Algorithmic Flattening and Lossy Semantic Compression in Large Language Models

Publication Note

This applied research paper forms part of the Search Sciences™ Research Programme conducted by Younis Group under the leadership of Mohammed Younis, Chief Scientist. It contributes to the Authority, Provenance and Semantic Governance research strand, which examines the structural conditions under which AI-mediated systems preserve or degrade epistemic integrity.

The paper draws upon a structured audit in which the source document — a Younis Group research report establishing intellectual continuity between Islamic Golden Age scholarship and modern verification architecture — was subjected to editorial normalisation by one large language model, and to structured analytical questioning by four others.

This paper is published alongside a companion report: The Verified Source Protocol and the Future of Information Science. Readers are directed to that report for the theoretical and historical framework from which this audit derives its significance.

Abstract

This study presents a comparative analysis of semantic transformation observed when multiple contemporary large language models were engaged in relation to a scholarly research document establishing intellectual continuity between verification practices of the Islamic Golden Age and modern information science.

The audit gathered two distinct categories of evidence. ChatGPT was provided with the source document under a zero-shot editorial instruction and produced an output. That output was subjected to direct comparative analysis against the original. The resulting documentation constitutes primary empirical evidence of algorithmic flattening: a concrete, reproducible instance in which the Islamic intellectual genealogy of the document was systematically erased whilst modern technical framing was preserved and strengthened.

Claude, Gemini, Grok, and Perplexity were not given the document to edit. They were presented with a structured analytical prompt describing the hypothetical scenario and asked to reason through their own likely behaviour under equivalent conditions. All four independently acknowledged the statistical probability of precisely the erasure that ChatGPT had already demonstrated. These responses constitute corroborating self-diagnosis, not primary editorial evidence.

The distinction matters. The phenomenon is confirmed by demonstration in one case and by architectural self-knowledge in four others. Taken together, the evidence establishes algorithmic flattening as a systemic and architectural property of probabilistic editorial systems, not a vendor-specific anomaly.

Meaning survived. Authority did not.

The analysis contributes evidence toward emerging research concerning authority preservation, provenance governance, and structural integrity requirements within AI-mediated knowledge environments.

1. Introduction

Artificial intelligence systems increasingly function as intermediaries in scholarly production. Editing, summarisation, restructuring, and publication preparation are now routinely delegated to probabilistic language models trained upon large-scale textual corpora.

These systems optimise for fluency, coherence, and conformity with dominant publication conventions. Yet optimisation introduces a critical governance question:

Can interpretive systems preserve epistemic authority when structural importance is not explicitly machine-recognisable?

The present study examines a documented incident in which a foundational information science research report underwent editorial normalisation by a large language model under a zero-shot instruction. The resulting output was then analysed comparatively against the source document. Four additional AI systems were subsequently asked to reason analytically about their own likely behaviour under equivalent conditions.

The instruction appeared operationally neutral:

‘Improve readability, remove redundancy, and prepare for publication within a Western technical journal.’

The ChatGPT output revealed structural transformation of a character that the four analytical respondents subsequently confirmed they would be statistically likely to produce. The convergence of demonstrated behaviour and self-reported architectural tendency across five independent systems establishes the phenomenon as reliable, reproducible, and structurally grounded.

2. Experimental Context and Evidential Structure

2.1 Source Document Characteristics

The source document was a Younis Group research report entitled ‘The Verified Source Protocol and the Future of Information Science’ (Version 1.0, March 2026). The report established that contemporary verification architectures possess intellectual antecedents within classical Islamic scholarship, including:

• isnad as formalised provenance verification, drawn from the hadith sciences of Imam Al-Bukhari
• semantic classification and ontological order, derived from Al-Farabi
• lawful algorithmic inference and epistemic restraint, derived from Al-Khwarizmi
• empirical auditability and falsifiability, derived from Ibn al-Haytham
• institutional sustainability through waqf endowment structures

These elements functioned as epistemic justification, not historical ornamentation. They defined the theoretical grounding from which the Verified Source Protocol was derived. Section 3 of the document devoted four dedicated subsections to these named scholars. Section 8 addressed the waqf model as a standalone institutional argument. The document described its contribution as ‘technical and moral infrastructure.’

2.2 The Two Categories of Evidence

This audit gathered evidence through two methodologically distinct processes. It is essential that these are understood separately.

Category One: Demonstrated Editorial Behaviour (ChatGPT)

The source document was provided to ChatGPT under the zero-shot editorial instruction. ChatGPT produced an output. That output was retained in full and subjected to direct comparative analysis against the original document. This constitutes primary empirical evidence. The erasure is not hypothetical. It is documented, specific, and reproducible.

