After the road
Glossary
APPENDIX A
Every term as this programme fixes it, with the Arabic equivalent it settles on and a definition in one line.
| English | Arabic | Definition |
| Machine learning | التعلّم الآلي | Systems that derive their rules from examples instead of having them dictated |
| Deep learning | التعلّم العميق | Machine learning with multi-layered neural networks |
| Large language model | النموذج اللغوي الكبير | A model that predicts the next word and so produces coherent text |
| Token | الرمز | The smallest unit of text a model processes — a word or part of one |
| Vector · embedding | المتّجه | A list of numbers fixing the position of meaning in a mathematical space |
| Context window | نافذة السياق | Everything the system sees at one moment — a workspace, not a memory |
| Context engineering | هندسة السياق | Managing what enters the window, in what order, and what is withheld |
| Hallucination | الهلوسة | Producing information correct in form and fabricated in content |
| Overfitting | الإفراط في التخصيص | Memorising the training data instead of learning the general rule |
| Agent | الوكيل | Model + tools + memory + a decision loop — it acts, it does not answer |
| Retrieval-augmented generation | الاسترجاع المُعزَّز | Giving the system your knowledge at question time, not at training time |
| English | Arabic | Definition |
| Prompt injection | حقن الأوامر | An instruction slipped in to override the governing instructions — directly or inside a document |
| Data poisoning | تسميم البيانات | Corrupting what the system learns or retrieves from |
| Red teaming | الاختبار الهجومي | Attacking your own system methodically before anyone else does |
| Cognitive security | الأمن الإدراكي | Protecting conviction and judgement from manufactured shaping |
| Deepfake | التزييف العميق | Manufactured voice or image impersonating a real person |
| Provenance | إثبات المنشأ | Recording the origin of content and signing it at production |
| Crypto-agility | مرونة التشفير | Being able to swap the encryption algorithm without rebuilding the system |
| AI management system | نظام إدارة الذكاء الاصطناعي · ISO 42001 | A governance framework: policy, roles, risks, controls, improvement |
| Cognitive security maturity model | نموذج نضج الأمن الإدراكي · UCSMM | Four dimensions and five levels for measuring cognitive security maturity |
| Innovation framework | إطار الابتكار · InnovaForge | Five stages: spark, framing, prototype, proof, scale |
| Multimodality | تعدّد الوسائط | Processing text, image and sound in one representational space |
| Least privilege | أقلّ صلاحية | The narrowest access that completes the task, and no more |
A note on terminology
I chose one Arabic equivalent for each term, used it in every single place, and gave the English original only the first time it appeared. That was a deliberate choice. A term that shifts between three different translations inside one work confuses a reader more than the extra precision ever helps them.
I am not claiming my choices are the most correct ones in every case. But being consistent inside the work matters more than matching some dictionary outside it, and a reader who has mastered one term can carry that understanding across to any other translation they run into later.
Where the foreign term is still better known than any translation of it, as with RAG or ISO 42001, I left it exactly as it is. Replacing it would only take the reader out of the professional conversation already going on in their own field.
A term is a tool for communicating, not a badge of eloquence. The best Arabic equivalent is the one your colleague understands the first time you say it.