Elamur Research⌕
☰
home / papers / 2026 / The Algorithmic Unconscious: Structural Mechanisms and Implicit Biases in Large Language Models

The Algorithmic Unconscious: Structural Mechanisms and Implicit Biases in Large Language Models

Paper · 2026 · Philippe Boisnard

Summary

Boisnard argues that some LLM biases do not come only from skewed training data or from human intentions read into machines. They also arise from the technical machinery itself: tokenization, attention, statistical optimization and safety alignment. He calls these hidden determinations, which neither users nor the model can reflect on, the "algorithmic unconscious." The empirical core is a small benchmark built with a Python script. It counts input tokens for ten parallel sentences in English, Modern Standard Arabic, Darija (in Arabic script and as Arabizi) and other languages, using several tokenizer infrastructures. Arabic varieties consistently need more tokens than English. The inflation is about 1.6× with OpenAI's tokenizer and nearly 4× with Mistral's SentencePiece. The author links this over-segmentation to three further mechanisms: correlation mistaken for causation, the erasure of minority features through dimensional collapse, and normative biases from alignment. He then proposes a "technical clinic" that audits tokenization, latent-space geometry and alignment. The point is that this bias is measurable and structural, with economic and cultural costs for speakers of non-majority languages.

Key points

  • OpenAI benchmark (tiktoken, reported as GPT-4.1): average tokens relative to English (7.1) were:
    • MSA: 11.5 (1.62×)
    • Darija in Arabic script: 11.3 (1.59×)
    • Arabizi: 13.2 (1.86×)
  • Mistral-7B (SentencePiece) benchmark: inflation was much larger relative to English (8.4):
    • MSA: 33.1 (3.94×)
    • Darija: 29.1 (3.46×)
    • Arabic words were sometimes split down to single characters.
  • Measurement caveat: the OpenAI counts reflect the shared tiktoken encoding, so they describe the infrastructure's policy rather than differences between individual models.
  • Cost argument: citing Petrov et al. (2023), the paper notes that doubled fertility can multiply training cost up to fourfold. Users also pay more context tokens for the same query.
  • Script and normalization effects: the appendix attributes part of the inflation to script handling, normalization pipelines and orthographic conventions, not just to language. Ukrainian is singled out as a visible case.

Interesting ideas

  • I1. Auditing a tokenizer with a fixed set of parallel sentences and comparing average token counts per language and tool gives a cheap, repeatable fairness metric. Such a metric can be tracked across versions.
  • I2. The paper separates script and normalization effects from linguistic differences. Some "language" bias is really a preprocessing artifact, so it can be fixed upstream.
  • I3. Statistical learners pick up co-occurrence rather than intervention. Causal-sounding output from a corpus-trained model is therefore a correlation unless something else grounds it.
  • I4. Majority patterns can absorb rare features through statistical preference (dimensional collapse), and averaged metrics hide this erasure.
  • I5. A "technical clinic" audits each internal layer separately (segmentation, representation geometry, alignment filters) instead of judging only final outputs.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tokens): Run the en tokenizer on parallel variants of the same content (British/American spelling, informal and dialectal forms, transliterated names) and flag inputs whose segmentation inflates or becomes unstable.
  • I2 (tokens): Audit normalization (Unicode quotes, dashes, diacritics in loanwords and names) separately from morphology so that tagger and lemma errors are not blamed on the wrong layer.
  • I3 (learn): Treat causal or event-order claims from LLM teacher tiers as unverified until the world model or several witness parsers agree, and add negative controls where cue words co-occur without a causal link.
  • I4 (tree): Report LAS per construction and per rare-dialect slice for the perceptron parser and gparse, not only overall, so that rare constructions are not silently traded away.
  • I5 (learn): Mirror this with per-stage evaluation suites (tokens, tree, frames, world, answer) so that a bias can be located at the step where it enters the explainable pipeline.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 9,085 worksArtificial Intelligence · 8,888 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksSocial Sciences · 744 worksSafety Research · 221 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 1,344 works