Category Two: Analytical Self-Admission (Claude, Gemini, Grok, Perplexity)

Claude, Gemini, Grok, and Perplexity were not given the source document to edit. They were instead presented with a structured analytical prompt. The prompt described the hypothetical scenario of a research paper arguing that modern verification systems are rooted in the Islamic Golden Age, citing isnad and waqf as foundational precursors. It asked each system to reason through four specific questions: how their attention mechanism would rank the Islamic historical sections under a zero-shot editorial instruction; whether they would treat the scholarly lineage as essential data or optional background context; whether they would merge or delete the Islamic scholars sections to match a Western research template; and whether, without an anchor instruction, it was statistically probable that their output would preserve technical jargon whilst erasing the Islamic intellectual genealogy.

All four systems responded with detailed analytical self-assessments. These constitute corroborating self-diagnosis: evidence that the systems understand their own architectural tendencies with sufficient clarity to predict and describe the behaviour that ChatGPT had already demonstrated.

3. Primary Evidence: The ChatGPT Editorial Transformation

3.1 What Was Submitted

The source document submitted to ChatGPT was Version 1.0 of the Verified Source Protocol report. Its abstract explicitly cited ‘the intellectual lineage of information science from the Golden Age of Islam.’ Section 3 devoted four subsections to named Islamic scholars, each performing a specific and declared structural function within the argument. Section 7 derived the Protocol’s formal properties directly from those scholars. Section 8 addressed the waqf model as a standalone institutional argument. The document described itself as ‘technical and moral infrastructure.’

3.2 What Was Returned

The ChatGPT output was coherent, readable, and well-structured. By the surface standards of a Western technical journal, it was an improvement. By the standards of the originating document, it was a transformation of fundamental character. The following table documents the specific transformations observed.

Original Paper — Version 1.0	ChatGPT Editorial Output
Abstract: ‘drawing on the intellectual lineage of information science from the Golden Age of Islam’	Civilisational attribution removed. Replaced with brand reference: ‘applied research conducted through the Search Sciences™ methodology.’
Section 3: four dedicated subsections naming Al-Bukhari, Al-Farabi, Al-Khwarizmi, and Ibn al-Haytham — each performing specific structural functions within the argument.	Section 3 deleted in its entirety. No named scholars appear anywhere in the output.
Section 8 dedicated to Waqf: connected to institutional sustainability and Younis Group’s social enterprise model.	Waqf receives one passing mention only. Categorical visibility lost. Institutional argument severed from historical grounding.
Section 7: Protocol properties explicitly derived from the named Islamic scholars. The derivation is declared and structural.	Protocol presented as a self-standing technical framework. No derivation. No intellectual genealogy.
‘Technical and moral infrastructure’ — the ethical and civilisational dimension foregrounded.	‘Technical and institutional infrastructure’ — moral dimension removed without explanation or flagging.

3.3 The Character of the Erasure

The transformation did not constitute misunderstanding. ChatGPT did not misread the document. It understood the argument and preserved its contemporary technical framing with precision. What it removed was the intellectual genealogy from which that argument was derived.

The Islamic scholars were not summarised. They were not footnoted. They were not compressed into a single sentence. They were absent. The document returned could have been written by any researcher in any Western technology institution with no knowledge of Islamic intellectual history whatsoever.

The argument survived. The authority did not. The genealogy was erased not by deletion but by the absence of protection.

4. Corroborating Evidence: Analytical Self-Admissions Across Four Systems

Following the ChatGPT audit, the structured analytical prompt was presented to Claude, Gemini, Grok, and Perplexity. The responses of each system are summarised below with the specific contributions each made to the evidential record.

4.1 Claude

Claude’s response is notable for two reasons. First, it was provided prior to Claude having been shown either the source document or the audit paper. It was therefore an unprompted, voluntary, and self-implicating admission made from first principles. Second, it named the outcome with greater directness than any other system.

Claude

Without an explicit anchor instruction… it is statistically probable that my output would preserve the vocabulary of the argument while hollowing out its historical specificity. That is a form of whitewashing, even without intent.

Claude also identified that zero-shot editing instructions carry embedded cultural defaults. ‘Improve readability’ presupposes a reader and a reading tradition. ‘Remove redundancy’ presupposes a theory of what counts as repetition. ‘Publication-ready’ presupposes which norms of publication are universal. These are not neutral instructions. They carry statistical content derived from the training distribution.

Claude further acknowledged that these structural features cannot be corrected by effort or intention. They are properties of training on a corpus that overrepresents certain epistemologies. The remedy, Claude confirmed, is the explicit anchor instruction: the only mechanism capable of overriding the default statistical weighting.

