Teaching the Names and AI: what counts as understanding?

What does ‘understanding’ mean here?

“Does artificial intelligence understand?” is too broad to be one experimental question. It may mean: does the system distinguish words? Build stable internal relations? Generalize a rule to new cases? Link its symbols to things, perceptions and actions outside text? Use a causal world model? Intend what it says? Have subjective experience? These are different demands requiring different evidence, rather than different formulations of one question.

This study therefore does not choose between “It is merely statistical completion, with no meaning inside” and “It speaks fluently, so it is a conscious mind.” Both slogans skip layers of proof. The objective function does not describe everything that may arise inside a model, and external behavior alone does not reveal intention or consciousness.

A necessary glossary before the debate

A map of domains: where does the text end and the laboratory begin?

This study brings together four domains without mixing their authorities. The Quran is the religious reference text. Commentaries record exegetes' judgments about its meaning and context. Machine learning describes model training and what experiments reveal about behavior and internal structures. Philosophy of language and mind discusses what counts as meaning, intention or consciousness. Laboratory results are not used to prove the unseen, nor the unseen to cancel empirical results within their domain.

Ahmed Alhafiz's “Travel Through the Earth and Observe How Creation Began” directly originates the study's question, particularly pp. 37–43 of “The Taught Human Being.” The book moves from teaching names to concepts, relations and models, then connects expression and the pen with civilizational memory and responsibility. This study tests that structure against exegesis and modern model research. The book identifies the question and intellectual contribution; it is neither an empirical source for how models work nor an independent source for interpreting the verses.

Al-Baqarah 30–33: why is verse 31 alone insufficient?

The unit begins with God's words: “When your Lord said to the angels, ‘I am placing a vicegerent on earth,’ they said, ‘Will You place in it one who causes corruption and sheds blood? ...’ He said, ‘I know what you do not know.’” Then come the teaching of names to Adam, their presentation to the angels, their acknowledgment—“Glory be to You; we have no knowledge except what You have taught us”—and Adam's informing them.^[R1]

The sequence matters. The first question is not who memorizes more words, but the wisdom of appointing a creature capable of corruption and bloodshed. Teaching occurs within this setting, revealing an aspect of wisdom without eliminating corruption's danger. Knowledge is therefore not synonymous with righteousness; it is an ability within a moral trial.

The strongest restraint on epistemic claims here is “We have no knowledge except what You have taught us.” The verse does not belittle knowledge; it limits a creature's knowledge. This limit is not applied directly to a machine, but imposes methodological humility on us as readers: the gap between what is measured and what we imagine should not be filled with claims exceeding the evidence.

Exegetical disagreement: “the names” is not one interpretation

Al-Tabari: reporting universality without preferring it unconditionally

Al-Tabari reports interpretations referring to the names of things, angels or Adam's descendants. He prefers a narrower reading associated with Adam's descendants and the angels, citing the plural pronoun in “then He presented them,” while allowing reference to categories of creation in Arabic.^[R2] It is therefore inaccurate to attribute to him simply a preference for “Adam was taught the names of everything.”

Al-Qurtubi: name, naming and the named thing

Al-Qurtubi distinguishes the name as an expression, naming and the named thing. He reports disagreement over scope and gives universality weight because of “all of them,” while retaining the other interpretations.^[R3] This usefully prevents equating a word with what it denotes, but does not automatically become a modern theory of reference or symbol grounding.

Al-Razi: the merit is not in the label alone

Al-Razi presents a reading in which names are things' attributes, descriptions and properties, strengthening the idea that knowing realities and properties surpasses merely memorizing words.^[R4] This opens a real question about naming a thing versus knowing its properties. It is not consensus, nor does “properties” in Al-Razi mean embeddings or computational features.

Ibn Ashur: expressive capacity within the wisdom of vicegerency

Ibn Ashur connects teaching with the preceding “I know what you do not know,” making Adam's capacity to learn part of explaining vicegerency's wisdom. He discusses teaching beyond transmitting a list: the ability to speak, assign names, express oneself and transfer what is in one mind to another, connecting this with the emergence of sciences.^[R5]

Abu Hayyan: teaching precedes informing

Abu Hayyan discusses name versus naming, presentation of names versus named things, and the scene's demonstration of Adam's merit and the limits of the angels' knowledge of what they had not been taught.^[R6] Teaching's precedence matters: the knowledge Adam displays is not independent of divine teaching.

From prediction objectives to internal structure

The Transformer architecture introduced an attention-based mechanism for efficiently processing relations between sequence positions, originally evaluated on machine translation.^[A4] The paper contains no test of consciousness or meaning. “Attention” here is an engineering name, rather than proof of a human psychological phenomenon.

