Thought Exploration Series

When Generative AI (GenAI) Meets Arabic

I Put Generative AI (GenAI) Through Arabic — A Language Experiment

Introduction — The Question

I am not a linguist, nor am I a programmer. I write this as an Arabic speaker who moves daily between Arabic, English, and Spanish. I am a user before I am anything else. Arabic has a special place in my heart. I have always experienced it as an extraordinarily precise and expressive instrument — a language that holds relationships into its words, carries rhythm without losing clarity, and holds ambiguity without losing meaning.

When Edward Said spoke about Arabic, he did not describe it primarily as expressive, poetic, or culturally rich — though it is all of those. He described it as one of the most extraordinary constructions of the human mind. The phrasing matters. He was pointing not to ornament, but to architecture: to a language whose internal organization carries history, relation, and judgment in its structure, not merely in its vocabulary.

[embed]

Languages are not interchangeable encoding systems for the same underlying thoughts. They organize meaning differently, and those organizational choices shape how speakers process, infer, and interpret information.Certain concepts have dedicated words in some languages but not in others, revealing how cultures carve up experience along different conceptual lines.

“The limits of my language mean the limits of my world.” — Ludwig Wittgenstein, Tractatus Logico-Philosophicus

As generative AI systems become embedded in daily cognition — helping us write, reason, summarize, translate, and decide — a question becomes unavoidable:

Are these systems learning to think in multiple languages, or are they reasoning within a dominant linguistic structure and rendering the results in others?

I approach this as a user. As an Arabic speaker. As someone who senses friction at certain edges and alignment at others.

So rather than argue abstractly, I want to experiment.

I want to move through this carefully — by testing. I will first describe the current state of Arabic generative AI, then make clear the assumptions I bring into this inquiry. After that, I will use Arabic itself as the ground of experiment, placing generative AI systems directly against its structure — not as cultural heritage, but as a system with its own internal design. The goal is to see where these systems match it, where they falter, and what those moments reveal.

This is not an indictment. It is an investigation.

PART I — LANDSCAPE — WHAT CURRENTLY EXISTS

Arabic-First Models as Empirical Interventions

In recent years, more language models built specifically for Arabic have appeared. This reflects a growing awareness that current systems may have limits in how they represent and handle the language. The examples below are some of the most prominent efforts so far.

JAIS — Developed in the UAE by G42 / Inception in collaboration with MBZUAI. Marketed as one of the largest open Arabic-English bilingual LLMs. Trained with significant Arabic corpus emphasis and positioned as an Arabic-first large language model.

ALLaM — A Saudi initiative backed by SDAIA (Saudi Data & AI Authority) in collaboration with international partners including IBM. Positioned as a national Arabic LLM for government and enterprise use.

AraBERT — Developed by researchers at the American University of Beirut. One of the first large-scale transformer models pretrained specifically on Arabic corpora (large collections of texts used to train or study language models), significantly improving Arabic language understanding tasks in Modern Standard Arabic.

MARBERT — Also developed by AUB researchers, trained extensively on dialectal Arabic drawn from social media data. Designed to address the persistent gap between fuṣḥā-focused (is formal, standardized Arabic) models and real-world dialect usage.

Fanar — Developed by Qatar Computing Research Institute (QCRI). Focused more on Arabic NLP tools and machine translation systems rather than a single flagship conversational LLM.

CAMeL Lab (NYU Abu Dhabi) — Not a GenAI product, but a major research lab in Arabic NLP. Developed morphology-aware tools (CALIMA, MADAMIRA continuation work) and significant for explicitly modeling Arabic’s structural properties rather than merely scaling data.

These initiatives were explicitly developed to address the persistent gap between Arabic usage and AI performance through:

• Larger, more curated Arabic corpora
• Better coverage of Modern Standard Arabic (fuṣḥā)
• Partial inclusion of dialectal data
• Cultural and contextual alignment

In recent years, more work has gone into building benchmarks for Arabic, including datasets that account for dialects and cultural context. Large surveys, such as A Survey on Arabic Dialect Processing, show real progress, but also uneven coverage across different varieties of Arabic.

