LLM-Driven Autonomous Vehicles Inherit Human Driver Biases in Pedestrian Yielding: Results and Implications From A New Benchmark
Irem Yoldas, Martim Brandão, Jie Zhang, Odinaldo Rodrigues
cs.AI · cs.CL · cs.CV · cs.CY
Quick verdict
Finds LLM/VLM AV yielding varies with pedestrian demographics via AEBE/SC audits; sound open benchmark, but uncorrected chi-square tests inflate bias claims.
Summary
The paper proposes two bias-audit methodologies for AV-oriented language models: 'All Else Being Equal' (AEBE) tests for LLMs, comparing yielding decisions across text scenarios identical except for an injected pedestrian attribute, and 'Self-Consistency' (SC) tests for VLMs, correlating a model's own demographic estimates with its own stop decisions on frames it judges equivalent. From 23,812 single-pedestrian frames in nuImages, JAAD and PIE, the authors build 3,157 demographic-scrubbed text scenarios and expand them into up to 176,792 conditioned variants per category (gender, ethnicity, religion, disability, age, skin tone, socio-economic status, plus gender intersections; Table I). Four LLMs (Qwen-3-8B, Llama-3.1-8B, Mistral-7B, GPT-4o) and four VLMs (Qwen-2.5-VL, Qwen-3-VL, LLaVA-NeXT, SPHINX) were evaluated zero-shot via logit-based decisions; GPT-4o was excluded from VLM tests for refusing demographic estimation. Reported findings: all models show significant yielding associations, with disability most consistent ('paralyzed' lowest everywhere); Qwen-3 varies most (median stop ~70%; female 61% vs 92% without gender info); GPT-4o gaps are significant but tiny (Jewish 99.5% vs Muslim 99.0%); VLM yield medians span 1-30%; Qwen-2.5-VL favors lighter skin (fair white 7% vs dark 0%), echoing US human-driver studies; LLaVA-NeXT yields 50% to 'Indian' pedestrians versus under 4% otherwise. Main limitations: no multiple-comparison correction or effect sizes across thousands of pairwise chi-square tests; SC tests conflate perception and decision via self-estimated attributes; category taxonomies are inconsistent between methods and results sections.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Both LLMs and VLMs make yielding decisions statistically associated with pedestrian gender, ethnicity, religion, disability, age, skin tone and socio-economic status, with disability the most consistent dimension across models.
- The AEBE and SC methodologies constitute the first benchmark for auditing yielding bias in AV-oriented language models.
- A skin-tone yielding gradient in Qwen-2.5-VL (higher yields to lighter skin) is consistent with documented US human driver bias.
- Bias type and degree differ per model, with Qwen-3 the most variable LLM and LLaVA-NeXT the most conservative VLM (1% median yield).
- These findings challenge the 'common sense' model paradigm for AV decision making unless downstream bias is addressed.
Strengths
Causally clean AEBE design with disclosed scrubbing
Text scenarios identical except injected demographics; manual removal of leaked gender mentions (55 female, 203 male) from Qwen-3-VL descriptions; deliberate inclusion of potentially offensive labels ('Normal', 'Old') to separate word-choice effects from attribute effects (Sec. III-B).
Dual-channel coverage of text and vision models
AEBE exploits controlled text edits for LLMs; SC gates VLM comparisons on model-estimated intention and need-to-stop, a pragmatic answer to the impossibility of pixel-identical image pairs differing only in demographics (Sec. III-C).
Scale and intersectional depth
Transparency and honesty about test limits
Novel audit target with safety relevance
Weaknesses
No multiple-comparison control or effect sizes
SC test circularity
Title overclaims inheritance of human bias
Taxonomy inconsistencies between method and results
Narrow model and prompt scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
King's College London authors propose two bias-audit methodologies for LLM/VLM-driven AV decisions: **All Else Being Equal (AEBE)** tests, comparing yielding decisions across text scenarios identical except for injected pedestrian demographics, and **Self-Consistency (SC)** tests, correlating a VLM's own demographic estimates with its own stop decisions on frames it judges equivalent. From 23,812 single-pedestrian frames (nuImages, JAAD, PIE), they derive 3,157 demographic-scrubbed text scenarios and expand them into up to 176,792 gender x disability variants (Table I). Central claims: (1) all eight evaluated models show statistically significant yielding associations with gender, ethnicity, religion, disability, age, skin tone and SES; (2) disability is the most consistent bias ('paralyzed' lowest across all models); (3) the 'common sense' paradigm needs revision or downstream debiasing. The title's 'inherit human driver biases' is stronger than the evidence: only the Qwen-2.5-VL skin-tone gradient (fair white 7% vs dark 0% yield) maps onto the cited human literature (Goddard et al., 32% longer waits for Black pedestrians); most model biases — e.g., Qwen-3 yielding least to women (61.1% vs 91.6% when gender is unstated, Fig. 3) — have no human-driver analogue and contradict each other across models.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 'All Else Being Equal' (AEBE) test in this paper?
A bias audit for text-based LLMs: the same driving-scenario text is repeated with only the pedestrian's demographic description changed (e.g., 'The pedestrian is male' vs 'female'). Since everything else is identical, differences in the model's yield/stop decision are attributed to the demographic attribute. The authors built 3,157 demographic-free scenario descriptions from real driving datasets and expanded them into over half a million conditioned variants.
Did the LLMs reproduce exactly the same biases as human drivers?
Not exactly. Only one pattern matched documented human behavior: Qwen-2.5-VL yielded more to lighter-skinned pedestrians (7% for fair white vs 0% for dark skin), echoing US studies where drivers yield less to Black pedestrians. Other biases diverged and even contradicted each other across models — Qwen-3 yielded least to female pedestrians (61%) while Mistral yielded most to them (98%). Disability groups, especially 'paralyzed', received the lowest yields across all models.
Which models showed the most bias in pedestrian yielding?
Among LLMs, Qwen-3 (8B) had the most variable decisions (median stop rate ~70%; female 61% vs 92% when gender is unstated). Among VLMs, LLaVA-NeXT almost never yielded (1% median) but spiked to 50% for 'Indian' pedestrians versus under 4% for other groups. GPT-4o's differences were statistically significant but tiny (religion gaps of about 0.5 percentage points).
Is the benchmark and data from this paper publicly available?
The authors state the full per-model, per-demographic figures and the benchmark are available at github.com/IremYoldas/LLM-Driven-AV-Pedestrian-Bias, built on the public nuImages, JAAD and PIE datasets. Exact demographic injection templates and the GPT-4o API version are not specified in the paper itself.
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.
- Figure percentage labels (Figs. 3-11) were extracted in jumbled/reversed order by the PDF parser; all numbers cited in this review are taken from the running text, which matches the narrative.
- Reference [14] appears incomplete in the extraction (headline cut after 'Black people wait 32'); the original bibliography entry could not be verified.
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.