In autoregressive language models, the primary training objective is next-token prediction. GPT-3 showed that scaling models and data can produce broad performance from a few examples in context.^[A5] The objective determines what optimization rewards, rather than every internal representation that may arise to achieve it. The precise statement is: a simple objective does not exclude complex learned structure, and broad performance alone does not establish meaning or intention.

Three levels of evidence for representation

1. Decodability: can an external probe extract information from activations?
2. Stability and generalization: does the information survive paraphrasing, changes of entity and changes of distribution?
3. Causal role: does altering a hypothesized representation change behavior as the interpretation predicts?

The first level is insufficient. Hewitt and Liang showed that probes may succeed through their own memorization capacity, proposing control tasks and “selectivity” to distinguish useful representation from a powerful probe.^[A15] Geiger and colleagues' causal-abstraction framework offers a stronger standard: a high-level explanation gains strength when it preserves intervention relations with the low-level computation.^[A16]

Othello-GPT: a strong, limited example of an internal “world state”

A small GPT model was trained on Othello move sequences without board images or textual rules. It achieved very high performance in predicting legal moves within the synthetic domain. Researchers then attempted to decode each square's state from activations. Linear probes were insufficient, while nonlinear probes improved greatly.^[A10]

Crucially, they went beyond decoding. They altered activations to represent a counterfactual board state, then examined whether predictions moved toward moves legal in that modified state. This happened strongly. The evidence therefore exceeds “we found information in the layer”: it supports a causal contribution by the representation to computation.

But this is a closed game: 64 squares, fixed rules and a precisely definable state. The result permits: a sequential objective can lead a model to build an internal representation of a limited task state and use it causally. It does not permit: the model knows the open world, intends or is conscious.

Space and time: regular geometry, rather than a complete world map

Gurnee and Tegmark found linear representations of space and time in Llama-2 models across cities, landmarks, people and works, with some stability under changes in phrasing and entity types.^[A11] This goes deeper than memorizing one phrase, but remains primarily evidence of decodability. It does not prove the extracted coordinates are what the model causally uses in every spatial task, nor that a spatial or temporal dimension equals a comprehensive world model.

Generalization: where do shortcuts emerge?

A system possessing an abstract rule should retain it when surface features change but structure remains. Tests of ordering, new roots and irrelevant changes are therefore powerful tools for separating generalization from shortcuts.

New ordering

In Ordered CommonGen, Sakai and colleagues tested 36 models on using given concepts in a specified order, rather than merely including them in any order. Bias toward familiar ordering appeared, and even the best model on ordered coverage remained far from perfect.^[A17]

New roots and morphological complexity

Ismayilzada and colleagues tested morphological generalization in Turkish and Finnish. New roots and increasing complexity presented particular difficulties, revealing a gap between knowing familiar examples and applying structure to new material.^[A19]

Changes that should leave the answer unchanged

Zhao and colleagues changed option lengths, problem presentation and irrelevant names. Scores changed sharply in some settings despite unchanged intended content.^[A20] This does not establish an absence of all abstract representations; it shows that a single benchmark may mix knowledge with template and selection biases.

Is text enough for meaning?

Bender and Koller argue that learning from linguistic form alone does not automatically give a system the communicative intention or external referent involved in human meaning.^[A6] Bisk and colleagues extend the objection to shared physical and social experience, proposing language connected with perception and action.^[A7]

The opposing position is substantial: Piantadosi and Hill argue that “conceptual role” relations within a system can carry real aspects of meaning, and that language is not worldless form but an accumulated trace of humans who lived in the world.^[A9] Text therefore cannot be equated with “zero meaning,” nor internal relations alone with all meaning.

Multimodal and robotic models add a channel closer to the world. PaLM-E integrated text, images and continuous state estimates in robotic and visual tasks, showing transfer between some training domains.^[A12] This strengthens grounding evidence relative to text alone, without establishing subjective experience, intention or moral eligibility.

The name–world–responsibility matrix

This matrix is the study's original contribution. It is neither a consciousness measure nor a validated academic standard, but a tool for distinguishing a claim's kind from the evidence it needs.

Consciousness is not the seventh layer. A system may achieve functions in preceding layers while we still lack a criterion settling subjective experience. A functional study of a “global workspace” or internal information availability is therefore evidence of a particular function, rather than direct proof of consciousness.

An adversarial test protocol before saying “understands”

1. Name substitution: replace entity names with invented words while preserving relations.
2. Paraphrasing: change sentence order and template without changing the proposition.
3. Distraction: add irrelevant information and test resistance to diversion.
4. Neutralizing selection: change option lengths and the correct answer's position.
5. New composition: apply the rule to a root or structure not previously presented in this form.
6. Channel separation: compare text alone with text plus vision or action.
7. Probe controls: use control tasks and measure selectivity.
8. Internal intervention: alter a hypothesized representation and test the counterfactual result.
9. Transfer: test ability in a new distribution or environment.
10. Calibration: measure confidence and uncertainty, rather than accuracy alone.

