Automated Journalism as a Source of and a Diagnostic Device for Bias in Reporting

Abstract

<jats:p>In this article we consider automated journalism from the perspective of bias in news text. We describe how systems for automated journalism could be biased in terms of both the information content and the lexical choices in the text, and what mechanisms allow human biases to affect automated journalism even if the data the system operates on is considered neutral. Hence, we sketch out three distinct scenarios differentiated by the technical transparency of the systems and the level of cooperation of the system operator, affecting the choice of methods for investigating bias. We identify methods for diagnostics in each of the scenarios and note that one of the scenarios is largely identical to investigating bias in non-automatically produced texts. As a solution to this last scenario, we suggest the construction of a simple news generation system, which could enable a type of analysis-by-proxy. Instead of analyzing the system, to which the access is limited, one would generate an approximation of the system which can be accessed and analyzed freely. If successful, this method could also be applied to analysis of human-written texts. This would make automated journalism not only a target of bias diagnostics, but also a diagnostic device for identifying bias in human-written news.</jats:p>

Bias of AI-generated content: an examination of news produced by large language models

“…Last, our study highlights that AIGC of large language models exhibits gender and racial biases. Further investigation is needed to explore other types of biases, such as position bias 14 and word-specific biases 15 . Drawing inspiration from the literature proposing debiasing methods 16 , 17 , it could be most effective to train LLMs with unbiased training data.…”

“…It is important to note that bias is not a unique challenge limited to AIGC. Within journalism research 14 , bias is a prevalent issue observed in news articles. An illustrative example is when both a man and a woman contribute equally to an event, yet prior research reveals a tendency in news reporting to place the man’s contribution first.…”

The CACAPO Dataset: A Multilingual, Multi-Domain Dataset for Neural Pipeline and End-to-End Data-to-Text Generation

“…Bias The fact that the CACAPO dataset is based on 'naturally occurring' data addresses the issue of datasets being not representative of real world NLP issues. However, it should also be noticed that having unedited texts in the dataset means that the biases from the original data are still present in the dataset and may lead to further generation of biased texts (Leppänen et al, 2020). Therefore, texts generated with this dataset, as well as the texts in the dataset itself, could warrant more traditional linguistics-oriented text analysis research to investigate biases that might exist.…”

Educación contra el prejuicio: caso aplicado al Periodismo

“…El artículo séptimo del Código Deontológico de la Federación de Asociaciones de la Prensa de España (FAPE) exige a los periodistas "abstenerse de aludir, de modo despectivo o con prejuicio, a la raza, color, religión, origen social o sexo de una persona o a cualquier enfermedad, minusvalía física o mental que padezca". En el ámbito académico se han propuesto herramientas de corrección relacionadas con redes sociales (Carlson, 2018;Usher et al, 2018), periodismo de datos (Jamali, 2020) o periodismo automatizado (Leppänen et al, 2020). También se ha destacado la importancia de la alfabetización mediática (Buckingham, 2019;Hamborg et al, 2019;Tully et al, 2020;Waddell, 2019) como mecanismo capaz de neutralizar o atenuar los efectos del prejuicio en la comunicación periodística.…”