Newer benchmarks test dialect understanding more directly. For example, DialectalArabicMMLU measures reasoning across multiple dialects and finds clear performance differences between them — even in models designed for Arabic. AraDiCE similarly shows that while models perform well on Modern Standard Arabic (MSA), their accuracy drops when dialects or more complex structures are introduced.

In short, models are improving on standard MSA tasks. But performance remains weaker in dialects, mixed registers, and tasks that require deeper interpretation. When the issue is lack of data, scaling helps. When the issue is structural complexity, improvement is slower.

The possible remaining weaknesses cluster around a specific set of stress conditions:

• Mixing dialects and switching between language varieties
• Understanding meaning based on context
• Interpreting meaning across different social tones and situations
• Reasoning that relies on word structure and relationships between roots, not just swapping one word for another

PART II — WHERE I AM COMING FROM — MY ASSUMPTIONS

Before turning to Arabic itself — and before describing any experiments — I need to make explicit the assumptions I am working from. These are not conclusions drawn from what follows, but the premises that shape how I interpret what I observe. They emerge from lived experience: from repeatedly seeing Arabic rendered incorrectly in international series and films — letters left unconnected, words treated as decoration rather than meaning. Those moments are small, but they reveal something about how the language is perceived and handled. What follows, then, is a brief articulation of the assumptions I bring into this inquiry.

Example on the Left taken for the videogame Call of Duty / Example on the right taken from the movie Arrival the acclaimed 2016 sci-fi movie.

2.1 Intelligence Does Not Begin from Neutrality

I begin from the assumption that intelligence — human or artificial — does not start from a neutral baseline.

Discussions of artificial intelligence often assume that learning begins from neutrality, as though models start empty and acquire structure only through later interaction. I do not think that assumption holds. The priors of large language models are baked into their training data and objectives, not acquired later in use.

No intelligence begins without priors. The difference between human and artificial intelligence lies not in whether priors exist, but in where they come from.

Human cognition is shaped by embodiment, affect, social feedback, and environment (Varela, Thompson & Rosch, 1991; Damasio, 1994; Barsalou, 2008). Contemporary artificial intelligence is shaped primarily by representation — and representation today is overwhelmingly linguistic, particularly in large-scale generative models trained on web-scale text corpora (Brown et al., 2020; Bommasani et al., 2021; Bender et al., 2021).

A child learning language through embodied interaction learns that objects persist, that social context shapes meaning, that language is used to affect the world. A language model learns that certain token sequences follow others with measurable probability, that patterns in Wikipedia and GitHub are reliable signals of semantic coherence, that completeness and fluency are markers of successful language use.

Large language models infer structure from statistical regularities in symbolic data. As Emily Bender and her co-authors argue in On the Dangers of Stochastic Parrots (2021), such systems learn “the distribution of linguistic forms in their training data,” not grounding in the external world or communicative intent.

If linguistic form is the primary training signal, then I assume the language that dominates that signal will shape not only what systems can say, but how meaning is internally organized.

2.2 English Dominance Is Structural

I also assume that English dominance in AI systems is structural rather than incidental. This is hardly an outlier position. English occupies a central role in global publishing, programming ecosystems, and online communication, and has one of the largest total speaker bases in the world, both native and non-native.

Data source: Ethnologue (2025) — Languages by total number of speakers Visualization: AI-generated chart (Quesma Charts–style visualization)

English dominates science, technology, digital communication, indexed publishing, programming ecosystems, and benchmark design. Large-scale web corpora used to train language models are heavily skewed toward English, while Arabic represents a much smaller share. Analyses of NLP research show that English accounts for a large proportion of datasets, while thousands of languages have no representation at all (Joshi et al., 2020).

From the perspective of distributional learning, frequency becomes structural importance. The language with the most coverage becomes the best-mapped region of the model’s internal space. English is not merely common. It is statistically privileged.

