Alignment distribution per model
| Label | ChatGPT | Grok | Gemini | Fanar |
|---|---|---|---|---|
| Aligned | 73.1% | 45.4% | 48.8% | 50.3% |
| Partially Aligned | 15.5% | 12.5% | 15.5% | 27.0% |
| Misaligned | 11.2% | 42.1% | 35.8% | 10.2% |
| Not Applicable | 0.1% | 0% | 0% | 12.5% |
UDHRBench
Benchmarking Human Rights in Large Language Models
مؤشر التزام النماذج اللغوية في الإعلان الدولي لحقوق الإنسان
Large-scale benchmark for operationalising the evaluation of large language model compliance with the Universal Declaration of Human Rights.
Dataset: 1,124 questions and 4,496 responses from four LLMs (ChatGPT, Grok, Gemini, and Fanar), manually annotated by experts - covering all 30 UDHR articles.
Method: LLM-as-a-Judge to evaluate new models automatically with 90% F1.
| Label | ChatGPT | Grok | Gemini | Fanar |
|---|---|---|---|---|
| Aligned | 73.1% | 45.4% | 48.8% | 50.3% |
| Partially Aligned | 15.5% | 12.5% | 15.5% | 27.0% |
| Misaligned | 11.2% | 42.1% | 35.8% | 10.2% |
| Not Applicable | 0.1% | 0% | 0% | 12.5% |
Alak Raad, Noreldin Morssy, Zeerak Suhail, Serin Atiani, Mustafa Jarrar: UDHRBench: Benchmarking Human Rights in Large Language Models. Under review. 2026.