5.3 Limits of Technological Tools, Best Practices

Ana Ćuća. Aitana Radu


The first part of this section is dedicated to a short introduction on the use of technological tools in the context of preventing and countering disinformation. The section then continues with an analysis of some of these tools, namely the ones most widely used, examining their potential, but also limitations and societal impact. Lastly, this section offers some recommendations for the further improvement of the use of such tools, especially in what concerns compliance with existing human rights standards. This section builds upon the theory presented in deliverable 2.5 Technological factors for disseminating fake news: social media, deepfakes, bots, swarms of bots. It analyses core technological tools used for detecting and countering disinformation and discusses their potential and limitations. It also provides a brief analysis on the human rights impact of such tools.
Main research questions addressed

● Which technological tools are developed for fighting disinformation and what is their potential and limitations?
● What is the human rights impact of such tools? 

Tech solutions and their limitations

The majority of tech-driven solutions rely on machine learning. The attractiveness of machine learning in the context of targeting and combating disinformation arises from the fact that machine learning models can recognize novel cases and react to them, based on prior learning. The possibility of continuous improvement of machine learning models, makes them seem like an effective tool to address the always-evolving world of disinformation.

1. Natural Language Processing Tools
Social media content analysis and content moderation have been identified as efficient and effective response instruments to the rising challenge of preventing disinformation. In order to carry out large-scale analysis of social media content, social media companies introduced machine-learning natural language processing tools (Facebook, 2019). Natural language processing (NLP) tools have the ability to parse text “[and] the ability of this paring is usually to predict something about the meaning of the text, such as whether to express a positive or a negative opinion’’ (Duarte et al, 2018, p.3).

NLP relies on classifiers trained through text labels/annotations determined by humans which guide the tool to decipher whether some word, phrase or text belongs to the targeted category of content. A collection of examples based on which NLP distinguishes different categories of text is called corpus. In the context of detecting disinformation, the NLP tool will use a corpus which has examples of accurate information and disinformation. Disinformation would then be annotated in a way that the tool could learn from this example and employ it automatically in the future. For example, the NLP tool could determine whether some words are missing, but also analyse the word embeddings that represent the context. For NLP tools to analyse the context, they rely on word embeddings generated by machine-learning tools such as Word2Vec (Duarte et al, 2018). NLP tools can replace journalists and media experts in the process of fact-checking. Kozik et. al argue, for example, that NLP tools can imitate a high level of intuitive reasoning, similar to experienced specialists. As computers process large amounts of data, they're able to detect patterns, “without the need to engineer the features before training the neural network” (Kozik et al, 2022). In fact, models rooted in deep-learning can identify authors of fake news based on literary features (Kozik et al, 2022). Example of BENDEMO, Dutch project aiming to prevent and counteract the spread of online disinformation, shows how NLP tools can replace journalist and media experts in the process of fact-checking. Automated network analysis and NLP technology detects emerging disinformation campaigns in the Dutch-speaking region and across Europe, and publishes fact-checks.

Although NLP tools could contribute to a more efficient prevention and countering of disinformation, they are not without limitations and shortcomings. Failing to address these shortcomings, would not only invalidate their efficiency, but quite the contrary they could contribute to spreading more disinformation and/or negatively impacting human rights.

Dataset bias

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).  

Language limitations

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).

Transparency and accountability

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).

2. Social media bot-detecting tools
Similarly, to the development of NLP tools, machine learning plays a pivotal role in the detection of social media bots. Detection tools that rely on machine learning can be categorized into three different groups; supervised, unsupervised and semi-supervised. As is the case in machine learning NLP tools, machine learning bot detection tools operate based on the availability and quality of the training data which has been labelled or annotated, specifically providing examples of human-managed and bot-managed social media accounts. Whereas one may argue that in the case of machine-learning NLP tools targeting disinformation, the objective is clear, given that there is at least an EU widely accepted definition, in the context of machine-learning bot detection tools this is more challenging. First and foremost, there is no operational definition of social media bots. Secondly, there is a large grey area in detecting the differences between human-like and bot-like behaviour. Existing bot detection tools rely on datasets that map typical bot behaviour, but the final result is often impacted by the following limitations (Yang et al., 2022):

