Cultural Bias in Large Language Models: A Comprehensive Analysis and Mitigation Strategies DOI
Z Y Liu

Journal of Transcultural Communication, Journal Year: 2024, Volume and Issue: unknown

Published: Sept. 16, 2024

Abstract This paper delves into the intricate relationship between Large Language Models (LLMs) and cultural bias. It underscores significant impact LLMs can have on shaping a more equitable culturally sensitive digital landscape, while also addressing challenges that arise when integrating these powerful AI tools. The emphasizes immense significance of in contemporary research applications, underpinning many systems algorithms. However, their potential role perpetuating or mitigating bias remains pressing issue warranting extensive analysis. Cultural stems from various intertwined factors; following analysis categorizes three dimensions: data quality, algorithm design, user interaction dynamics. Furthermore, impacts identity linguistic diversity are scrutinized, highlighting interplay technology culture. advocates responsible development, outlining mitigation strategies such as ethical guidelines, diverse training data, feedback mechanisms, transparency measures. In conclusion, is not solely problem but presents an opportunity. enhance our awareness critical understanding own biases fostering curiosity respect for perspectives.

Language: Английский

Philosophical Transactions of the Royal Society of London. Series A, Mathematical and Physical Sciences DOI

Jairo F. Gudiño,

Umberto Grandi, César A. Hidalgo

et al.

Published: Dec. 22, 2019

We explore the capabilities of an augmented democracy system built on off-the-shelf LLMs fine-tuned data summarizing individual preferences across 67 policy proposals collected during 2022 Brazilian presidential election.We use a train-test cross-validation setup to estimate accuracy with which predict both: subject's political choices and aggregate full sample participants.At level, out predictions lie in range 69%-76% are significantly better at predicting liberal college educated population we using adaptation Borda score compare ranking obtained from probabilistic participants LLMs.We find that predicts than samples alone when these represent less 30% 40% total population.These results indicate potentially useful for construction systems democracy.

Language: Английский

Citations

275
A survey on large language model (LLM) security and privacy: The Good, The Bad, and The Ugly DOI Creative Commons
Yifan Yao, Jinhao Duan, Kaidi Xu

et al.

High-Confidence Computing, Journal Year: 2024, Volume and Issue: 4(2), P. 100211 - 100211

Published: March 1, 2024

Language: Английский

Citations

222
Bias in medical AI: Implications for clinical decision-making DOI Creative Commons
James M. Cross,

Michael A. Choma,

John A. Onofrey

et al.

PLOS Digital Health, Journal Year: 2024, Volume and Issue: 3(11), P. e0000651 - e0000651

Published: Nov. 7, 2024

Biases in medical artificial intelligence (AI) arise and compound throughout the AI lifecycle. These biases can have significant clinical consequences, especially applications that involve decision-making. Left unaddressed, biased lead to substandard decisions perpetuation exacerbation of longstanding healthcare disparities. We discuss potential at different stages development pipeline how they affect algorithms Bias occur data features labels, model evaluation, deployment, publication. Insufficient sample sizes for certain patient groups result suboptimal performance, algorithm underestimation, clinically unmeaningful predictions. Missing findings also produce behavior, including capturable but nonrandomly missing data, such as diagnosis codes, is not usually or easily captured, social determinants health. Expertly annotated labels used train supervised learning models may reflect implicit cognitive care practices. Overreliance on performance metrics during obscure bias diminish a model's utility. When applied outside training cohort, deteriorate from previous validation do so differentially across subgroups. How end users interact with deployed solutions introduce bias. Finally, where are developed published, by whom, impacts trajectories priorities future development. Solutions mitigate must be implemented care, which include collection large diverse sets, statistical debiasing methods, thorough emphasis interpretability, standardized reporting transparency requirements. Prior real-world implementation settings, rigorous through trials critical demonstrate unbiased application. Addressing crucial ensuring all patients benefit equitably AI.

Language: Английский

Citations

22
ReelFramer: Human-AI Co-Creation for News-to-Video Translation DOI Open Access
Sitong Wang, Samia Menon, Tao Long

et al.

Published: May 11, 2024

Short videos on social media are the dominant way young people consume content. News outlets aim to reach audiences through news reels—short conveying news—but struggle translate traditional journalistic formats into short, entertaining videos. To reels, we support journalists in reframing narrative. In literature, narrative framing is a high-level structure that shapes overall presentation of story. We identified three framings for reels adapt norms but preserve value, each with different balance information and entertainment. introduce ReelFramer, human-AI co-creative system helps print articles scripts storyboards. ReelFramer supports exploring multiple find one appropriate AI suggests foundational details, including characters, plot, setting, key information. also visual framing; character detail designs before generating full storyboard. Our studies show introduces necessary diversity various establishing details generate more relevant coherent. discuss benefits using content retargeting.

Language: Английский

Citations

16
Rethinking machine unlearning for large language models DOI
Sijia Liu,

Yuanshun Yao,

Jinghan Jia

et al.