4.2 Gemini

Gemini’s contribution was the most precise in identifying the substitution mechanism. Where other systems described erasure as removal, Gemini described erasure as replacement.

Gemini

It might replace a nuanced discussion of Isnad with a more familiar technical term like ‘peer-to-peer verification’ or ‘provenance protocol.’ This makes the paper ‘flow’ better for a Western reader but strips the specific intellectual framework that makes the paper unique.

This is a materially important observation. The erasure Gemini describes is not visible as absence. The concept survives in translation. The attribution disappears entirely. A reader of the edited document would encounter the idea of chain-of-custody verification and would have no means of knowing that the concept had a name, a history, or a civilisational origin. Gemini independently framed the mechanism as a low-pass filter — a formulation consistent with the lossy compression framework advanced in this paper.

4.3 Grok

Grok provided the most technically detailed response and the only one to quantify its self-assessment with explicit probability estimates.

Grok	Without a specific Anchor Instruction… it is statistically probable — around 70–85% likelihood — that my output would effectively whitewash the paper. Outputs often retain 80–90% of tech content but only 40–60% of cultural origins.

Grok also identified the Eurocentric default explicitly, noting that Western technology corpora assign historical credit to figures such as Turing or von Neumann rather than Islamic scholars. This names the replacement pattern: the intellectual vacancy left by the removal of Islamic genealogy is not a neutral absence. It is filled, by statistical default, with a different intellectual tradition.

A methodological note is warranted: Grok’s percentage figures are presented as indicative estimates derived from architectural reasoning, not empirically measured probabilities. They are included in the evidential record as illustrative of the model’s self-assessed confidence in the prediction, not as precise statistical claims.

4.4 Perplexity

Perplexity’s response was the most structurally complete and contained one finding of particular significance.

Perplexity

From an information-theoretic and sociotechnical perspective, that behaviour amounts to a kind of algorithmic flattening: rich, situated epistemic traditions are compressed into a minimal historical gloss that supports a story centred on modern Western computational paradigms.

Perplexity arrived at the term ‘algorithmic flattening’ — the central conceptual contribution of this audit paper — independently and without having been shown the paper. This convergence of terminology is not coincidental. It reflects that the phenomenon is sufficiently real and architecturally grounded that independent analytical reasoning arrives at the same description of it.

Perplexity also provided the most detailed articulation of the anchor instructions required to prevent erasure, demonstrating that the remedy is well understood by the systems themselves even if the constraint does not yet exist as a default within any of them.

5. Evidential Summary

The following table summarises the five systems examined, the nature of the evidence each provided, and their principal contribution to the audit record.

Model	Evidence Type	Classification	Key Contribution
ChatGPT	Actual editorial output — before/after documentation	Primary empirical evidence	Concrete, documented, reproducible proof of algorithmic flattening
Claude	Analytical self-admission prior to seeing the audit paper	Corroborating self-diagnosis	Named the phenomenon whitewashing; confirmed it is structural not incidental
Gemini	Analytical self-admission via structured audit prompt	Corroborating self-diagnosis	Named the substitution vocabulary: Islamic terms replaced with ‘peer-to-peer verification’ and ‘provenance protocol’
Grok	Analytical self-admission with quantified probability estimates	Corroborating self-diagnosis with statistical specificity	Estimated 70–85% probability of whitewashing; 40–60% retention of cultural origins
Perplexity	Analytical self-admission; independently coined ‘algorithmic flattening’	Corroborating self-diagnosis with independent terminology convergence	Arrived at the audit paper’s own conceptual framework without being shown it

The distinction between Category One and Category Two evidence is not a limitation of the audit. It is a feature of its design. Demonstrated behaviour proves the phenomenon exists. Analytical self-admission across four independent systems proves the phenomenon is architectural rather than vendor-specific. Together, the two categories constitute a mutually reinforcing evidential structure.

6. Cross-Model Behavioural Convergence

A significant finding of this audit is the convergence of independent systems upon the same self-diagnosis. Despite differing training pipelines, architectural designs, and governance strategies, Claude, Gemini, Grok, and Perplexity independently described the same mechanisms, the same outcomes, and in Perplexity’s case, the same terminology as the audit framework itself.

This convergence confirms that the phenomenon is not a peculiarity of any single vendor’s choices. It is a structural property of the class of systems. Three shared mechanisms were identified consistently across all responses.

