February 13, 2026
AUB
researcher Wael Khreich’s new research-based bias detection system exposes
subtle discrimination embedded in language and machine learning.
In recent months, debates over bias in artificial intelligence have moved from academic journals to front-page news. The latest flashpoint came after Grok, the AI system developed by Elon Musk's company, generated sexualized and non-consensual images of women, a reminder that generative models not only reflect society's norms but also can magnify its worst impulses. As AI systems increasingly shape hiring, credit decisions, content moderation, and everyday communication, the stakes of such incidents are growing.
And the real threat—gender bias—is even more pervasive and difficult to detect, says Wael Khreich, assistant professor of business information and decision systems at the American University of Beirut's Suliman S. Olayan School of Business.
Khreich has been studying machine bias since long before generative AI entered the mainstream. A researcher whose work spans responsible AI, natural language processing, business analytics, and computer security, he focuses on how bias enters AI systems and works on methods to detect, measure, and mitigate this bias before it causes harm.
“The real risks aren't dramatic misuses or deepfakes that get a lot of attention," Khreich says. “Those are damaging, but it's what happens when small, quiet biases are repeated millions of times without being noticed that can potentially create harm for generations."
Now, Khreich and his team have built a tool to address such bias: Genderly, a research-based system built on one of the field's most comprehensive datasets for measuring gender bias in language and AI systems.
Genderly scans text and AI system outputs for hidden gender bias and flags problematic language for the researcher using it before that language can spread. The platform trains and fine-tunes a combination of traditional learning, pre-trained models, and large language models on its carefully labeled datasets to expose biases that might otherwise go unnoticed. Created primarily for researchers and developers who evaluate and train AI systems, it allows them to identify patterns of bias and measure whether models are improving over time.
“The tool was built around high-quality data that has been carefully reviewed by human annotators," Khreich explains. “This allows us to test different models—from traditional machine learning to large language models—and measure bias in a reliable way. With that kind of information, we can tell whether a system is actually improving or just avoiding scrutiny."
But such tools, he argues, are only a starting point for broader changes in how AI systems are built and governed. In a conversation with AUB@Work, Khreich reflected on what his research reveals about the limits of current “debiasing" efforts and why the choices technologists make now will shape the information ecosystem for generations.
Public attention has recently focused on Grok and the generation of sexualized images of women. How do you see that episode in relation to your research on bias in AI?
What we work on is a quieter, more subtle form of bias, the kind that influences decisions about hiring, credit scoring, or how people are described in everyday language. Those biases are usually unintentional, inherited from data and then amplified by models.
What we've seen with Grok is different in tone but related in consequence. The sexualization of women through AI-generated images is far more aggressive and deliberate. It can be used to silence, harass, or blackmail. Technologically, this has been possible for years, but now it's faster and more accessible. When systems are released with minimal guardrails, the damage scales instantly, and the response often lags far behind the harm.
Can you say more about minimal guardrails? Has your own work tested how robust the safeguards actually are?
One of my graduate students, Ghina Baassiri, tackled exactly that question in her master's thesis. She developed an automated method for stress-testing AI safety filters called Dynamic Adversarial Prompting. Using a single, carefully constructed query, she bypassed the safety mechanisms of every major commercial model tested—including GPT-4, Gemini, and Claude—with success rates ranging from 73 to 95 percent. The dedicated guardrail systems designed to catch harmful outputs failed to block more than three-quarters of the attempts.
Consider what that means: a graduate student with a university's resources can systematically defeat the safety measures of billion-dollar companies. Now imagine what a well-funded malicious actor could do. It tells you something about the gap between what the industry promises about safety and what it actually delivers. And it's why detection tools like Genderly matter—you can't fix what you can't measure, and right now the industry's ability to measure is lagging far behind its ability to deploy.
Your team began studying gender bias in language data before large language models became dominant. What did you find when you followed patterns over time?
We found something striking: the patterns barely change. Male-exclusive terms appear far more frequently than female-exclusive ones, and benevolent sexism persists regardless of the author's gender. Even with increased awareness and public discussion around gender bias, the language itself remains remarkably stable over time.
The problem is that modern AI systems are trained on this material at an enormous scale. When large language models absorb these patterns, they don't just reflect them; they create feedback loops, where model-generated text becomes part of the data used to train future systems.
Many companies claim they are “debiasing" their models. Based on your research, how effective are those efforts?
Most current approaches focus on treating the symptoms rather than the root causes. You might reduce bias in one direction but create it somewhere else. We've seen models overcorrect, reducing bias against women, for example, while introducing new distortions or overlooking intersectional issues entirely.
Bias isn't something you can simply switch off. Without high-quality evaluation datasets, it's hard to know whether a system is genuinely fairer or just optimized to avoid public backlash. In some cases, companies are responding more to reputational pressure than to rigorous measurement.
Looking ahead, as AI becomes ever more embedded in daily life, what bias-related issues should we be paying the most attention to?
The danger isn't only spectacular failures like deepfake abuse, but slow, systemic distortions that affect opportunity, representation, and trust. For everyday users, that means learning to approach AI-generated content with healthy skepticism, especially when it appears neutral or authoritative.
One of the most important defenses against bias is learning to question AI-generated content rather than treating it as neutral or authoritative. We also need diverse teams involved in design, stronger governance frameworks, and accountability that goes beyond voluntary guidelines. Either way, the speed of AI development means the choices being made now will shape how future generations access information and make decisions.