Willemijn Kornelius is a graduate of Civil Law (Intellectual Property) (LL.M.) Leiden University and a current Legal Research Master student at Utrecht University, the Netherlands. Her research focus lies within the field of law & technology from a European perspective. During her research internship at CEPRI (University of Copenhagen), she is looking into digital platforms and liability for illegal content.
In Europe, the EU legislature looks for possibilities to tackle disinformation, intentionally spread false content (see an overview here). One of the ways the EU fights disinformation is through a cooperation with online platforms like Twitter and Facebook. They are considered to be best placed to reduce, disable access to or remove content containing disinformation, because they directly control how their platforms are used. Yet, these platforms have to navigate in a complicated environment. Removal of content is at odds with their users’ freedom to express themselves. Online platforms are expected to respect this freedom when they act against disinformation. In the discussion on how platforms should go about this, there is increasing attention for the use of artificial intelligence technologies to detect disinformation. This blogpost critically examines existing artificial intelligence systems for the detection of disinformation online. It argues that these systems are not fit to be used in the complex disinformation framework. Yet before we look at specifics of the existing AI systems, let us dive in what disinformation actually is and what obligations rest upon online platforms.
1. What is disinformation online?
Internet users increasingly encounter intentionally spread “fake news”. During the US elections in 2016, the conspiracy theory Pizzagate was spread and went viral, with the intention to blacken Hillary Clinton and support the election of Donald Trump. We saw a boost of COVID-19 related disinformation during the pandemic: incorrect scientific claims about the virus, unfunded conspiracy theories and dangerous tips were widely spread. Currently, there are considerable concerns about the power of online disinformation in the war in Ukraine (see e.g. here).
For a well-functioning democracy, it is crucial that citizens have access to diverse and most accurate information. The spread of “fake news”, or disinformation, potentially leads to distrust in governmental institutions and manipulation of public opinion. Social media platforms have the potential to function as online public spaces, where people share their thoughts and opinions by uploading information (“content”), such as Tweets on Twitter.
The EU Commission expressed in its 2020 Democracy Action Plan that combatting disinformation is one of three pillars to strengthen democratic resilience. In the 2018 Action Plan against Disinformation, disinformation is understood as “verifiably false or misleading information that is created, presented and disseminated for economic gain or to intentionally deceive the public, and may cause public harm”.
2. Voluntary action? Yes, but need to respect users’ freedom to expression
The wish of the EU Commission to get social media platforms “on board” in the fight against disinformation, resulted in the development of a Code of Practice on Disinformation in 2018. The Commission evaluated the effectiveness of the Code of Practice, which resulted in a Strengthened Code of Practice this year. It is a voluntary set of industry standards to fight disinformation, signed by 34 actors, such as major online platforms like Facebook and Twitter. In Part IV, ‘Integrity of Services’, the Code states that the signing actors recognise the importance to ensure the integrity of their services, by implementing safeguards against disinformation. To tackle this, the signatories agreed to adopt clear policies clarifying behaviours and practices that are prohibited on their services (Commitment 14).
Although the Code is signed voluntarily and online platforms are, as private parties and in line with the law of contract, free to determine what information they do not allow on their services, their actions are governed by the fundamental rights at stake. Following measure 15.2, the platforms can use algorithms for detection and moderation of impermissible conduct and content. This measure emphasises that these algorithms are trustworthy and respect the rights of end-users. In several other places, the Code emphasises the need to give ‘due regard to freedom of expression and information’. For example, in Preamble (c): “As stressed in the Communication [Communication “Tackling online disinformation: a European approach“], fundamental rights must be fully respected in all the actions taken to fight Disinformation. The Signatories are mindful of the fundamental right to freedom of expression, freedom of information (…)”. Moreover, the Digital Services Act, a new EU regulation, introduces a framework of responsibilities for online platforms. It influences the way online platforms need to govern their users activities. Under the Digital Services Act online platforms can only impose restrictions on the use of their service (e.g. removal of content) with “due regard of the freedom of expression” (Article 14(4) and recitals 47 and 54 DSA)..
So what does that mean? Freedom of expression is a fundamental right that can be defined in various ways and its scope can vary even within the EU Member States. In EU law it is enshrined in Article 11 EU Charter of Fundamental rights (“Charter”), with a corresponding scope to Article10 ECHR. Article 11 Charter provides that everyone has the right to freedom of expression, which includes “freedom to hold opinions and to receive and impart information and ideas without interference by public authority and regardless of frontiers”. The ECtHR has confirmed in i.a. Vladimirov Kharitonov v. Russia that the internet is now a ‘principal means’ by which individuals express themselves and communicate on the internet (the CJEU repeated this in its recent Case C-401/19, para 46). Consequently, any restriction imposed on the (dissemination) of content (“content moderation”) interferes with this right. It could be that interfering is only justified when content is indeed verifiably false or misleading and shared intentionally. If artificial intelligence technologies are employed, this interpretation of freedom of expression seems to require that they should be specifically targeted. The question is whether they can.
3. Artificial intelligence to detect disinformation online
In the context of this blogpost, artificial intelligence is understood as the ability of a technology to perform a task with a human characteristic, such as decision-making. Such technologies are based on machine learning: algorithms and large datasets are used to train the systems to recognise patterns and extract relevant information (Kertysova 2018; Surden 2019). In this case, we are looking at technologies that aim to identify disinformation among textual information shared online (“content”). Techniques that for example detect manipulated images are not considered. In recent years, different techniques have been developed (Bontridder and Poullet 2021). In this post, broadly four techniques are identified and discussed below.
