Nationality Bias in Text Generation

Narayanan, Venkit, Pranav; Gautam, Sanjana; Panchanadikar, Ruchi; Ting-Hao,; Huang, Biao; Wilson, Shomir

doi:10.48550/arxiv.2302.02463

Cited by 3 publications

(4 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A UNESCO policy report (West et al, 2019) highlighted the problematic ways that AI perpetuates gender biases, such as using feminized voices in virtual assistants (e.g., Amazon's Alexa, Apple's Siri) that reinforces a submissive or obliging stereotype. Along with racism and sexism, AI may also perpetuate biases based on religion (Abid et al, 2021), nationality, or disability (Venkit et al, 2023).…”

Section: Ai Bias Reliability and Accuracymentioning

confidence: 99%

Artificial Intelligence in Practice: Opportunities, Challenges, and Ethical Considerations

Farmer,

Lockwood,

Goforth

et al. 2024

Preprint

View full text Add to dashboard Cite

This project explores the rapidly evolving integration of Artificial Intelligence (AI) into psychological practice, focusing on health service psychologists. The study critically examines the potential, limitations, and ethical/legal considerations of AI use in this field.Artificial Intelligence, particularly Large Language Models (LLMs), presents both promising applications and significant ethical concerns for psychological practice. This paper aims to provide a balanced view of AI's integration into psychology, advocating for careful oversight, ethical practices, and legal compliance to harness AI's potential responsibly. The manuscript details AI's practical uses in administrative tasks, therapeutic services, and enhancing service provision, while also addressing the potential risks such as bias introduction, deskilling, and privacy concerns. Special attention is given to the ethical and legal challenges posed by AI technologies, emphasizing the need for ongoing evaluation and updating of ethical codes and legal frameworks to align with technological advancements. This project raises awareness among psychologists about the complexities of AI application in their practice, urging a cautious yet optimistic approach to integrating these technologies. It serves as a call to action for the psychological community to actively engage in shaping the future of AI in their field, ensuring that advancements improve service quality without compromising ethical standards or client welfare.

show abstract

Section: Ai Bias Reliability and Accuracymentioning

confidence: 99%

Artificial Intelligence in Practice: Opportunities, Challenges, and Ethical Considerations

Farmer,

Lockwood,

Goforth

et al. 2024

Preprint

View full text Add to dashboard Cite

show abstract

“…Studies are starting to investigate how cultural biases are expressed and propagated through LLMs. They highlight that language models tend to mimic the majority viewpoints found on the Internet (Venkit et al, 2023), misrepresenting those of minorities, possibly due to the use of English training data and prompts that could reduce the variability in model responses (Cao et al, 2023;Naous et al, 2023). Further evidence comes from studies that compare multi-and mono-lingual LLMs.…”

Section: Llms Are Culturally Biasedmentioning

confidence: 99%

“…Previous studies show that stereotypes and cultural biases are inherently embedded in language (Marinucci, Mazzuca, Gangemi, 2023). LLMs designers aim for universal solutions but, despite the efforts in debiasing, actual models still engage in stereotypes (Venkit et al, 2023). Having a large amount of training data does not guarantee diversity and training data need to be de-biased from sets of potentially harmful words.…”

Section: Llms Do Not Capture Individual Characteristicsmentioning

confidence: 99%

Exploring The Prospects And Challenges Of Large Language Models For Language Learning And Production

Borghi,

De Livio,

Mannella

et al. 2023

Preprint

View full text Add to dashboard Cite

The success of Large Language Models (LLMs) in many application domains suggests that they may also change how we conceive cognition. LLMs possess capabilities traditionally considered exclusively human. Can the experience with language alone facilitate the acquisition of other complex cognitive abilities? Do linguistic, sensorimotor, and interoceptive experiences need to be integrated? Are there domains, like that of abstract concepts (e.g., freedom), where linguistic experience suffices to capture meaning? After introducing what LLMs are, we address their potential impact, discussing five differences from human cognition: they are not grounded, lack action, hardly capture pragmatics, are culturally biased, and do not reflect individual characteristics.

show abstract

“…Regarding linguistics in particular, the abovementioned stochastic biases imply that the lexical subtleties and overall diversity found in language generated by these models will be poor. Concerns have been raised about how human biases present in the input of a model can lead to direct, indirect, or intersectional discrimination (Abid et al 2021;Venkit et al 2023), and engender AI biases after the learning phase (Wirtz et al 2019). Such discrimination is visible in the development of sexist stereotypes when AI is trained on uncontrolled open access data, where men are described as maestros and women as homemakers (Caliskan et al 2017).…”

Section: Diversitymentioning

confidence: 99%

Minds

Backus,

Cohen,

Cohn

et al. 2023

AVT

View full text Add to dashboard Cite

It is an interesting time to be a linguist. The advent of large language models (LLMs) like GPT-4 has raised fundamental questions about language and its nature, such as whether artificial systems are able to "use" language in a similar way to humans. While such issues are at the core of ongoing scientific and societal debates, the role of linguistics in the development of these technologies has been surprisingly limited. LLMs are suddenly widespread because of the opportunities they provide, but also come with several risks. Schools are therefore regulating their use, and some countries-like Italy-even prohibited them. LLMs can be wrong, for example, in the sense that they can "hallucinate" and come up with structurally correct, but false statements (Ji et al. 2023). This behavior is due to the way in which these models learn to use language; not by assigning meaning to form but rather by learning statistical regularities about how words and sentences typically co-occur. As a result, the 'meaning' that these models assign to language is generally not grounded in any experience with the world. 1 In other words, LLMs can be described as 'stochastic parrots' that semirandomly reproduce linguistic forms without actually being related to the outside world (Bender et al. 2021). 2 We believe linguists can pick up a much larger role in these discussions clarifying how LLMs could be adapted to become more similar

show abstract

Nationality Bias in Text Generation

Cited by 3 publications

References 0 publications

Artificial Intelligence in Practice: Opportunities, Challenges, and Ethical Considerations

Artificial Intelligence in Practice: Opportunities, Challenges, and Ethical Considerations

Exploring The Prospects And Challenges Of Large Language Models For Language Learning And Production

Minds

Contact Info

Product

Resources

About