From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale
Chowdhury Mohammad Abdullah, Rita Orji
cs.CL · cs.AI
Quick verdict
Extends covert LLM dialect-bias probing to Nigerian Englishes across ten models; consistent but small effects and LLM-generated stimuli temper the claims.
Summary
The paper adapts the matched-guise paradigm of Hofmann et al. (2024) to probe covert dialect bias in ten open-weight LLMs across four English varieties (SAE, AAVE, Nigerian Standard English, Nigerian Pidgin) in three housing scenarios of increasing social proximity: tenant screening, neighbor acceptance, roommate selection. Using 260 meaning-matched sentence quadruples (AAVE/SAE pairs from Groenwold et al. 2020; NSE/NP renderings produced by an LLM and validated by two native speakers), the authors score length-normalized log-probabilities of 17-20 housing adjectives per scenario and compute q-scores against the SAE baseline, yielding 582,400 scored observations. AAVE and NP show elevated negative-adjective associations across all ten models (mean q +0.154 and +0.123) with distinct stereotype content: AAVE via unreliability and threat (shady, unstable), NP via corruption and foreignness (foreign +134.6% in the neighbor context). NSE, sharing NP's identity but carrying institutional prestige, is favored over SAE in formal tenant screening yet penalized in closer contexts, implicating prestige as a mechanism independent of racial identity. Mean perplexity correlates with penalization (r = 0.923). Key limitations: LLM-generated Nigerian-variety stimuli validated by only two speakers, statistical tests treating massively correlated observations as independent, very small effects (epsilon-squared up to 0.0026), and a null baseline the signal exceeds by only 1.03-1.51x.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- AAVE and NP are consistently associated with more negative housing-relevant adjectives than semantically identical SAE inputs across all ten evaluated models and all three housing scenarios.
- AAVE and NP are penalized through distinct stereotype clusters (AAVE: unreliability and social threat; NP: corruption and foreignness) rather than a generic non-standard-dialect penalty.
- NSE, sharing NP's national and racial identity but carrying institutional prestige, is favored over SAE in formal tenant screening but penalized as social proximity increases, implying prestige operates as a penalty mechanism independent of identity.
- Per-dialect perplexity tracks penalization (r = +0.923), suggesting training-data underrepresentation of non-standard varieties is connected to covert attitudinal bias.
Strengths
First covert-bias evaluation of Black African postcolonial Englishes
Extends the Hofmann et al. (2024) SAE-AAVE design to NSE and NP (Sec. 1, 3); the NSE-NP pair holds national and racial identity constant while varying institutional prestige, a clean factorial lever for disentangling the two hypothesized penalty pathways that prior work could not separate.
Scale and diversity support the model-agnostic claim
582,400 scored observations across ten open-weight models from Western, European, Middle Eastern, and East Asian training regimes (Sec. 3.3-3.4); per-model consistency of 9-10 out of 10 models (Sec. 4.5, Fig. 5) shows the effect is systemic rather than an artifact of one model family.
Stereotype-content analysis goes beyond aggregate scores
Unusually candid limitations and null baselines
Weaknesses
LLM-generated minority-dialect stimuli introduce circularity
Pseudoreplication makes the significance testing uninformative
Thin margin over the null-attribute baseline
Internal inconsistency on which dialect is penalized most severely
Headline perplexity-bias correlation rests on four data points
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the 'Shibboleth problem' of AI according to this paper?
The paper borrows the biblical shibboleth episode, where pronunciation alone decided who was excluded at a river crossing, to name how AI systems can use dialect features as markers to exclude people regardless of what they actually say. It argues that LLM probability distributions encode such judgments invisibly, so downstream screening systems can discriminate by dialect without any discriminatory intent on the operator's part.
Did the study find that LLMs are biased against Nigerian English?
It found a split. Nigerian Pidgin was consistently linked to more negative housing adjectives than Standard American English across all ten models, mainly via corruption and foreignness stereotypes. Nigerian Standard English, the institutionally prestigious variety, was actually favored over Standard American English in formal tenant screening but penalized in more personal contexts (neighbor, roommate), suggesting prestige rather than race alone drives the penalties.
How did the researchers measure covert dialect bias in LLMs?
They used a matched-guise design: 260 sentences with identical meanings written in four English varieties (Standard American English, AAVE, Nigerian Standard English, Nigerian Pidgin) were fed to ten open-weight models. For each sentence they measured the log-probability the model assigned to positive and negative housing-related adjectives, then compared each dialect against the Standard American English version of the same sentence (q-scores), so any difference is attributable to dialect alone since content is held constant.
What are the main limitations of this dialect bias study?
The Nigerian-variety sentences were generated by an LLM and checked by only two native speakers, so the stimuli may reflect model-internal caricatures of these dialects. The statistical tests treat highly correlated observations as independent, making p-values uninformative at this scale. Effect sizes are small and only modestly above a trivial-attribute null baseline, and probability-level bias is a necessary but not sufficient indicator of real-world discriminatory outcomes in deployed systems.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are present only as captions in the extracted text; heatmap values, bar magnitudes, and per-model patterns in Fig. 5 could not be independently verified.
- Prompt templates, adjective lexicons, the translation protocol, native-speaker agreement scores, and the null-baseline numbers are all deferred to the external GitHub repository and could not be checked from the text.
- The abstract supplied in the metadata is truncated mid-sentence; this review is based on the full extracted text, which otherwise appears complete through the references.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.