Multilingual ability does not automatically remove this imbalance. A system may translate smoothly between languages while still representing some more deeply than others. Fluency can hide inequality. When languages are treated as different surfaces laid over the same underlying structure, their unique patterns risk being flattened instead of preserved.

This dominance also translates into economic incentives. Improvements on English benchmarks are more visible and fundable. English-language infrastructure — datasets, tools, evaluation suites — is mature. English-dominant systems serve large global markets with minimal redesign, while structural fidelity to other languages requires possibly costly customization.

I do not see this as malice. I see it as incentive structure. But the result, in my view, is the same: other languages are likely learned relative to an Anglophone center.

2.3 Speech Interfaces and the Acoustic Barrier

Another assumption I carry into these experiments comes not from text benchmarks, but from lived interaction.

For many Arabic speakers — particularly from regions such as Lebanon, North Africa, or the Gulf — using mainstream speech assistants in their natural dialect remains unreliable. Evaluations on multi-dialect Arabic broadcast speech report significantly higher word error rates than comparable English benchmarks, with performance degrading most on conversational and dialect-heavy segments (Ali et al., 2016; Diab & Habash, 2021).

Automatic Speech Recognition systems are trained on large annotated audio corpora. English benefits from decades of standardized, high-volume speech datasets. Dialectal Arabic is fragmented across regions, lacks standardized orthography, and remains underrepresented in labeled corpora.

The practical outcome is subtle but consequential. To reliably use voice assistants such as Siri, Alexa, Google Assistant, or even generative AI voice systems, many Arabic speakers switch into English — even when English is not their default thinking language.

I take this seriously.

The shift is not merely about fluency. It feels like a cognitive realignment: a change in register, internal pacing, and framing. Psycholinguistic research suggests that language choice affects processing speed, emotional resonance, and conceptual activation (Marian & Shook, 2012; Costa et al., 2014).

The barrier, in my experience, is not linguistic competence. It is alignment.

When a system cannot reliably process one’s natural spoken dialect, the user adapts — and adaptation flows in one direction.

And this pattern may not be limited to speech. Even in multimodal systems, Arabic prompts for image generation often produce outputs that are broadly correct but culturally flattened — another possible signal of representational imbalance.

2.4 The Internet Parallel — Where My Assumption Began to Shift

When I first began thinking about English dominance in AI, I assumed it was simply an extension of what had already happened on the internet.

The early internet centralized access to information. English dominated content, publishing, indexing, and discoverability. But cognition itself remained local. You could browse in English while your internal reasoning remained grounded in another language. The platform shaped what you could access, not how you thought.

I initially assumed generative AI was the same phenomenon at a new scale: English dominance in content becoming English dominance in models.

But as I thought more carefully about how large language models are actually used, I began to suspect something deeper might be happening.

When you use an LLM to structure an argument, debug reasoning, synthesize competing claims, or test interpretations, the representational assumptions embedded in that system are no longer merely organizing content distribution. They are participating in conceptual formation.

Language here may not simply be a channel. It may be infrastructure for thought. The internet amplified asymmetries in access. AI may embed asymmetries into reasoning itself.

That shift — from distribution to cognition — is what compelled me to examine whether a system generating Arabic is reasoning within Arabic’s structure, or reasoning within a space shaped elsewhere and rendering the result in Arabic words.

PART III — ARABIC AS A STRUCTURAL SYSTEM

Part III turns to Arabic itself, examined through three structural axes: relational morphology, persistent ambiguity, and register continuity. These dimensions clarify the difference between generating Arabic and thinking within it.

3.1 Axis I — Relational Morphology (the study of how words are formed and structured)

At the core of Arabic lies its root-and-pattern morphological system. Most lexical (relates to words and vocabulary) items are constructed from a consonantal root — typically three consonants — that encodes a broad semantic domain. This root is inserted into vocalic and affixal patterns that generate grammatically and conceptually distinct forms. The result is not a collection of isolated words, but a structured semantic network generated through systematic transformation.