Nature Machine Intelligence, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 17, 2025

Language: Английский

Citations

4
How understanding large language models can inform the use of ChatGPT in physics education DOI Creative Commons
Giulia Polverini, Bor Gregorcic

European Journal of Physics, Journal Year: 2023, Volume and Issue: 45(2), P. 025701 - 025701

Published: Dec. 11, 2023

Abstract The paper aims to fulfil three main functions: (1) serve as an introduction for the physics education community functioning of large language models (LLMs), (2) present a series illustrative examples demonstrating how prompt-engineering techniques can impact LLMs performance on conceptual tasks and (3) discuss potential implications understanding prompt engineering teaching learning. We first summarise existing research popular LLM-based chatbot (ChatGPT) tasks. then give basic account work, illustrate essential features their functioning, strengths limitations. Equipped with this knowledge, we some challenges generating useful output ChatGPT-4 in context introductory physics, paying special attention questions problems. provide condensed overview relevant literature demonstrate through selected be employed improve ’s Qualitatively studying these provides additional insights into ChatGPT’s its utility problem-solving. Finally, consider from inform use learning physics.

Language: Английский

Citations

38
Fluent but Not Factual: A Comparative Analysis of ChatGPT and Other AI Chatbots’ Proficiency and Originality in Scientific Writing for Humanities DOI Creative Commons
Edisa Lozić, Benjamin Štular

Future Internet, Journal Year: 2023, Volume and Issue: 15(10), P. 336 - 336

Published: Oct. 13, 2023

Historically, mastery of writing was deemed essential to human progress. However, recent advances in generative AI have marked an inflection point this narrative, including for scientific writing. This article provides a comprehensive analysis the capabilities and limitations six chatbots scholarly humanities archaeology. The methodology based on tagging AI-generated content quantitative accuracy qualitative precision by experts. Quantitative assessed factual correctness manner similar grading students, while gauged contribution reviewing article. In test, ChatGPT-4 scored near passing grade (−5) whereas ChatGPT-3.5 (−18), Bing (−21) Bard (−31) were not far behind. Claude 2 (−75) Aria (−80) much lower. all chatbots, but especially ChatGPT-4, demonstrated proficiency recombining existing knowledge, failed generate original content. As side note, our results suggest that with size large language models has reached plateau. Furthermore, paper underscores intricate recursive nature research. process transforming raw data into refined knowledge is computationally irreducible, highlighting challenges face emulating originality Our apply state affairs third quarter 2023. conclusion, revolutionised generation, their ability produce contributions remains limited. We expect change future as current model-based evolve model-powered software.

Language: Английский

Citations

34
(A)I Am Not a Lawyer, But...: Engaging Legal Experts towards Responsible LLM Policies for Legal Advice DOI Creative Commons
Inyoung Cheong, King Xia, K. J. Kevin Feng

et al.

2022 ACM Conference on Fairness, Accountability, and Transparency, Journal Year: 2024, Volume and Issue: 67, P. 2454 - 2469

Published: June 3, 2024

Large language models (LLMs) are increasingly capable of providing users with advice in a wide range professional domains, including legal advice. However, relying on LLMs for queries raises concerns due to the significant expertise required and potential real-world consequences To explore when why should or not provide users, we conducted workshops 20 experts using methods inspired by case-based reasoning. The provided realistic ("cases") allowed examine granular, situation-specific overarching technical constraints, producing concrete set contextual considerations LLM developers. By synthesizing factors that impacted response appropriateness, present 4-dimension framework: (1) User attributes behaviors, (2) Nature queries, (3) AI capabilities, (4) Social impacts. We share experts' recommendations strategies, which center around helping identify 'right questions ask' relevant information rather than definitive judgments. Our findings reveal novel considerations, such as unauthorized practice law, confidentiality, liability inaccurate advice, have been overlooked literature. deliberation method enabled us elicit fine-grained, practice-informed insights surpass those from de-contextualized surveys speculative principles. These underscore applicability our translating domain-specific knowledge practices into policies can guide behavior more responsible direction.

Language: Английский

Citations

15
Natural language processing in the era of large language models DOI Creative Commons
Arkaitz Zubiaga

Frontiers in Artificial Intelligence, Journal Year: 2024, Volume and Issue: 6

Published: Jan. 12, 2024

SPECIALTY GRAND CHALLENGE article Front. Artif. Intell., 12 January 2024Sec. Natural Language Processing Volume 6 - 2023 | https://doi.org/10.3389/frai.2023.1350306

Language: Английский

Citations

13
Utilizing large language models in breast cancer management: systematic review DOI Creative Commons
Vera Sorin, Benjamin S. Glicksberg, Yaara Artsi

et al.

Journal of Cancer Research and Clinical Oncology, Journal Year: 2024, Volume and Issue: 150(3)

Published: March 19, 2024

Despite advanced technologies in breast cancer management, challenges remain efficiently interpreting vast clinical data for patient-specific insights. We reviewed the literature on how large language models (LLMs) such as ChatGPT might offer solutions this field.

Language: Английский

Citations

12