Standard Language Ideology in AI-Generated Language

Technical Summary: Standard Language Ideology in AI-Generated Language

Problem Statement
Large Language Models (LLMs) and generative AI systems currently reproduce and normalize "standard language ideology," a sociolinguistic concept describing the bias toward specific language varieties (particularly Standard American English, or SAE) that are granted higher prestige, authority, and legitimacy than others. This is not merely a performance disparity where models struggle with non-standard dialects; it is a sociotechnical issue where AI systems actively reinforce linguistic hierarchies. These systems position "standard" varieties as the appropriate default while marginalizing minoritized varieties (e.g., African American English, Singlish, regional dialects) through mechanisms of stereotyping, appropriation, and erasure. The paper argues that these issues are structural rather than purely technical, reflecting power relations that determine whose language is preserved, filtered, or amplified in training data and model outputs.

Methodology
The authors employ a multidisciplinary approach, synthesizing empirical evidence from existing LLM evaluations with theoretical frameworks from sociolinguistics and critical language studies.

• Theoretical Grounding: The analysis is grounded in the mechanisms of standard language ideology identified by sociolinguists: iconization (linking linguistic features to social essence), fractal recursivity (projecting oppositions within subgroups), and erasure (obscuring facts that contradict the dominant narrative).
• Analytic Process: The team conducted a focused review of large-scale studies and systematic evaluations regarding linguistic bias in LLMs, complemented by a broader literature review on linguistic inequality. The analysis was both deductive (applying established sociolinguistic mechanisms as a scaffold) and inductive (iteratively mapping empirical findings to these categories).
• Synthesis: The authors categorized documented AI manifestations and linked them to potential societal implications, creating a structured framework to explain how linguistic biases in AI reflect and reinforce broader systems of power.

Key Contributions
The paper's primary contribution is the introduction of Standard AI-Generated Language Ideology and a faceted taxonomy of open problems that operationalize this concept. This taxonomy moves beyond a simple survey of model errors to categorize the structural mechanisms through which LLMs reproduce linguistic power hierarchies. The taxonomy identifies five specific open problems:

1. Normalization of "Standard" Language as Default: LLMs systematically default to prestige varieties (SAE), even when prompted with non-standard inputs. This reinforces the notion that standard varieties are the only "correct" or neutral modes of communication, potentially exacerbating automation bias where users trust AI outputs that align with these norms.
2. Unequal Quality of Service for Minoritized Varieties: Speakers of non-standard varieties face degraded performance, including lower comprehension, misclassification of inputs as toxic/hateful, and refusal to respond. This constitutes allocational harm, incentivizing "digital code-switching" and imposing psychological tolls on users.
3. Stereotyped and Racialized Productions: When models do generate text in minoritized varieties, they often produce stylized, caricatured, or demeaning content. This reflects racialization, where linguistic form becomes a proxy for negative social identity and moral judgment.
4. Appropriation and Commodification: LLMs enable the extraction of cultural and linguistic value from marginalized communities without accountability or benefit to those communities. This includes "linguistic minstrelsy" (non-native speakers using AI to appropriate dialects) and the erasure of the historical and ethnic origins of language (e.g., rebranding AAE as "Gen Z slang").
5. Suppression and Erasure: Efforts to "avoid harm" by suppressing non-standard outputs risk reproducing institutional erasure, where legitimate linguistic forms are pushed out of technological infrastructures, narrowing the linguistic future of AI ecosystems.

Results and Findings
The paper synthesizes existing empirical evidence to demonstrate that LLMs do not merely "struggle" with non-standard varieties but actively participate in constructing linguistic hierarchies:

• Performance Disparities: Studies show higher perplexity and lower accuracy for AAE and other non-standard varieties in tasks like text generation and sentiment analysis.
• Defaulting Behavior: Research indicates that even when prompted with non-standard inputs, model outputs revert to SAE orthography and syntax, with significant reductions in dialectal features (e.g., >96% reduction for Jamaican and African-American English inputs).
• Stereotyping: Native speakers rate AI outputs for minoritized varieties as significantly more demeaning and condescending compared to outputs for standard varieties.
• Data Inequities: The dominance of English in training data (approx. 60% of internet content vs. 17% of global speakers) and the filtering of non-standard varieties during data scraping are identified as material consequences of standard language ideology, not just technical limitations.
• Sovereign AI Context: The paper notes that while "sovereign AI" initiatives (e.g., in India) aim to counter English dominance, they risk reproducing internal linguistic hierarchies if they prioritize state-defined "standard" forms over dialectal and indigenous diversity.

Significance and Recommendations
The paper claims that these issues cannot be resolved through engineering fixes alone (e.g., better data balancing) because they are rooted in power relations and institutional priorities. The authors argue for a shift toward emancipatory outcomes that prioritize accountability, community agency, control, and ownership.

The paper proposes three concrete recommendations for researchers, funders, and practitioners:

1. Expand Evaluation Metrics: Model evaluation frameworks must explicitly include language variety as a category for assessing harm (alongside race and gender), developing benchmarks to detect erasure, stereotyping, and condescension.
2. Support Community-Led Development: Multilateral institutions should fund and support language communities in creating their own datasets and models (e.g., Masakhane, AI4Bharat, Te Hiku Media). This shifts epistemic authority back to speakers and challenges the logics of appropriation.
3. Implement Participatory Processes: Developers of dominant AI systems must engage minoritized communities as equal partners throughout the AI lifecycle (data collection, training, deployment). This includes fair compensation, respecting data sovereignty, and moving from extractive data practices to non-extractive participation.

The paper concludes that linguistic diversity should be recognized not as a technical challenge to be managed, but as a resource to be protected and sustained for a just AI future. It emphasizes that there is no singular "correct" behavior for language models; rather, the focus must be on who holds the authority to decide how these systems behave.