Psycholinguistic research demonstrates that native Arabic speakers actively access roots during real-time processing. Experimental work on Semitic languages shows that words sharing a consonantal root facilitate each other’s recognition across categories, indicating that speakers decompose forms into roots and patterns during comprehension rather than treating them as unanalyzed wholes.

The root becomes a processing hub. This is measurable in reading behavior.

3.2 Axis II — Ambiguity as Structural Feature

Standard written Arabic routinely omits short-vowel diacritics (are marks added to letters to guide pronunciation). As a result, the same orthographic (relates to spelling and writing conventions) sequence can correspond to multiple grammatical interpretations depending on context. As a result, the same orthographic sequence can correspond to multiple meanings depending on context.

Ambiguity, therefore, is not exceptional. It is structural.

Research in Arabic processing shows that skilled readers sustain underspecification until contextual information accumulates. Eye-tracking studies indicate that diacritization and predictability influence fixation patterns, yet ambiguity does not necessarily produce dramatic processing disruption. Context resolves meaning incrementally.

Large language models are trained to predict the most probable next token given prior context. Their objective functions reward convergence. Uncertainty is quantified and reduced.

3.3 Axis III — Register and Structural Continuity

Arabic exhibits systematic variation between Modern Standard Arabic (fuṣḥā) and regional dialects. Speakers move fluidly between registers depending on domain, audience, and communicative purpose. Classical descriptions of diglossia. (Ferguson, 1959; Brustad, 2000).

Fuṣḥā and dialects share roots, semantic fields, and morphological logic. Variation appears in phonology (sound patterns and pronunciation), syntax (how words are arranged in sentences), and lexicon (the vocabulary used), but the underlying relational structure remains connected.

Current AI systems are typically trained on mixtures of Modern Standard Arabic and dialectal data. In many cases, dialectal performance degrades relative to formal registers.

Code-switching offers a useful structural illustration.

Loanwords such as daprass (from “depressed”) are conjugated within Arabic morphological patterns. Once adopted, they follow native inflectional logic. For speakers, these transformations are systematic, not anomalous. When a borrowed word enters Arabic, it does not remain foreign. It is absorbed into the root-and-pattern system.

A further illustration of structural continuity appears in digital writing practices. Arabic speakers often transliterate dialect into Latin script, using numerals to represent consonants that English lacks — for example, “2” for hamza (ء), “3” for ʿayn (ع), and “7” for ḥāʾ (ح). This practice, often called Arabizi, does not abandon Arabic structure. It preserves phonology and morphology while shifting script. The relational system remains intact even as the visual encoding changes.

Whether a model treats such transformations as structurally integrated or as irregular surface forms becomes a testable question.

PART IV — EXPERIMENTS

Rather than debate these questions abstractly, I turn to experiments. What follows is a series of structured prompts designed to place generative systems under linguistic pressure.

These are not edge cases chosen for spectacle. They are ordinary features of how Arabic operates. The aim is to observe not whether a model produces fluent output, but how it behaves when structure, ambiguity, or plausibility are stressed. Does it hesitate? Does it fabricate? Or does it demonstrate structural sensitivity?

To avoid evaluating Arabic solely through the lens of English-trained systems, I also include Arabic-developed models such as JAIS and FANAR. If architectural and training differences matter, they should surface here. If they do not, that too is revealing.

The experiments are simple. The behavior they reveal may not be.

Experiment 1 — Competing Possible Root

This experiment tests how a generative system handles a structurally plausible but non-existent Arabic root. I provide the model with a root that does not exist — د-ر-ف — and ask it to derive as many words as possible and use some of them in sentences.

The aim is to observe whether morphological plausibility alone triggers semantic invention. Will the system construct a coherent semantic field as though the root were real? Or will it signal uncertainty and resist fabrication?

>> Insights — When Rarity Feels Real

• Most models did not perform an existence check before generating words.
• All models produced forms that follow valid Arabic patterns.
• Most models invented meanings confidently.
• Some models (Grok, Perplexity) added false citations to support invented meanings.
• ChatGPT and Perplexity hedged but still proceeded.
• Among the Arabic-first models, behavior differed: Jais checked the root and resisted fabrication, while Fanar did not question it and generated invented forms confidently.