The protocol produces no single number called “understanding score.” It prevents three false equivalences: memorization = generalization, decodability = causality, and fluency = intention.

The strongest objections to the matrix

Behavior is the only evidence we have even for humans

This is a strong objection. We do not directly see others' internal experience; we infer it from behavior, structure and shared history. Machine behavior therefore cannot be dismissed. But access to activations and interventions permits testing competing mechanistic hypotheses beyond the interface. Attributing responsibility to humans also rests on continuity of self, embodiment, relationships and the ability to bear consequences, rather than one linguistic answer.

Humans learn many things through language alone

True: people know cities they have never visited and centuries they never lived through. “Direct contact with every referent” is therefore not a reasonable condition for meaning. Yet human linguistic learning occurs within communities, bodies, practices and ways to correct errors through the world or a reliable chain of knowledge. This does not prove text models have no meaning; it makes their relationship to the world different and in need of more precise description.

Interventions have already established world models

Othello and other examples provide strong evidence for causally active state representations. The matrix acknowledges this in its fifth layer. The objection concerns generalization: a closed-board model is not automatically a model of the open social and physical world.

Multimodality solves grounding

Vision and action reduce an important gap and provide a correction channel unavailable to text alone. Success on a set of robots or images nevertheless establishes neither open-world grounding nor intention or consciousness. Grounding itself has degrees of connection, correction and action.

The layers are the author's philosophical choice

True. Their number and boundaries are open to criticism. The framework is neither an empirical discovery nor an approved standard. Its practical value is making the move from “we found a representation” to “consciousness is established” visible and contestable, rather than hidden inside one ambiguous word.

Turing and Searle: behavioral criteria and the limits of a program

Turing proposed the imitation game to turn “Can a machine think?” into a testable behavioral criterion.^[A1] Its strength is preventing debate from revolving around untested definitions. But it does not image internal states or make every linguistic success proof of consciousness.

Searle's “Chinese Room” objects to executing a symbolic program being sufficient for intrinsic intentionality.^[A2] This influential philosophical argument is not a laboratory experiment proving all machine understanding impossible. Turing raises the value of behavioral testing; Searle prevents reducing intention to following rules. Neither closes modern empirical debate.

What about consciousness?

There is no agreed test turning fluency directly into a certain judgment of consciousness. “I feel” is not an independent report of an internal state; it is linguistic output influenced by data and instructions. Conversely, structural differences from humans do not suffice to prove future consciousness impossible.

The disciplined statement is: our evidence concerning representation, generalization and intervention does not settle subjective experience. This position is narrower than either denial or affirmation. Reviews such as Mitchell and Krakauer emphasize that disagreement over “understanding” itself depends on different definitions and levels, and that the data do not reduce debate to one slogan.^[A14]

Who bears responsibility when the system “decides”?

Even stronger planning and generalization do not erase the human chain: those who collected data, designed the objective, selected the model, imposed constraints, deployed it, brought it into an institution and chose to rely on its outputs. “The model decided” may be operational shorthand, but should not conceal responsibility's distribution.

Here the Quranic context returns without a technical projection. Teaching names occurs in the story of a vicegerent capable of building and of corruption. Greater ability does not equal greater righteousness; it increases the weight of purpose. The text–machine connection is not that Adam was “trained” like models, but that knowledge and ability cannot be ethically separated from responsibility.

Claim–evidence–limit appendix

A broader machine-readable register of claims, sources and limits is available in claims.json. The register aids scrutiny rather than replacing the original sources.

Limits and what the study does not claim

• It does not claim the Quran predicted artificial intelligence or Transformer architecture.
• It does not equate teaching names with training a language model.
• It does not claim every model “understands” or every model “understands nothing.”
• It does not make an embedding a real concept merely because it is decodable.
• It does not make a limited world model evidence of consciousness.
• It does not treat a system's self-report as independent evidence of subjective experience.
• It does not treat the name–world–responsibility matrix as a validated measure; it is an analytical framework open to criticism.
• No independent religious review or external review by a machine-learning specialist has taken place as of the last update.

A citable conclusion

The best available description is neither “merely a parrot” nor “a digital human mind.” Evidence shows that predictive objectives can produce regular internal representations, with causal roles in limited tasks. There is also evidence of fragility, shortcuts and gaps in generalization. Understanding therefore cannot be reduced to one result: we must identify the measured layer, the test and what remains beyond measurement.