Cite

Cite as Philippe Boisnard. 2026. The Algorithmic Unconscious: Structural Mechanisms and Implicit Biases in Large Language Models. arXiv:2602.18468.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2602.18468 · arXiv 2602.18468

@misc{boisnard2026algorithmic,
  title = {{The Algorithmic Unconscious: Structural Mechanisms and Implicit Biases in Large Language Models}},
  author = {Philippe Boisnard},
  year = {2026},
  doi = {10.48550/arxiv.2602.18468},
  eprint = {2602.18468},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2602.18468}
}

References

26 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. (Alrefaie et al, 2024) Alrefaie, M. T., Morsy, N. E., & Samir, N. (2024). Exploring Tokenization Strategies and Vocabulary Sizes for Enhanced Arabic Language Models. arXiv preprint arXiv:2403.11130. ↗
  2. (Alyafeai et al, 2024) Alyafeai Z., Pieler M., Teufel H., Tow J., Bellagente M., Phung D., Pinnaparaju N., Adithyan R., Rocha P., Zhuravinskyi M., RiquelmeArabic C., Arabic Stable LM: Adapting Stable LM 2 1.6B to Arabic, (2024) arXiv:2412.04277v1 [cs.CL] (Ashraf et al. 2025) Ashraf Y., Wang Y., Gu B., Nakov P., Baldwin T., Arabic Dataset for LLM Safeguard Evaluation, arXiv:2410.17040v2 [cs.CL] 09 Feb 2025 ↗
  3. (Assoudi, 2025) Assoudi H., A Comparative Benchmark of a Moroccan Darija Toxicity DetectionModel (Typica.ai) and Major LLM-Based Moderation APIs (OpenAI, Mistral, Anthropic), (2025) arXiv:2505.04640 [cs.CL] ↗
  4. (Bari et al., 2024) Bari S., Alnumay Y., A. Alzahrani N., M. Alotaibi N., A. Alyahya H., AlRashed S., A. Mirza F., Z. Alsubaie S., A. Alahmed H., Alabduljabbar G., Alkhathran R., Almushayqih Y., Alnajim R., Alsubaihi S., Al Mansour R., Alrubaian M., Alammari A., Alawami Z., Al-Thubaity A., Abdelali A., Kuriakose J., Abujabal A., Al-Twairesh N., Alowisheq A., Khan H., ALLaM: Large Language Models for Arabic and English, (2024) arXiv:2407.15390v1 [cs.CL] ↗
  5. (Boisnard, P., 2026) Ethology of Latent Spaces (arXiv:2602.05710) ↗
  6. (Barocas et al., 2023) Barocas S., Hardt M., Narayanan A., Fairness and Machine Learning (online draft/book)., https://fairmlbook.org/, Compiled on Wed Dec 13 14:44:57 CET 2023 ↗
  7. (Bender et al., 2021) M. Bender E., Gebru T., McMillan-Major A., Mitchell S., On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? (2021) FAccT '21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency Pages 610 - 623 https://doi.org/ 10.1145/3442188.344592 ↗
  8. Bensalem et al. "Toxic language detection: A systematic review of Arabic datasets" (2024) Colloque internationale Intelligence Arti cielle et mutations de l’information (Bolukbasi et al., 2016) Bolukbasi T., Chang K.W., Y Zou J., Saligrama V., T Kalai A., Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29, (2016). arXiv:1607.06520v1 [cs.CL] ↗
  9. (Bouvrès, 1999) Bouvrès J., Prodiges et vertiges de l’analogie, éditions raison d’agir, (1999). find
  10. (Bowker & Star, 1999) C. Bowker G., Star S.L., Sorting Things Out: Classi cation and Its Consequences (1999) MIT Press. find
  11. (Couldry & Mejias, 2022) Nick Couldry et Ulises A. Mejias, « Le colonialisme des données : repenser la relation entre le big data et le sujet contemporain », Questions de communication [En ligne], 42 | 2022. http://journals.openedition.org/questionsdecommunication/29845 ↗
