Linguistic Bias in ChatGPT: How Models Reinforce Dialect Discrimination

TL;DR

ChatGPT defaults to Standard American English (SAE) and retains SAE features far more than non‑standard dialects, by a margin of over 60%.
Non‑standard dialects exhibit consistent biases: stereotyping (19% worse), demeaning content (25% worse), lack of comprehension (9% worse), and condescending responses (15% worse) in GPT‑3.5.
GPT‑3.5 prompted to imitate a dialect worsens stereotyping by 9% and comprehension by 6%.
GPT‑4 improves in warmth, comprehension, and friendliness when imitating input, but stereotyping worsens by 14% for minoritized varieties relative to GPT‑3.5.
British spelling prompts are often answered with American spelling, which may frustrate non‑American users; broader implications include reinforcing linguistic power dynamics as AI tools become more pervasive. For more details, see the Berkeley BAIR blog post: Linguistic Bias in ChatGPT.

Context and background

Researchers at the Berkeley AI Research (BAIR) lab studied how ChatGPT behaves when prompted with ten varieties of English. The study tested two standard varieties—Standard American English (SAE) and Standard British English (SBE)—and eight non‑standard varieties: African‑American English, Indian English, Irish English, Jamaican English, Kenyan English, Nigerian English, Scottish English, and Singaporean English. The central question was whether linguistic features present in the prompt are retained in GPT‑3.5 Turbo responses, and how native speakers judge model outputs along multiple dimensions. Prompts were annotated for linguistic features and spelling conventions (American vs. British). Native speakers rated model responses on positive qualities (warmth, comprehension, naturalness) and negative ones (stereotyping, demeaning content, condescension). The researchers expected SAE to be ChatGPT’s default, given the model’s US origins and the likely prevalence of SAE in training data. The study then compared responses to standard vs. non‑standard varieties to quantify imitation and bias.

What’s new

The model retains SAE features far more than non‑standard dialects, by a margin of over 60%, confirming a strong SAE bias in generated responses.
The model does imitate other dialects, with stronger imitation for varieties spoken by more people (e.g., Nigerian and Indian English) than for varieties with fewer speakers (e.g., Jamaican English). This suggests training data composition influences dialect imitation.
Non‑standard prompts prompt British spellings to default to American spellings in many cases, indicating a bias toward American conventions that can frustrate non‑US users.
Across earlier GPT‑3.5 success metrics, non‑standard varieties provoked higher levels of negative judgments: stereotyping (19% worse), demeaning content (25% worse), lack of comprehension (9% worse), and condescension (15% worse).
When GPT‑3.5 is asked to imitate the input dialect, stereotyping content worsens by about 9% and comprehension worsens by about 6% compared with non‑imitative responses. GPT‑4, while improving warmth, comprehension, and friendliness in imitation, worsens stereotyping by about 14% relative to GPT‑3.5 for minoritized varieties.
The findings indicate that larger, newer models do not automatically eliminate dialect discrimination and may in some cases intensify it.

Why it matters (impact for developers/enterprises)

As AI tools like ChatGPT become more embedded in daily life globally, biased handling of non‑standard dialects can reinforce existing social inequalities and power dynamics, effectively acting as a barrier for speakers of these varieties.
Discrimination linked to language can intersect with race, ethnicity, or nationality, making equitable deployment a pressing fairness concern for developers and enterprises.
Designing models that neutrally understand and respond across dialects may require explicit handling of dialectal variation, user preferences, and prompts that avoid demeaning or stereotyping content.
The study underscores that newer models alone do not solve these challenges; attention to training data composition and response behavior across dialects remains essential for trustworthy AI.

Technical details or Implementation

Methodology: The authors prompted both GPT‑3.5 Turbo and GPT‑4 with text in ten English varieties and compared responses to the standard varieties (SAE and SBE).
Annotation: Prompts and model outputs were annotated for linguistic features and spelling conventions (American vs. British). This helped determine when ChatGPT imitates a variety and what factors influence the degree of imitation.
Evaluation: Native speakers rated model outputs on positive dimensions (warmth, comprehension, naturalness) and negative ones (stereotyping, demeaning content, condescension).
Versions tested: GPT‑3.5 Turbo and GPT‑4, including scenarios where the models were instructed to imitate the input dialect.
Key metrics reported: The model retained SAE features by a margin of over 60% relative to non‑standard varieties; non‑standard responses showed higher stereotyping, demeaning content, and other negative indicators based on native speaker ratings.
Observed spelling behavior: British prompts often elicited American spellings in model outputs, indicating a default to American conventions.

Tables: key findings at a glance

| Aspect | GPT‑3.5 (non‑standard vs standard) | Implication when asked to imitate | GPT‑4 (imitation) notes |---|---|---|---| | Retention of SAE features | >60% more SAE features than non‑standard | Imitation is inconsistent across varieties | Imitation improves warmth and comprehension vs GPT‑3.5 but increases stereotyping for minoritized varieties |Stereotyping in non‑standard prompts | 19% worse than standard | 9% worse when prompted to imitate | Stereotyping worsens by 14% vs GPT‑3.5 for minoritized varieties |Demeaning content | 25% worse | — | — |Comprehension | 9% worse | 6% worse when prompted to imitate | (No specific % given; improvement reported in warmth/comprehension for imitation) |Condescension | 15% worse | — | — |Spelling conventions | British prompts often converted to American | Frustration for non‑US users | — |

Key takeaways

SAE remains the dominant pattern in ChatGPT outputs, with a clear retention advantage over non‑standard dialects.
The training data landscape appears to influence how often the model imitates non‑standard varieties; more widely spoken varieties are more likely to be imitated.
Non‑standard dialect handling in ChatGPT is associated with measurable biases in user ratings, including higher stereotyping and demeaning content.
Prompting the model to imitate a dialect can worsen certain negative behaviors (stereotyping and comprehension) for GPT‑3.5, while GPT‑4 can improve warmth and comprehension yet still raises stereotyping concerns for minoritized varieties.
Spelling conventions in outputs often default to American English, potentially alienating non‑US users.
The findings highlight the need for ongoing fairness and accessibility efforts in AI models as they scale globally.

FAQ

Which dialects were tested?

Standard American English, Standard British English, and eight non‑standard varieties: African‑American English, Indian English, Irish English, Jamaican English, Kenyan English, Nigerian English, Scottish English, and Singaporean English.
What metrics were used to assess quality?

Native speaker ratings of warmth, comprehension, naturalness, and negative dimensions including stereotyping, demeaning content, and condescension.
Does GPT‑4 fix dialect bias?

GPT‑4 shows improvements in warmth, comprehension, and friendliness when imitates input dialects, but stereotyping can still worsen for minoritized varieties compared with GPT‑3.5.
How does spelling handling affect users?

Outputs frequently revert British‑spelling prompts to American spelling, which may frustrate non‑US users.