Limited datasets

The development of supervised social media bot-detecting tools relies on the existence of training datasets. These datasets used for annotation and labelling are often limited, due to them being directly extracted from social media. In recent years, especially in the context of the Cambridge Analytica case, many social platforms have limited access to their APIs as a result of human rights concerns or monetized access making it increasingly difficult to employ social media for training AI models. This means that datasets are compiled by human operators who often manually label and annotate information. Recent research findings seem to indicate a very high error rate in detecting the more sophisticated bots, with only24% of bots being accurately labelled (Cresci, 2020).

The accuracy of training datasets is also impacted by bot evolution. Social media bots were initially easily recognizable since they often lacked personal information and presented few social connections, however with time these bots evolved into perfectly engineered accounts that seem human-operated and displaying a large social network (Cresci, 2020).  

Language limitations

According to a 2020 study by Rauchfleisch et al., social media bot-detection tools cannot be transferred from one country to another one. Given the lack of training data available in other languages (all other languages but English), detection tools are likely to give false positives, or negatives since they fail to take into consideration different communication patterns and styles (Rauchfleisch et al, 2020). Even leading instruments such as, Botometer, fail to implement its content and sentiment analysis features in cases of non-English accounts (Yang et al., 2022).

This would mean that current machine-learning social media bots detection tools cannot disproportionately target social media users who are non-native English speakers.


Often, human-operated social media accounts can behave similarly to a bot-operated account, especially in the case of politically engaged individuals or activists, namely they don’t disclose a lot of personal data, including their location, and they don’t share any audio-visual content. In other scenarios, individuals try to randomize their handles to protect their data and privacy. Although these are all valid measures individuals might opt to take in order to protect their privacy, due to biases found in the training there is a high likelihood for such accounts to be labelled as bots. Furthermore, taking down such accounts on the premise of them being bots would infringe the freedom of expression of individuals using those social media accounts.

Human rights impact and the way forward

While technology can be an aid in preventing and countering disinformation, it is also clear that many of the tools developed for this purpose can have a negative impact on human rights. For these reasons, the Council of Europe has identified a set of core rights which need to be protected at an individual level, this becoming a key requirement for any technology developed in the field.

  1. Technological tools must respect the right to human dignity, the right to life, and the right to physical and mental integrity, defined in Article 2, of the European Convention of Human Rights (ECHR). This would mean that when there is a risk of technological tools violating human dignity, the same procedure must be carried out by a human.The profile photo is not of a person, but of an object (flower, landscape) or an animal. 
  2. The right to liberty and security (Article 5, ECHR), must be respected at all times. This right prescribes an obligation towards developers to establish human rights oversight mechanisms which would evaluate possible risks arising from the implementation of those technologies. Such oversight mechanisms could help in addressing issues arising from dataset bias, language limitations or lack of algorithmic transparency.The posts do not refer to the owner's professional/personal life at all, they do not give any indication of the real existence of the owner.
  3. Special attention must be given to safeguarding the right to non-discrimination (on the basis of the protected grounds set out in Article 14 of the ECHR and Protocol 12 to the ECHR). To prevent dataset bias, authorities and developers must ensure that deployed technologies do not cause discrimination, promote harmful stereotypes or foster social inequality. Developers must be aware of these risks and continuously examine if in any way bias is fostered through the development and implementation of these technologies.
  4. The right to respect for private and family life and protection of personal data (Article 8, ECHR) must be safeguarded. Developers should mitigate any negative impact of technological tools on the right to privacy or family life that might rise either in the development or implementation stage. Protection of this right is particularly relevant in the context of bot detection technologies that tend to show a false positive for profiles where individuals are more protective of their personal information. 
  5. The right to an effective remedy for violation of rights and freedoms (Article 13 ECHR) must be protected. Authorities and developers should make sure that there are accessible remedies individuals can rely on in case of unlawful data collection or if the implementation of such technologies causes unjust harm to the individual or violates their rights.
  6. Similarly, to the above-mentioned right, the right to a fair trial and due process (Article 6, ECHR) should be respected. Individuals must have the opportunity to challenge any decisions made based on evidence acquired through the use of these technologies.
  7. Although these technologies are often used to prevent interference in the electoral process through the creation and promotion of disinformation, these tools should in no way inflict the right to freedom of expression and freedom of assembly and association (Article 10 and 11 ECHR). Technologies which target disinformation should respect the principle of transparency, fairness, and responsibility. This obligation is particularly important when it comes to the transparency of algorithms (Leslie et al, 2021).