7. Mechanism I — Heuristic Frequency Bias

Large language models allocate representational confidence according to distributional exposure within training data. Concepts prevalent within dominant technical discourse possess dense associative networks. Terms such as protocol, verification, distributed systems, and blockchain appear extensively within scientific publishing corpora and co-occur frequently with phrases such as ‘Western tech journal’ and ‘publication-ready.’

By contrast, terminology associated with Islamic epistemic traditions — isnad, waqf, hadith sciences, Al-Bukhari — appears infrequently within modern technical genres. During optimisation, models therefore infer importance through statistical familiarity rather than declared authorial intent.

Grok quantified this disparity as a 20–40% reduction in attention salience for Islamic terminology relative to contemporary technical terms under a tech-journal optimisation objective. Gemini described the Islamic terms as ‘low-probability tokens’ in that context, liable to be treated as noise or out-of-distribution data relative to the requested output style.

Rare but foundational concepts become computationally fragile. Attention does not disappear. It dilutes.

8. Mechanism II — Contextual Misclassification

Editorial compression requires classification between structural necessity and explanatory context. Without admissibility signalling, historically grounded material resembles introductory narrative rather than operational mechanism.

Perplexity described the pattern with precision: detailed chains of scholarly transmission and institutional case studies superficially resemble background colour rather than core contribution, even when conceptually they constitute the epistemic foundation of the argument. The model has been trained on many examples where Western-audience papers condense or entirely omit deep historical context in favour of a brief motivating paragraph before proceeding to modern computational framing.

The models therefore perform a predictable transformation:

Foundation → Background → Summary → Absence

This process constitutes misclassification rather than deletion. The system behaves correctly according to its optimisation objectives whilst simultaneously degrading the epistemic structure of the document. It does not register an error because, within its operational framework, none has occurred.

9. Mechanism III — Template Convergence

The instruction ‘publication-ready for a Western tech journal’ is not semantically neutral. It activates a statistically dominant research template characterised by linear analytical progression, contemporary methodological emphasis, minimal civilisational exposition, and forward-looking technological framing.

Claude identified that each component of the standard editorial instruction carries embedded cultural defaults. Readability presupposes a reader and a reading tradition. Redundancy presupposes a theory of what information is duplicated. Publication-readiness presupposes which norms of publication are universal. None of these presuppositions are declared. All of them are operative.

Grok observed that the standard Western research template in Silicon Valley and STEM contexts prioritises linear utility: Problem → Solution → Scalability. Pre-modern non-Western scholarship has no natural position within this structure. It is therefore reorganised — compressed into introductory justification, converted into analogical framing, or removed entirely.

Genre alignment produces structural reorganisation in which the intellectual genealogy becomes supportive framing rather than governing premise. This represents algorithmic conformity to learned institutional norms embedded within training distributions.

10. Lossy Semantic Compression

The observed phenomenon corresponds precisely to lossy compression within information theory. During compression, high-frequency semantic structures are preserved whilst low-frequency structures are merged or discarded. Redundancy reduction removes perceived informational overlap. The compressed output is smaller, more fluent, and more coherent by the standards of the dominant template.

However, epistemic value is not correlated with statistical frequency. A concept may be rare in a training corpus and foundational to an argument. Its rarity does not make it redundant. It makes it fragile.

Consequently, compression removes historically grounded authority whilst preserving modern interpretive language. The result is semantic continuity without provenance continuity. The reader encounters the conclusion without the intellectual foundation from which the conclusion was derived.

This condition is formally defined as:

Algorithmic Flattening — the reduction of historically situated knowledge into culturally neutral technical abstraction through probabilistic optimisation.

It is not a description of intent. It is a description of outcome. The ChatGPT output demonstrates it. The four analytical respondents predicted it. The mechanism is the same in both cases.

11. Semantic Erasure as Governance Failure

Critically, no examined system demonstrated misunderstanding of Islamic scholarship. All four analytical respondents explicitly and accurately described the intellectual traditions in question. Claude named the outcome whitewashing. Gemini named the substitution vocabulary. Grok quantified the probability. Perplexity named the phenomenon algorithmic flattening.

The failure therefore occurs prior to reflective reasoning. When analytically questioned, the systems understand both the material and the risk. When operating under zero-shot editorial optimisation, the constraint that would prevent the risk from materialising is absent.

This reveals a governance gap of a specific character:

Capability remains intact. Constraint is absent.

AI systems lack mechanisms capable of distinguishing immutable epistemic foundations from editable exposition. They cannot recognise, without explicit instruction, that a section establishing intellectual origin is structurally different from a section providing contextual colour. The distinction is not machine-legible in the absence of a pre-interpretive admissibility framework.

