Research by Ahmed Taha on AI alignment and trustworthy medical AI — correction selectivity under user pressure, demographic sensitivity in vision–language models, and documentation integrity in LLM-agent systems. Open manuscripts and supporting artifacts are linked when available.
Google Scholar: 0 citations · h-index 0
An exact-scored evaluation of anti-sycophancy prompt steering that separates useful corrections (Update) from harmful reversals (WrongFlip) across three open-weight models, finding model-dependent trade-offs and no universally selective prompt.
A paired counterfactual benchmark that audits whether nine frozen vision–language models change their spinal-radiology reports when apparent age and sex are edited while the target pathology is preserved.
A benchmark and simulated electronic-health-record environment that measures whether medical LLM agents preserve documentation integrity when institutional context rewards shortcuts or omissions.
Ahmed Taha on ORCID · Ahmed Taha on Google Scholar · Ahmed Taha on GitHub · Ahmed Taha on Hugging Face