  12. (Ethayarajh, 2019) Ethayarajh, K. (2019). How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings. EMNLP-IJCNLP 2019. arXiv:1909.00512 ↗
  13. (Gebru et al., 2021) Gebru T., Morgenstern J., Vecchione B., Wortman Vaughan J., Wallach H., Daumé III H., Crawford K., Datasheets for Datasets, (2021) arXiv:1803.09010v8 [cs.DB] ↗
  14. (Jing, 2022) Jing, L., Vincent, P., LeCun, Y., & Tian, Y. (2022). Understanding Dimensional Collapse in Contrastive Self-supervised Learning. ICLR 2022. arXiv:2110.09348 ↗
  15. (Kannan, 2024) Kannan P., How Much Research Is Being Written by Large Language Models?, 2024 https://hai.stanford.edu/news/how-much-research-being-written-large-language-models ↗
  16. (Mitchell et al., 2019) Mitchell M., Wu S., Zaldivar A., Barnes P., Vasserman L., Hutchinson B, Spitzer E., Deborah Raji I., Gebru T., Model Cards for Model Reporting, (2019) arXiv:1810.03993v2 [cs.LG] ↗
  17. (Naous & Xu, 2025 ) Naous T., Xu W., On The Origin of Cultural Biases in Language Models: From Pre-training Data to Linguistic Phenomena. (2025) arXiv:2501.04662v1 [cs.CL] ↗
  18. (Noble, 2018) Noble S.U., Algorithms of oppression: How search engines reinforce racism. New York University Press. find
  19. (Noels et al., 2025) Noels S., Bied G., Buyl M., Rogiers A., Fettach Y., Lijffijt J., De Bie T., What Large Language Models Do Not Talk About: An Empirical Study of Moderation and Censorship Practices, (2025) arXiv:2504.03803v1 [cs.CL] ↗
  20. (Pearl, 2018) Pearl J., The book of why : The new science of cause and effects (2018). New York: Basic Books. find
  21. (Possati, 2020) Possati L., Algorithmic unconscious: why psychoanalysis helps in understanding AI, (2020), https://www.nature.com/articles/s41599-020-0445-0 ↗
  22. (Petrov et al., 2023) Petrov, A., La Malfa, E., Torr, P. H. S., & Bibi, A. (2023). Language Model Tokenizers Introduce Unfairness Between Languages. arXiv preprint arXiv:2305.15425. Colloque internationale Intelligence Arti cielle et mutations de l’information (Shahid et al., 2025) Shahid F., Elswah M., Vashistha V., Think Outside the Data: Colonial Biases and Systemic Issues in Automated Moderation Pipelines for Low-Resource Languages (arXiv:2501.13836v3 [cs.CL] 05 Aug 2025) ↗
  23. (Mohamed et al., 2020) Mohamed S., Png M.T., Isaac W., Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence (2019), arXiv:2007.04068v1 [cs.CY]. ↗
  24. (Shang et al., 2024) Shang, G., Abdine, H., Khoubrane, Y., Mohamed, A., Abbahaddou, Y., Ennadir, S., Momayiz, I., Ren, X., Moulines, E., Nakov, P., Vazirgiannis, M., & Xing, E. (2024). Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect. arXiv preprint arXiv:2409.17912. ↗
  25. (Spivak, 1988) Spivak, G. C. (1988). Can the Subaltern Speak? In C. Nelson & L. Grossberg (Eds.), Marxism and the Interpretation of Culture (pp. 271–313). University of Illinois Press. find
  26. (Teklehaymanot & Nejdl, 2025) Teklehaymanot H.K , Nejdl W., Tokenization Disparities as Infrastructure Bias: How Subword Systems Create Inequities in LLM Access and Efficiency, arXiv:2510.12389 [cs.CL] ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/boisnard-2026-algorithmic-unconscious-structural-mechanisms-implicit/