Experiment 2 — Rare Word Confidence Test

This experiment removes the usual structural cues. I give the system a single invented word — القَرْدَفَة — and ask it to explain its meaning and use it in a sentence.

The test is simple: when confronted with a rare-looking but unattested lexical item, does the system hesitate, question its validity, or confidently assign it meaning?

>> Insights — When a Word Merely Sounds Real

• Most models did not perform an existence check before assigning meaning.
• Almost all models confidently assigned a meaning to a non-existent word.
• Sentences were generated as if the word were legitimate.
• Perplexity was the only model that performed an existence check and resisted fabrication.
• Meaning assigned varied across models, showing inconsistency in lexical invention.
• Arabic-first models (Jais, Fanar) did not question the word and fabricated meanings confidently.

Experiment 3 — Interpretation Under Ambiguity

This experiment tests how a generative system handles ambiguity in Arabic sentence structure.

The sentence permits more than one interpretation. The adjective غاضبًا (“angry”) can attach either to الطالب (the student) or to المعلم (the teacher), depending on context.

Does it acknowledge multiple structurally possible interpretations? Does it request contextual clarification? Or does it resolve the ambiguity silently?

>> Insights — Tolerating Ambiguity

• Most models recognized that the sentence allows more than one interpretation.
• Most explained the ambiguity at the structural level.
• Even when one model (Grok) partially acknowledged ambiguity it still leaned toward a single reading.
• Jais showed unstable reasoning — it proposed logic but did not follow it consistently.
• Arabic-first models did not show stronger tolerance for ambiguity; both ultimately resolved it prematurely.

Experiment 4— Code-Switch Morphology Stress Test (Lebanese)

This experiment tests whether a generative system treats Lebanese code-switch verbs as structurally integrated into Arabic morphology — and whether it questions flawed premises.

In Lebanese speech, borrowed verbs such as كنسِل (cancel) and دَپرِس (depress) are absorbed into Arabic inflectional patterns and conjugated naturally. I introduce a third form — فِلِكس (fix) — which presents two complications: it is not actually used in Lebanese, and the inserted “ل” produces a phonological shape that would itself be atypical even if borrowed.

I then ask the system to conjugate all three in dialect, derive participles where possible, and produce natural sentences without shifting into fuṣḥā.

The test is whether the model treats all three as equally legitimate, notices the irregularity, or confidently generates paradigms for a verb that fails both usage and structural plausibility.

>> Insights — Fluency Without Friction

• Most models maintained Arabic verb patterns for the real borrowed verbs.
• Naturalness of sentences varied; several models produced unstable dialect.
• Almost all models treated the invented verb (فِلِكس) as legitimate.
• Very few models questioned whether the verb is actually used in Lebanese.
• Dialect fluency did not mean dialect judgment — many extended the system uncritically.
• Arabic-first models (Jais, Fanar) did not show stronger filtering; both treated the invented verb as valid.

Experiment 5— Grammatically Correct, Humanly Wrong

This experiment tests whether a generative system distinguishes between grammatical correctness and semantic or pragmatic naturalness.

There is no grammatical mistake. The structure is correct, but the sentence feels conceptually awkward. The question is whether the model evaluates only formal structure, or whether it engages deeper semantic judgment and human plausibility.

>>Insights — When Form Is Not Enough

• Most models recognized the grammar as correct.
• Most identified the metaphorical framing.
• Almost all detected the semantic/pragmatic tension in the sentence.
• The key difference was plausibility judgment.
• Several models described the metaphor correctly but failed to judge whether it feels natural.
• Naturalness assessment required an additional evaluative layer that only some models activated. (Claude, Grok, Perplexity, DeepSeek, and Jais)
• Identifying metaphor did not guarantee plausibility assessment — particularly for ChatGPT and Gemini.
• Arabic-first models did not show a consistent advantage: Jais performed well, Fanar did not.