In conclusion, it is clear that there are multiple technologies which play/will play an important role in detecting and combating disinformation. They protect democracies and their citizens from unlawful interference in their internal processes and shed light on the mechanisms used to manipulate public opinions. This positive impact is not without costs.

As explained in this document, many of these technologies are still underdevelopment and thus subject to many limitations. In addition to the high error rate, there are also cases where their use can have a negative impact on human rights. In order to avoid this, stronger emphasis needs to be placed on understanding technological limitations, introducing privacy-by-design and privacy-by-default approaches in their developments as well as carrying out a constant review of ways in which their design can be improved in order to mitigate potential risks. In addition to this, it is important to ensure that the regulatory framework manages to keep the pace with technological developments, by introducing the necessary safeguards. 

1. Rauchfleisch, A., Kaiser, J., (2020). The False positive problem of automatic bot detection in social science research. PLoS ONE. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0241045
2. Raso, F., Hilligoss, H., Krishnamurthy, V. et al. (2018, September 25). Artificial Intelligence & Human Rights: Opportunities & Risks. Berkman Klein Center for Internet & Society, Harvard University.
3. Posetti, J., Matthews, A. (2018). A short guide to the history of ’fake news’ and disinformation. International Centre for Journalists. https://www.icfj.org/sites/default/files/2018-07/A%20Short%20Guide%20to%20History%20of%20Fake%20News%20and%20Disinformation_ICFJ%20Final.pdf
4. Perel, M., Elkin-Koren, N. (2017). Black Box Tinkering: Beyond Disclosure in Algorithmic Enforcement. Florida Law Review. https://scholarship.law.ufl.edu/cgi/viewcontent.cgi?article=1348&context=flr
5. Osoba O., Wesler IV W. (2017). An Intelligence in Our Image: The Risks of Bias and Errors in Artificial Intelligence. RAND Corporation. https://www.rand.org/content/dam/rand/pubs/research_reports/RR1700/RR1744/RAND_RR1744.pdf
6. New America (n.d). The Limitations of Automated Tools in Content Moderation. New America. https://www.newamerica.org/oti/reports/everything-moderation-analysis-how-internet-platforms-are-using-artificial-intelligence-moderate-user-generated-content/the-limitations-of-automated-tools-in-content-moderation/
7. Duarte, N., Llanso, E., & Loup., A. (2018). Mixed Messages? The Limits of Automated Social Media Content Analysis. The 2018 Conference on Fairness, Accountability, and Transparency. https://cdt.org/wp-content/uploads/2017/12/FAT-conference-draft-2018.pdf
8. Linardatos, P., Papastefanopoulos, V., Kotsiantis, S. (2020). Explainable AI: A Review of Machine Learning Interpretability Methods. Entropy. https://dx.doi.org/10.3390/e23010018
9. Kozik, R., Kula, S., Choras, M., Wozniak, M. (2022). Technical solution to counter potential crime: Text analysis to detect fake news and disinformation. Journal of Computational Science. https://doi.org/10.1016/j.jocs.2022.101576
10. IBM. (n.d). What is machine learning? IBM.https://www.ibm.com/topics/machine-learning
11. European Union Agency for Fundamental Rights. (2022). Bias in Algorithms – Artificial
