5. TECH-DRIVEN SOLUTIONS AND EMERGING TECHNOLOGIES TO COUNTER DISINFORMATION

NLP tools used for combating disinformation are highly dependent on the quality of the training data or in other words, “with limited human direction, an artificial agent is as good as the data it learns from” (Osoba et al., 2017, p.17). This would also mean that, if the data used for training is biased, the automated tool will reproduce these biases, or according to Raso et al. will exacerbate them (Raso et al., 2018). There are several stages in which biases can be introduced in the dataset. In the majority of the cases, bias is introduced during the data collection process, specifically during the annotation process. Individuals building the corpus can integrate their judgement when defining annotations by deciding “what specific type of speech and demographic groups, and so on are prioritized in the training data” (New America, 2020).

An example of dataset bias, that specifically targets one demographic group can be found in a case study put forth by the European Union Agency for Fundamental Rights (FRA). In this particular example, the automatic system indicated a correlation between text being labelled as offensive when written in the African American English dialect, proving that content may be misclassified based on the expressions certain ethnic groups are using (FRA, 2022, p. 69).  

Machine-learning NLP tools cannot parse text in all languages. Given the fact that there are hundreds of languages in the world, machine-learning NLP tools will be effective in the case of high-resource languages (HRLs) such as English Spanish, German, and Chinese, while their accuracy will be significantly lower in the case of low-resource languages (LRLs) such as Bengali, Punjabi, Indonesian, although these languages are spoken by millions of people (Hirschberg, Manning, 2016).

This would consequently mean that the machine-learning NLP used for detecting disinformation could have “disproportionately harmful outcomes for non-English speakers”, especially if the outcome of the machine analysis is to be used as part of a decision-making process (Duarte et al, 2018). According to Xiao Mina, there are a billion people only in Asia who speak thousands of languages and who on the one hand cannot actively participate in conversations in the online world due to language barriers or on the other hand, can experience their posts being flagged as disinformation due to the same language barriers. This would also mean that machine-learning NLP tools would struggle with detecting disinformation in LRLs (Xiao Mina, 2015).

A recently published report by FRA shows how a seemingly neutral corpus covering HRLs, can be biased. During their research, FRA developed several algorithms for offensive speech detection for different languages; English, German and Italian and have found that “for example, in English, the use of terms alluding to ‘Muslim’, ‘gay’ or ‘Jew’ often lead to predictions of generally non-offensive text phrases as being offensive. In the German-language algorithms developed for this report, the terms ‘Muslim’, ‘foreigner’ and ‘Roma’ most often lead to predictions of text as being offensive despite being non-offensive. In the Italian-language algorithms, the terms ‘Muslims’, ‘Africans’, ‘Jews’, ‘foreigners’, ‘Roma’ and ‘Nigerians’ trigger overly strong predictions in relation to offensiveness” (FRA, 2022, p. 11). 

The accuracy of these NLP tools is significantly lowered in cases where the context can completely transform the meaning of the claim which the system could mark as disinformation. Since these tools interact with an environment they might not be familiarized with, there is a higher error probability. As Asudeh et al., explain, one may claim that “[she] has never lost a game of chess” which can be truthful information for an experienced chess player, but also for someone who has never played chess (Asudeh et al., 2020).

NLP tools have a difficulty distinguishing whether similar claims are truthful or not, since they are entirely context dependent. Furthermore, machine-learning NLP tools experience difficulties in detecting “context, subtlety, sarcasm, and subcultural meaning” (Gillespie, 2020, p. 3). These difficulties were shown in the case of the machine-learning NLP tool used by YouTube which misclassified 150 000 videos as disinformation (Vincent, 2020).

In the words of Frank Pasquale, we live in the Black Box Society in which “hidden algorithms can make (or ruin) reputations, decide the destiny of entrepreneurs, or even devastate an entire economy” (Pasquale, 2016). Given that automated processes such as NLP tools can almost autonomously manage online behaviour as well as enforce rights it is important to apply scrutiny over these and other similar technological solutions.

There is often little to no information on how automated tools make correlations or decisions, nor is there any data on their accuracy and reliability. According to Perel et al., “as passive, transparency-driven observations of algorithmic enforcement systems are limited in their capacity to check the practices of non-transparent, constantly evolving algorithms, it is essential to encourage the active engagement of the public in challenging unknown and possibly biased systems…” (Perel, et al., 2017, p.41).

As a response to the lack of transparency significant investment has been made in the field of Explainable Artificial Intelligence (XAI) – “a field focused on the understanding and interpretation of the behaviour of AI systems” (Linardatos, 2020, p. 2). Linardatos et. al identified four different methods for explaining algorithms: “methods for explaining complex black-box models, methods for creating white-box models, methods that promote fairness and restrict the existence of discrimination, and, lastly, methods for analysing the sensitivity of model predictions” (Linardatos, 2020, p. 5). Methods for explaining the black-box models aim to interpret already developed complex models, such as deep neural networks (Linardatos, 2020).

The white-box models create understandable models which include a decision-tree, making it easier to follow the development process. Methods that promote fairness and restrict the existence of discrimination aim to detect inequities or discrimination which can be promoted by the algorithm. This method is applied with 3 primary goals in mind: controlling discrimination, limiting distortion in individual instances, and preserving utility (Linardatos, 2020, p.19). Lastly, methods for analysing the sensitivity of model predictions aim to ensure that predictions made by the algorithm are trustworthy and reliable (Linardatos, 2020, p. 26).