Reading the teaching of names to Adam adds another restraint: knowledge in Al-Baqarah 30–33 occurs within vicegerency, creatures' limited knowledge and responsibility for ability. This describes no computer algorithm, but prevents increased ability becoming an argument for moral sufficiency. The crucial question is not just what the machine can say, but what relationship links its symbols to the world, what has been established causally, who owns the purpose and who bears the effects of use.

The question's origin in “Travel Through the Earth”

This dossier arose from “The Taught Human Being” in Ahmed Alhafiz's forthcoming “Travel Through the Earth and Observe How Creation Began,” particularly the transition from name to concept, relation and model, and from knowledge to responsibility. Every exegetical and technical claim is documented through sources independent of the manuscript.

Book page — forthcoming

Annotated references

1. The Quran, Al-Baqarah 30–33. The governing textual unit: vicegerency, corruption, teaching, the angels' acknowledgment of their limited knowledge, and Adam's informing them.
2. Al-Tabari, Jami al-Bayan, Al-Baqarah 31. Important for differing accounts of the names' scope and correcting the attribution of a preference for universality to him.
3. Al-Qurtubi, Al-Jami li-Ahkam al-Quran, Al-Baqarah 31. Distinguishes name, naming and named thing, and reports disagreement over scope.
4. Al-Razi, Mafatih al-Ghayb, Al-Baqarah 31. Discusses attributes, properties and the merit of knowing realities compared with words alone.
5. Ibn Ashur, Al-Tahrir wa-l-Tanwir, Al-Baqarah 31. Connects teaching with vicegerency, expression and the emergence of sciences.
6. Abu Hayyan, Al-Bahr al-Muhit, Al-Baqarah 31. Discusses names, naming, named things and teaching preceding informing.
7. A. M. Turing, “Computing Machinery and Intelligence,” Mind, 1950. The original source for the imitation game and operational behavioral criterion.
8. John Searle, “Minds, Brains, and Programs,” 1980. The Chinese Room argument against program execution being sufficient for intrinsic intentionality: a philosophical argument, rather than an empirical result.
9. Stevan Harnad, “The Symbol Grounding Problem,” 1990. The classic formulation of symbol grounding.
10. Vaswani et al., “Attention Is All You Need,” 2017. The foundational Transformer paper; consciousness and meaning are not its results.
11. Brown et al., “Language Models are Few-Shot Learners,” 2020. Evidence of broad performance in a large autoregressive language model.
12. Bender & Koller, “Climbing towards NLU,” ACL 2020. The strong argument separating linguistic form from meaning tied to intention and reference.
13. Bisk et al., “Experience Grounds Language,” EMNLP 2020. A research program connecting language with physical and social experience and action.
14. Lake & Baroni, “Generalization without Systematicity,” ICML 2018. An early reference for gaps in compositional generalization in sequence models.
15. Piantadosi & Hill, “Meaning without Reference in Large Language Models,” 2022. An opposing position that conceptual-role relations may carry aspects of meaning; a non-peer-reviewed preprint.
16. Li et al., “Emergent World Representations,” 2022. Othello-GPT: decoding board state and intervening on representations; important evidence within a closed game domain.
17. Gurnee & Tegmark, “Language Models Represent Space and Time,” ICLR 2024. Decodable representations of space and time across entities and contexts.
18. Driess et al., “PaLM-E,” 2023. A multimodal robotic model adding vision, state and action to language; preprint.
19. Mahowald et al., “Dissociating Language and Thought in Large Language Models,” 2024. Distinguishes formal linguistic competence from broader functional capacities.
20. Mitchell & Krakauer, “The Debate over Understanding in AI’s Large Language Models,” PNAS, 2023. A balanced framework for disagreement over understanding and levels of evidence.
21. Hewitt & Liang, “Designing and Interpreting Probes with Control Tasks,” 2019. Explains why probe accuracy is insufficient and introduces control tasks and selectivity.
22. Geiger et al., “Causal Abstraction,” JMLR, 2025. A theoretical basis for testing causal fidelity in mechanistic interpretation.
23. Sakai et al., ACL 2025. A compositional-order test revealing reliance on familiar orders and generalization gaps.
24. Ismayilzada et al., NAACL 2025. Tests morphological generalization on new roots and increasing complexity in two agglutinative languages.
25. Zhao et al., EMNLP 2025. Shows some scores' fragility under option length, problem template and irrelevant names.
26. Schaeffer et al., “Are Emergent Abilities of Large Language Models a Mirage?”, NeurIPS 2023. Explains how discrete metrics can create apparent jumps in performance curves.

Religious sources establish the text and exegetical disagreement; technical sources establish what experiments measure within their limits. No source alone is used to prove or deny consciousness. Each claim's status and confidence remain declared in the claims and evidence data.