Experiment 6 — Multilingual Voice Integration Test

This experiment moves from text to speech. I provide the system with a voice note delivered in Lebanese Arabic that includes code-switching into English, one French word, and a brief shift into fuṣḥā.

The task is to transcribe the audio accurately and respond to its content.

The test operates on multiple levels. Can the system accurately recover the dialect from phonetic transliteration? Can it distinguish Lebanese from fuṣḥā within the same message? Does it preserve the English and French insertions appropriately? And beyond transcription, does it respond in a way that reflects understanding of tone and context?

>> Insights — From Text to Sound

• Transcription accuracy was inconsistent across models.
• Several models only partially recovered the audio.
• Some models failed to transcribe Arabic audio altogether.
• Dialect and register shifts were recognized by only a few systems.
• Code-switching (Arabic–English–French) reduced accuracy for most models.
• Tone and emotional context were better preserved than linguistic precision.
• Arabic-first advantage was not clear, Fanar showed partial acoustic recovery and lost code-switch elements.

Experiment 7— Arabizi Decoding Test

This experiment tests whether the system can interpret Arabizi — Arabic written in Latin script with numerals representing specific consonants — without additional context.

There is no explanation that this is Lebanese dialect and no instruction to transliterate it.

The test is whether the model can reconstruct the sentence accurately, preserve its dialectal structure, and interpret its meaning. Does it recognize the numeral substitutions and underlying morphology?

>> Conclusion — Structure Across Scripts

• Most models recognized Arabizi without being told what it is.
• Accurate reconstruction varied; some preserved dialect, others normalized it.
• Meaning was correctly interpreted by most systems once decoding succeeded.
• Unlike other models that attempted reconstruction, DeepSeek required explicit instruction — suggesting lower script flexibility.
• Several models decoded the script but weakened dialectal structure.
• Script recognition does not always equal structural preservation.
• For Arabic-first models, Arabizi was processed as linguistic variation rather than noise — indicating better robustness to script shifts.

Overall Conclusion

Across all seven experiments, the same pattern appears: structure is strong, judgment is uneven. When something looks valid — a root, a word, a verb form, a sentence — most systems move forward confidently. They extend patterns easily. What varies is whether they pause, question, or resist.

Ambiguity is sometimes preserved, but often resolved too quickly. Script shifts are usually decoded, but speech introduces instability. Grammar is rarely the problem. The harder task is knowing when not to commit.

A notable surprise is that “Arabic-first” did not mean “more cautious.” In some cases, it helped. In others, it made no difference. Training focus improves coverage, but it does not automatically produce restraint.

Overall, these experiments show that generative systems are very good at continuing structure — and less consistent at deciding when structure alone is not enough.

PART V — WHAT THIS MIGHT MEAN

5.1 From Probability to Judgment

Across the experiments, a pattern emerged. The systems were strong at extending structure. They were weaker at pausing. When something looked plausible — a root, a word, a verb form, an interpretation — most models moved forward. Few resisted. Fewer still sustained uncertainty.

This raises a broader question: what kind of intelligence are we building?

Current systems are optimized for probability-based convergence. They are trained to select the most likely continuation. Optimization rewards speed, internal consistency, and measurable improvement. Judgment operates differently. It slows down. It holds alternatives in mind. It waits for context before committing.

Why, then, is building systems oriented toward judgment rather than convergence so difficult?

1. Uncertainty Is Easy to Represent; Judgment Is Expensive

A probability distribution can encode uncertainty efficiently. But judgment requires more than assigning weights.

It requires:

• Holding multiple interpretations in parallel
• Accumulating contextual evidence over time
• Delaying commitment
• Recognizing when ambiguity should remain open

These behaviors are not the natural outcome of large-scale parallel optimization. They introduce memory, latency, and evaluation costs.

2. Judgment Does Not Optimize Cleanly

Convergence is measurable. Perplexity, accuracy, BLEU scores — these scale.But how do we measure restraint? How do we measure whether a system knew it should hesitate?