3.1 End-to-end content detection models
This technique is understood as an end-to-end-model trained with machine learning on a dataset containing labelled information, such as articles (Akers e.a. 2018). It uses natural language processing to convert natural text into computer-readable texts. In its most simple form, information is either labelled as “disinformation” or “accurate information”. The system is trained to differentiate between these two types of information. OpenAI recently launched a content moderation tool (Moderation endpoint). This tool detects illegal content, and is thus in principal not made to detect disinformation, but the idea is the same. Their tool also works on the basis of natural language processing
The disadvantage of these systems is that for the system to be accurate, a large amount of classified data is necessary. Moreover, these datasets, even when they are very large, are vulnerable to contain biases.
3.2 Automated fact-checking
Artificial intelligence technologies can also be used to detect factual inaccuracies. Fact-checking is the task of assessing whether claims made are true. It is typically performed by journalists by researching different sources of information (Thorne and Vlachos 2018). Initiatives to automate fact-checking of uploaded content have resulted in technologies that compare content with evidence or some form of knowledge available to the system, the dataset. An example of a created dataset is the one developed under the EU funded FANDANGO project (Martín-Gutiérrez e.a. 2020). FANDANGO provides a dataset that combines and links several European open data sources. This year, Meta (Facebook) announced that it developed Sphere – an AI-system capable of scanning citations to check whether they support corresponding claims. Interestingly enough, this system is trained on claims from Wikipedia.
Automated fact-checking has two important limitations. Firstly, its assessment whether a claim is true or false is dependent on the (quality of the) dataset. Do we trust Wikipedia to be a reliable source of information? Second, natural language used to make the claims under assessment contains a lot of nuances that the technology is not always able to detect. Certainly, new innovations keep changing the environment. With chatbot ChatGPT on the rise, launched by OpenAI last November, we can certainly see artificial intelligence technologies becoming more advanced and the datasets used becoming more all-encompassing. However, although OpenAI presents ChatGPT as a tool to challenge incorrect premises (which can be useful for tackling disinformation), experts say that its answers still have flaws (see e.g. here). It is generally assumed that fact-checking is still better performed by humans.
3.3 Detection of misleading style
Some technologies are designed to analyse the style of a certain post. By doing so, they attempt to construe the intent of the post based on linguistic features of the text. This is also described as the psycholinguistic analysis of texts (Choras e.a. 2021). An indication for disinformation could be e.g. the use of manipulative wording or indirect forms of expression (Rashkin e.a. 2017). However, the absence of specific language that can be associated with disinformation does not mean that the information in the post is indeed true or accurate.
3.4 Metadata analysis
A trend can be witnessed in technologies that depart from this textual based analysis. These technologies instead focus on a post’s metadata, such as the user profile (e.g. geolocalization) or user activity of the person that uploaded the post or social media sharing patterns (the way a post is received by e.g. its shares and likes) (Akers e.a. 2018). The project GoodNews developed a model that looks at how disinformation spreads (for its technical description: see Monti e.a. 2019). It is based on existing research that shows that online disinformation is shared differently than true claims.
Other types of metadata analysis look e.g. at the reputation of the individual that posted the content or the social group in which it is shared (Choras e.a. 2021). The ‘reputation’ is based on suspicious domain names, IP addresses and feedback of readers.
4. The thin line
Despite the numerous initiatives of promising content detection technologies, especially when they are combined, and the high rate of accuracy they promise (GoodNews claims an accuracy of 93%), there are some limitations. The first problem is that these models are trained on datasets. That makes them dependent on the quality and scale of these datasets. If certain information is wrongly labelled as disinformation, the models are trained in the wrong way. Moreover, these datasets are vulnerable to contain human biases. A second limitation is the text dependency of most technologies. The uploaded content contains natural language, written to be understood by humans. However, the technologies are not (always) capable to capture the nuances or implied claims in long sentences (Kertysova 2018; Study on disinformation European Parliament Research Service (EPRS) 2018).
As a consequence, the use of these technologies can lead to false positives (content is labelled as disinformation, but it is not) or false negatives (content is not detected as disinformation, but it is). It is a thin line between over-blocking or under-blocking.
A third issue is the lack of a uniform definition of disinformation. The Commission has defined disinformation, but it has to be established from a case-to-case basis whether information is actually false and whether it is shared with the wrong intention. It relates to a more fundamental issue: who do we want to judge what is the truth and what is not? Francesco Nucci, principal researcher for the FANDANGO project, states: “Fake news is not a mathematical question of algorithms and data, but a very philosophical question of how we deal with the truth.” (see this blogpost at the website of the European Commission).
5. AI fit for the fight against disinformation?
This blogpost outlined to what extent artificial intelligence detection systems are fit to be used in the complex disinformation framework. The limitations addressed above makes that these technologies are not fit for this. First of all. they are expected to effectively and accurately detect disinformation so that the online platforms can act (moderate). However, they are unable to do so due to the dependency on datasets, the nuances in natural language and the lack of a uniform definition of disinformation. Second, they should be used in a way that respects the freedom of expression and information of users. As said before, this requires strict targeting of content containing disinformation. Content that is compatible with the applicable user terms and conditions should in principal remain unaffected. However, in light of the risk of over-blocking and the lack of a uniform definition, it is inevitable that users’ are restricted in their freedom to share information online. Detection technologies could therefore support online platforms in their difficult task to identify disinformation, but should not be used on their own.