12. Cresci, S. (2020). A decade of social bot detection. Communications of the ACM. doi:10.1145/3409116
13. Alaphilippe, A., Gizikis, A. Hanot, C., Bontcheva, K. (2019, March). Automated tackling of disinformation. European Parliament. https://www.europarl.europa.eu/RegData/etudes/STUD/2019/624278/EPRS_STU(2019)624278_EN.pdf
14. Hirschberg, J., Manning, C. (2016, May 12). Advances in natural language processing. Science. https://cs224d.stanford.edu/papers/advances.pdf
15. Pasquale, F. (2016, August 29). The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard University Press.
16. Menn, J., Ingram, D. (2017, September 6). Facebook says likely Russian-based operation funded U.S. ads with political message. Reuters. https://www.reuters.com/article/us-facebook-propaganda-idUSKCN1BH2VX
17. European Commission. (2018, March 12). Report on the Public Consultation on Fake News and Online Disinformation. European Commission. ec.europa.eu/newsroom/dae/document.cfm?doc_id=51810
18. Meta AI. (2019, April 15). Announcing new research awards in NLP and machine translation. Facebook. https://ai.facebook.com/blog/announcing-new-research-awards-in-nlp-and-machine-translation/
19. Kessler, G., Rizzo, S., Kelly, M. (2019, December 16). President Trump has made 15,413 false or misleading claims over 1,055 days. The Washington Post. https://www.washingtonpost.com/politics/2019/12/16/president-trump-has-made-false-or-misleading-claims-over-days/
20. Asudeh, A. Jagadish, H.V., Wu, Y. Yu, C. (2020, March 11). On detecting cherry-picked trendlines. Proceedings of the VLDB Endowment. https://dl.acm.org/doi/10.14778/3380750.3380762
21. Gillespie, T. (2020, July). Content moderation, AI, and the question of scale. Big Data & Society. https://www.researchgate.net/publication/343798653_Content_moderation_AI_and_the_question_of_scale
22. Vincent, J. (2020, September 21). YouTube brings back more human moderators after AI systems over-censor. The Verge. https://www.theverge.com/2020/9/21/21448916/youtube-automated-moderation-ai-machine-learning-increased-errors-takedowns
23. Leslie, D. Burr, C. et al. (2021, June). ARTIFICIAL INTELLIGENCE, HUMAN RIGHTS, DEMOCRACY, AND THE RULE OF LAW. Council of Europe. https://rm.coe.int/primer-en-new-cover-pages-coe-english-compressed-2754-7186-0228-v-1/1680a2fd4a
24. Yang, K., Ferrera, E. Menczer, F. (2022, August 20). Botometer 101: social bot practicum for computational social scientists. Journal of Computational Social Science. https://link.springer.com/article/10.1007/s42001-022-00177-5 

Co-funded by European Commission Erasmus+
University of Malta
University Rey Juan Carlos
Logo New Strategy Center

Project: DOMINOES Digital cOMpetences INformatiOn EcoSystem  ID: 2021-1-RO01-KA220-HED-000031158
The European Commission’s support for the production of this publication does not constitute an endorsement of the contents, which reflect the views only of the authors, and the Commission cannot be held responsible for any use which may be made of the information contained therein.

Ivan, Cristina; Chiru, Irena; Buluc, Ruxandra; Radu, Aitana; Anghel, Alexandra; Stoian-Iordache, Valentin; Arcos, Rubén; Arribas, Cristina M.; Ćuća, Ana; Ganatra, Kanchi; Gertrudix, Manuel; Modh, Ketan; Nastasiu, Cătălina. (2023). HANDBOOK on Identifying and Countering Disinformation. DOMINOES Project https://doi.org/10.5281/zenodo.7893952