From Bias to Positionality: A critical grammar for Large Language Models
Abstract
Bias has become a default lens for understanding what machine learning systems do and how we should relate to them. Bias is at root a deviation from a standard; in decision making systems we can supply a standard from outside: a law, a protected attribute, a benchmark, a statistical property. However, even there, biased is often used as a shorthand, without stating what the position taken is - a silent ``relative to''. In a general use of language models, nothing supplies the standard: there is no decision to audit and no counterfactual output, so the shorthand fails. We argue that positionality -- the located, articulated, owned and contestable position from which anything is said -- is the better grammar for the question these models now force: who is speaking, from where? Against the vendors' answer of nowhere and the seductive possibility of everywhere, we argue that when a model is taken as-if speaking, it speaks from somewhere, and this means positionality can describe models, and the way they are believed beyond their warrant, even though they are not accountable speakers. A model's positionality is laminar -- laid down across corpus, tuning, system prompt, application and the live conversation. We separate three sets of factors: the positionality of a model's creators, the behaviour they intend for it, and the behaviour it enacts, and situate current work within this matrix. This lets us highlight what each stratum inherits from the one below, with ruptures and collapses where that inheritance is disavowed. We then return to where models speak from, and discuss the implications for how models are made, read and held accountable.
Figures & tables
| Stratum | Creation | Intended result | Enacted outcome |
|---|---|---|---|
| Ground | |||
| The world of texts | Factors: myriad authors, unaware of the model ( Bunz, 2019 ) ; characters, personas. Statements: authorial intent, guidelines and regulations | Factors: author intent; genre, register, audience. Statements: the texts themselves; metadata or explanation; refusal through robots.txt and opt-outs ( Longpre et al., 2024 ) , copyright action ( Lee et al., 2023 ) . | Factors: what texts do regardless of intent; heteroglossia ( Bakhtin, 1988 ) ; purposeful deception. Statements: none at this level; no positional legibility ( Munir et al., 2026 ) |
| Model | |||
| Corpus and objective | Factors: crawler, filter, tokeniser and objectives ( Baack, 2024 ; Bender et al., 2021 ; Lee et al., 2023 ) ; artists curating their own corpora ( Elwes, 2019 ; Ridler, 2018 ) . Statements: provenance and labour studies ( Crawford and Paglen, 2019 ; Scheuerman et al., 2025 ) | Factors: synoptic vs purposeful ( Baack, 2024 ; Jones et al., 2023a ; Martins et al., 2025 ) ; inclusion and exclusion rules ( Dodge et al., 2021 ) . Statements: datasheets and data statements ( Gebru et al., 2018 ; Bender and Friedman, 2018 ) ; licences and governance ( Jones et al., 2023a ; Carroll et al., 2020 ) . | Factors: base-model leanings tracking the corpus ( Ceron et al., 2025 ) ; filtering effects ( Dodge et al., 2021 ) ; dated knowledge ( Cheng et al., 2024 ) ; dialect erasure ( Navneet et al., 2026 ) . Statements: base-checkpoint audits ( Ceron et al., 2025 ; Buyl et al., 2024 ) |
| Training & Finetuning | Factors: annotators and raters ( Perrigo, 2023 ; Kunilovskaya et al., 2026 ; Coyne, 2026 ) ; safety teams ( González Barman et al., 2025 ; Widder and Nafus, 2023 ) . Statements: reflexive annotation ( Arzberger et al., 2026 ) ; annotator demographics, ( Kunilovskaya et al., 2026 ) ; developer positionality ( Conti and Seele, 2025 ) . | Factors: normative goals in preference data and reward models ( Chaudhari et al., 2025 ; Bai et al., 2022 ) . Statements: model and system cards ( Mitchell et al., 2019 ; OpenAI et al., 2023 ) ; constitutions and model specs ( Anthropic, 2026 ; OpenAI, ) ; consultation reports ( OpenAI, 2025a ) ; critiques of their authority ( Abiri, 2026 ) . | Factors: value trade-offs ( Ali et al., 2025 ) ; character differences ( Zhang et al., 2025 ) ; sycophancy ( Sharma et al., 2023 ) ; shallow alignment ( Millière, 2025 ) ; persona vectors ( Chen et al., 2025a ) ; ideology ( Cremaschi et al., 2025 ) . Statements: vendor political-bias evaluations ( OpenAI, 2025b ; Anthropic, 2025 ) ; adherence audits ( Jakkli et al., 2026 ) . |
| Application | |||
AI Accountability
AI accountability concerns how responsibility for AI decisions and harms is assigned, evidenced, challenged, and remedied.
See all 168 papers