Learnijoy
ResearchAI ResearchAI Engineering & DevToolsAI News & Tools

Autonomous Vehicles Inherit Human Biases in Pedestrian Yielding Decisions

Irem Yoldas, Martim Brand\~ao, Jie Zhang, Odinaldo Rodrigues· September 2, 2026 View original

Key takeaways

  • LLM-driven autonomous vehicles can inherit and exhibit human-like biases in critical decisions like pedestrian yielding.
  • Bias in AVs is influenced by pedestrian demographics, raising ethical and fairness concerns.
  • New methodologies are needed to rigorously test and identify biases in AI models for autonomous systems.
  • Addressing these biases is crucial for public trust, ethical deployment, and regulatory acceptance of AVs.

Who benefits

AutomotiveAI DevelopmentRegulatory BodiesUrban PlanningInsurance

Summary

This research reveals that LLM and VLM-driven autonomous vehicles inherit human driver biases, showing varied pedestrian-yielding rates based on factors like gender, ethnicity, and socio-economic status. The study introduces new bias testing methodologies and questions the "common sense" model paradigm for AVs without addressing downstream bias.

A new study highlights a critical issue for autonomous vehicles (AVs): the potential for AI systems, particularly those driven by Large Language Models (LLMs) and Visual-Language Models (VLMs), to inherit and perpetuate human biases in decision-making. Researchers investigated how these models make pedestrian-yielding decisions, revealing that factors such as a pedestrian's gender, ethnicity, religion, disability, age, skin tone, and socio-economic status can influence the AV's behavior. To assess this, the paper proposes two novel bias testing methodologies: "All Else Being Equal" tests and "Self-Consistency" tests. These methods were applied to various LLMs and VLMs, consistently demonstrating that biases exist, although their specific nature and degree differ across models. The findings raise significant concerns about the current paradigm of using general-purpose "common sense" models to guide AV decisions. The authors argue that a thorough analysis of model bias must be an integral part of AV evaluation, emphasizing the need to either fundamentally revise this paradigm or implement robust strategies to mitigate these inherited biases to ensure fair and equitable outcomes.

Why it matters

Professionals in AI development, automotive, and policy must address inherited biases in AVs to ensure public trust, ethical deployment, and regulatory compliance, as biased decisions could lead to significant societal and legal repercussions.

How to implement this in your domain

  1. 1Integrate bias testing methodologies, like "All Else Being Equal" and "Self-Consistency" tests, into the AV development lifecycle.
  2. 2Develop and apply debiasing techniques to LLMs and VLMs used in AV decision-making to mitigate discriminatory outcomes.
  3. 3Establish diverse and representative datasets for training and validation to reduce the perpetuation of human biases.
  4. 4Collaborate with ethicists and social scientists to understand and address the societal implications of AI biases in autonomous systems.
  5. 5Advocate for industry standards and regulations that mandate bias assessment and mitigation in AV technology.

Original post by Irem Yoldas, Martim Brand\~ao, Jie Zhang, Odinaldo Rodrigues

"arXiv:2609.00192v1 Announce Type: new Abstract: Public trust in Autonomous Vehicles (AVs) may depend not only on technical success but also on the fairness of their decision making. While a recent trend in AV research involves using general purpose "common sense" models to guide…"

View on X

Originally posted by Irem Yoldas, Martim Brand\~ao, Jie Zhang, Odinaldo Rodrigues on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI Engineering & DevToolsAI Research

Subspace Levenberg-Marquardt Algorithms Boost Neural Network Training

This research evaluates subspace Levenberg-Marquardt (LM) algorithms, such as KSLM and HSLM, for training neural networks on regression and classification tasks. These methods address the high computational and memory costs of classical LM, offering more efficient second-order optimization compared to first-order methods like SGD and Adam.

M. Duc HoangSep 2, 2026
AI ResearchAI Engineering & DevTools

Neural Networks Show Varied Conceptual Separation Internally

A study examined "conceptual separation" in CNNs and LLMs, analyzing how internal activations represent concepts. It found that CNNs form coherent representations for familiar concepts, while LLMs show clear separation for distinct domains but collapse distinctions for ambiguous topics.

Jaee Ponde, Roshni Agarwal, Subhashis BanerjeeSep 2, 2026
AI Engineering & DevToolsAI Research

FCCA Boosts Low-Resource LLM Adaptation with Fewer Parameters

Researchers introduced Frozen-Core Cross-Covariance Adaptation (FCCA), a new parameter-efficient fine-tuning method that significantly improves LLM performance with a minimal trainable parameter budget. FCCA achieves results comparable to LoRA and DoRA while optimizing orders of magnitude fewer parameters by carefully selecting fixed core coordinates.

Wentao Ye, Zhanming Shen, Zhiqing Xiao, Yao Ding, Haobo Wang, Gang ChenSep 2, 2026