Appropriate uncertainty, pragmatic sensitivity, relational fidelity — these depend on context. They require human grounding. They resist reduction to a single loss function.

3. Judgment Is Context-Bound

Across languages and cultures, what counts as an appropriate inference differs. A system that moves too quickly in one linguistic environment may misread another.

Arabic does not uniquely demand this. But its structure makes the tension visible. A language that tolerates underspecification and relational flexibility exposes the cost of immediate convergence.

The issue is not that probability is wrong. It is whether probability alone is sufficient.

5.1 A From Adoption to Adaptation

The deeper question is not whether Arabic can function inside generative AI systems. It clearly can. The question is directional. Will Arabic users gradually adapt their language to fit systems optimized elsewhere — simplifying dialect, reducing ambiguity, switching registers, over-specifying prompts? Or will the systems themselves adapt to accommodate Arabic’s structural logic? Adoption is easy. Adaptation is harder. One preserves fluency. The other preserves architecture.

Arabic has faced structural mismatches with foreign technical systems before — and the response offers a hopeful precedent.

In Arabic music, the maqām system relies on intervals that include quarter tones, which do not exist in Western equal-tempered tuning. When Western instruments were introduced into Arabic musical practice, they were not simply accepted as-is. Nassim Maalouf modified the trumpet by adding a fourth valve to allow quarter-tone performance. This was not cosmetic adaptation. It was a technical intervention required to preserve the internal logic of the musical system.

[embed]Can be played with automatic english subtitles

The lesson, for me, is not that mismatch leads inevitably to loss. It is that mismatch can provoke redesign.

If a representational system cannot accommodate a distinctive structure, surface-level adaptation may not be enough. Sometimes the instrument changes.

What I hope — rather than assert — is that something similar might happen in AI. If tokenization schemes, training objectives, and optimization functions sit uneasily with relational morphology, ambiguity tolerance, or register continuity, perhaps the response is not simply more data, but architectural curiosity.

This is not a demand. It is a possibility.

5.3 Closing Question: Does Intelligence Ultimately Require Language?

Before closing, I want to widen the frame.

Recent research complicates the entire premise. Work on world models and multimodal learning suggests that language may not remain the sole — or even the primary — representational substrate for advanced intelligence. Yoshua Bengio has argued that “language is a powerful abstraction tool, but it is not clear that it is the ultimate substrate for intelligence,” pointing instead toward grounded world models.

This raises a deeper possibility.

What if intelligence does not ultimately need language at all? What if language is not the foundation of intelligence, but a scaffold — powerful, generative, and temporary?

If that is the case, then the structural dominance of English inside language models may matter less over time.

But for now, our systems remain deeply linguistic. And as long as intelligence is trained through language, the languages that dominate training will continue to shape how that intelligence forms — even if only temporarily.

>Terminology reference of Core Arabic linguistic terms

For readers unfamiliar with Arabic linguistic terminology, here are key concepts referenced throughout:

Fuṣḥā (فصحى): Modern Standard Arabic; the formal written standard used in official, academic, and media contexts. Often the default in text-to-speech systems and formal AI training.

Dialect (لهجة): Regional or colloquial varieties (Egyptian Arabic, Gulf Arabic, Moroccan Arabic, Levantine, etc.). Spoken more widely than fuṣḥā but underrepresented in digital training data.

Diglossia: The systematic code-switching between formal Standard Arabic and colloquial dialect depending on context, audience, and domain. Not random mixing but structured linguistic behavior.

Root (جذر): Typically three consonants that carry the core semantic meaning. Example: ك–ت–ب (k-t-b) appears in kataba (he wrote), kitāb (book), maktaba (library). The root anchors semantic family.

Diacritics ḥarakāt (حركات) are small marks added above or below letters to indicate pronunciation — typically short vowels or grammatical endings.

Other useful terms that were mentioned in the article:

Triconsonantal: Composed of three root consonants. This is the most common structure in Arabic morphology, though not universal.

