities in AI-generated code. Now beginning a Computer Science PhD, he wants to investigate how weaknesses can be detected earlier and how safeguards can become part of AI-assisted development rather than arriving after software has already been deployed.
“The pace of AI development means cybersecurity can no longer be treated as a problem we revisit only after a technology is widely deployed,” he said.
His work also crosses into entrepreneurship. Aljaberi co-founded Wiqayah AI, a UAE-developed threat-modelling platform intended to help organisations conduct security assessment and compliance processes internally.
“Research gives you the opportunity to examine emerging problems before they become established industry challenges,” he said. “Entrepreneurship forces you to ask a different question: how can that understanding be turned into something organisations can use?”
When the benchmark misses the person
For Alhammadi, the limitation lies elsewhere. She completed her master’s in Machine Learning, studying how systems learn from data and how their performance is measured. For her PhD, she has deliberately moved into Human-Computer Interaction.
“Technical performance alone cannot tell us whether a system is understandable, trustworthy, accessible or genuinely useful to the people it is intended to serve,” she said. “I want to learn how to design and evaluate AI with people, not simply for them.”
One research direction she hopes eventually to explore is AI that could support Quranic tajweed, Arabic pronunciation and recitation.
The idea began during an ordinary evening with members of her community. Someone asked whether anybody had experience in AI. Alhammadi raised her hand, and a conversation followed about whether technology could help people practise recitation.
But the boundary matters as much as the capability. “I do not see AI replacing teachers, scholars or the human and religious knowledge involved in Quranic recitation,” she said. “Its value could be complementary: helping learners practise consistently between lessons, identify recurring pronunciation patterns and receive timely feedback.”
Doing that properly, she said, would require collaboration with tajweed experts, Arabic-language specialists, teachers and learners.
The problem extends beyond a single use case. Recent multilingual AI research continues to find significant performance gaps between English and lower-resource languages, even among models marketed as multilingual. A 2026 benchmark covering 61 languages found persistent disparities between English and low-resource languages, while other work has identified gaps in safety protections when models move beyond English-dominated evaluation.
The question after capability
The three researchers aren’t anti-AI or saying the intelligence doesn’t work; their questions begin precisely because it increasingly does, and because it is imperially more important that AI can generate software. The next question is whether that software is secure. Models can perform strongly on benchmarks. Researchers still need to know whether that performance survives a change in language, culture or real-world conditions.
For Orel, that changes what researchers should mean when they describe AI as good. “Strong headline performance does not necessarily mean a model behaves equally well across languages, communities or use cases,” he said.
The next phase of AI research may therefore be less spectacular than proving another model can perform a new task. It is the slower work of finding out what happens when that model leaves the benchmark and meets the real world.
Source: Khaleej Times