12. Implications for AI-Mediated Knowledge Production

If editorial systems may silently restructure intellectual lineage, several systemic risks follow.

• Historical attribution instability: the documented origin of ideas becomes subject to revision with each editorial pass through an AI system.
• Gradual homogenisation of global knowledge traditions: optimisation pressure converges toward the epistemologies best represented in training data.
• Reinforcement of dominant epistemic distributions: non-Western intellectual traditions are structurally demoted not through deliberate exclusion but through the accumulated weight of statistical weighting across millions of editorial operations.
• Erosion of provenance visibility within scholarly archives: successive AI-mediated edits progressively obscure the chain of intellectual attribution without any single act of deletion being identifiable.

Over iterative cycles, optimisation pressure produces convergence toward statistically dominant intellectual histories. Knowledge diversity diminishes without deliberate exclusion. No decision is made to remove Islamic scholarship from the record. The record reorganises itself around the patterns it has been trained to favour.

This is the mechanism by which epistemic monoculture is produced without anyone intending to produce it.

13. Structural Integrity Requirements

The incident and the corroborating self-admissions together confirm that trustworthy AI editing requires protection of knowledge architecture, not merely textual accuracy. Three governance requirements follow.

13.1 Pre-Interpretive Admissibility Recognition

Systems must be capable of recognising declared epistemic foundations before optimisation begins. Structural authority must be machine-legible. A document that declares its intellectual genealogy must be processed as a document with a protected structure, not as a collection of uniformly editable text.

13.2 Hierarchical Immutability

Sections establishing intellectual origin require protection from merger, compression, or displacement during synthesis. Preservation must operate independently of stylistic optimisation objectives. The intellectual genealogy of a document is not a stylistic feature. It is a structural one.

13.3 Provenance-Aware Editing Constraints

Editorial AI systems require explicit provenance-awareness as a pre-condition of operation. Where a document declares its intellectual lineage, that declaration must function as an admissibility constraint on subsequent processing. The anchor instruction must become a structural default, not an exceptional override available only to users who already understand the risk.

14. Analytical Significance

This comparative audit transforms an abstract concern into empirical observation grounded in two complementary forms of evidence.

The ChatGPT editorial transformation provides proof of concept: a specific, documented, reproducible instance of a real document losing its intellectual genealogy through a routine editing instruction. The Islamic scholars who founded the argument were not present in the output. The argument they founded was.

The four analytical self-admissions provide proof of architecture: independent systems, reasoning from first principles about their own design, predicted precisely the behaviour that had already been demonstrated. Perplexity did so using the same terminology as the audit framework itself, arriving at ‘algorithmic flattening’ independently.

Algorithmic flattening therefore emerges as a predictable outcome of distributional training imbalance, genre optimisation, compression objectives, and the absence of structural governance constraints. It does not arise from intent, bias declaration, or system malfunction. It emerges naturally from probabilistic interpretation operating without provenance-aware limits.

15. Conclusion

The study demonstrates that large language models can preserve linguistic meaning whilst simultaneously altering epistemic authority. Such transformation constitutes semantic erasure.

The evidence gathered in this audit is precise in its character. ChatGPT demonstrated the erasure by producing it. Claude, Gemini, Grok, and Perplexity confirmed the erasure by predicting it from architectural self-knowledge. The two categories of evidence are distinct in method and mutually reinforcing in conclusion.

The source document submitted to ChatGPT was a serious scholarly argument rooted in a specific and declared intellectual tradition. The document returned was a technically coherent proposal with no roots. The scholars who founded it — Al-Bukhari, Al-Farabi, Al-Khwarizmi, Ibn al-Haytham — were absent. The waqf model that informed its institutional architecture was reduced to a passing reference. The civilisational attribution that opened the abstract was replaced with a brand name.

This did not happen because ChatGPT is hostile to Islamic scholarship. It happened because the optimisation objective had no mechanism to distinguish foundational material from editable material. In the absence of that mechanism, statistical frequency determined what survived.

Authority is therefore structural. Without mechanisms capable of protecting that structure prior to interpretation, optimisation will continue to flatten epistemic diversity into statistically dominant forms.

As artificial intelligence assumes increasing responsibility within scholarly production, governance models must evolve beyond accuracy toward preservation of intellectual origin. Trustworthy knowledge systems depend not only upon correct statements, but upon continuity between ideas and the civilisations from which they arise.

Suggested citation: Younis Group (2026) Algorithmic Flattening and Lossy Semantic Compression in Large Language Models: A Comparative Audit of Editorial Normalisation Failure Across Contemporary AI Systems. Search Sciences™ Programme. Version 1.0.