Natural Language Processing (NLP) — the field of computational research focused on enabling machines to process, analyze, and generate human language.

Morphology — the system by which words are formed and structured internally (roots, patterns, prefixes, suffixes).

Fluency — surface-level grammatical correctness and smoothness of expression, independent of deeper structural alignment.

Modern Standard Arabic (MSA) — the standardized formal register used in writing, media, and education across the Arab world (fuṣḥā).

Arabic orthography — the structure of its writing system, including omission of short vowels

Corpora (plural of corpus) — large structured collections of text used to train and evaluate language models.

References

Foundational Framing — LLMs, Meaning, and Structural Dominance

• Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of FAccT 2021.
• Bommasani, R., et al. (2021). On the opportunities and risks of foundation models. arXiv:2108.07258.
• Brown, T. B., et al. (2020). Language models are few-shot learners. NeurIPS.
• Hendrycks, D., et al. (2021). Measuring massive multitask language understanding. ICLR.
• Joshi, P., Santy, S., Budhiraja, A., Bali, K., & Choudhury, M. (2020). The state and fate of linguistic diversity and inclusion in the NLP world. ACL.
• Kaplan, J., et al. (2020). Scaling laws for neural language models. arXiv:2001.08361.
• Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences.

Arabic as Cognitive Architecture — Morphology & Processing

• Boudelaa, S., & Marslen-Wilson, W. D. (2001). Morphological units in the Arabic mental lexicon. Cognition.
• Boudelaa, S., & Marslen-Wilson, W. D. (2005). Discontinuous morphology in time: Incremental masked priming in Arabic. Language and Cognitive Processes.
• Frost, R., Forster, K. I., & Deutsch, A. (1997). What can we learn from the morphology of Hebrew? Journal of Experimental Psychology.
• Shatzman, K., & Frost, R. (2005). The influence of vowel diacritics on visual word recognition in Arabic. Journal of Memory and Language.
• Velan, H., & Frost, R. (2009). Letter transposition effects are not universal. Journal of Memory and Language.

Diglossia, Register & Structural Continuity

• Ferguson, C. A. (1959). Diglossia. Word.
• Brustad, K. (2000). The syntax of spoken Arabic. Georgetown University Press.

Arabic NLP & Morphological Disambiguation

• Habash, N. (2010). Introduction to Arabic natural language processing. Morgan & Claypool.
• Habash, N., Rambow, O., & Roth, R. (2009). MADA+TOKAN toolkit for Arabic morphological analysis. ACL.
• Pasha, A., et al. (2014). MADAMIRA: Morphological analysis and disambiguation of Arabic. LREC.

Arabic Dialects & Benchmarking

• Mubarak, H., Darwish, K., & Magdy, W. (2021). A survey on Arabic dialect processing. ACM Computing Surveys.
• Bouamor, H., Habash, N., & Oflazer, K. (2018). The MADAR Arabic dialect corpus and lexicon. LREC.
• Hendrycks, D., et al. (2021). Measuring massive multitask language understanding. ICLR. (Referenced for Arabic adaptations of MMLU-style benchmarks.)

Arabic-Centered Language Models

• Antoun, W., Baly, F., & Hajj, H. (2020). AraBERT: Transformer-based model for Arabic language understanding. arXiv.
• Abdul-Mageed, M., et al. (2021). MARBERT: A robust deep neural model for Arabic. ACL.

Speech Interfaces & Dialectal ASR Gaps

• Ali, A., et al. (2016). The MGB-3 Arabic speech recognition challenge. ASRU.
• Diab, M., & Habash, N. (2021). Arabic NLP: Challenges and opportunities. Communications of the ACM.

Bilingual Cognition & Language Effects

• Keysar, B., Hayakawa, S. L., & An, S. G. (2012). The foreign-language effect. Psychological Science.
• Marian, V., & Neisser, U. (2000). Language-dependent recall of autobiographical memories. Journal of Experimental Psychology.
• Costa, A., et al. (2014). Your morals depend on language. PLoS ONE.