1.1 Perceived Need for Legal Intervention

This section considers the growing array of legislative efforts geared (in whole or in part) to addressing the harms of misinformation and disinformation. In terms of the sheer number of enactments, these laws are noteworthy. For instance, a study conducted by Gabrielle Lim and Samantha Bradshaw for the Center for International Media Assistance and the National Endowment for Democracy reports that, globally, ninety-one items of legislation were enacted between 2016 and 2022 to address misinformation and disinformation (Lim and Bradshaw, Reference Lim and Bradshaw2023). Another study from Simon Chesterman at the National University of Singapore reports that, as at the start of 2024, more than half the countries of the world had adopted laws designed to ‘reduce the impact of false or malicious information online’ (Chesterman, Reference Chesterman2025: 934).

While touched on in the second study, we pay particular attention in this section to two significant pieces of regional and national legislation adopted in 2022 and 2023 which are geared specifically to platforms that host and circulate misinformation and disinformation posted by users, namely the EU DSA and the UK OSA. (The Australian version of the OSA was passed in 2021, albeit in more restricted form.) As Chesterman points out, and as we come back to later, in practice ‘much [in this space] has depended on technical interventions by platforms’ (Chesterman, Reference Chesterman2025: 944). Laws such as the DSA and OSA provide some legal oversight but bring their jurisdictions into sharp relief with the US where § 230 of the Communications Decency Act (CDA) of 1999 provides that platforms are generally immunised from legally responsibility for content posted by users,Footnote ³ and ‘good Samaritan’ interventions by platforms are further protected from legal responsibility (Chesterman, Reference Chesterman2025: 946).

Another focus of this section is the EU AI Act of 2024, with its provisions currently scheduled to come into operation in stages from February 2025 to August 2027. This will also impact on (some) AI-fuelled misinformation and disinformation. Indeed, its impact could be very significant, although its precise reach is still not clear with much depending on how the provisions (once in force) are deployed in practice (see Duivenvoorde, Reference Duivenvoorde2025). But the EU is not the only jurisdiction to address AI-related misinformation and disinformation specifically, even if its AI Act is the most comprehensive legislation to date. For instance, in the US now, there is new federal legislation in the form of the TAKE IT DOWN Act of 2025 to address sexually explicit deepfakes (see Ortutay, Reference Ortutay2025). Other laws, such as Singapore’s Protection from Online Falsehoods and Manipulation Act of 2019 (POFMA), are broad enough in their terms to address deepfakes (Shanmugam, Reference Shanmugam2024).

No doubt there will be more legislative efforts to come in this space. Some may be controversial, as indeed are many of the legislative efforts noted previously. However, in this section we do not seek to critique particular legislative initiatives (see, for instance, Coe, Reference Coe2023 on the OSA), or particular types or features of legislation (for a comprehensive review, see Katsirea, Reference Katsirea2024), or the definitional challenges of embedding vague terms such as ‘misinformation’ and ‘disinformation’ into law (see Roberts et al., Reference Roberts, Morris and Sorial2025; OECD, 2024: 17), or the fact of using legislation to manage what some might say is ‘a sense of urgency expressed in polemical terms’ (Burgess, Reference Burgess2023). Instead, we think it is useful to treat misinformation and disinformation as a problem that can lead to real harms and then consider broadly the legislation being introduced to address concerns around these harms. At the same time, we acknowledge concerns about actual and potential impacts of these laws on other rights, freedoms and interests, such as freedom of speech and innovation. And we point out the limits of these laws, even if well designed to address misinformation and disinformation in a properly balanced, or ‘proportionate’, way – and of law in general.

1.2 Continuing Reliance on Older Laws

Of course, the new laws do not operate in a vacuum. Older laws continue to regulate and their significance should not be underestimated. For instance, users of Facebook successfully settled class action claims in California inter alia for breach of various privacy torts and negligence following Cambridge Analytica’s harvesting of their data to target political and commercial advertising without their knowledge or consent (In Re Facebook, Inc. Consumer Privacy User Profile Litigation, 2018). Such torts might also be drawn on to address AI-related practices. Indeed, some have suggested that AI was already being used by Cambridge Analytica in 2016 (see Delcker, Reference Delcker2020). More recently Scarlett Johansson’s legal dispute with OpenAI after its launch of an AI-generated voice in ChatGPT ‘eerily’ resembling her own is reminiscent of the earlier California right of publicity case Midler v Ford Motor Co where Bette Midler succeeded in arguing that the use of a soundalike voice in Ford advertising for a nostalgic line of cars was a violation of her common law right of publicity under California law (Midler v Ford Motor Co, 1998).Footnote ⁴ Open AI has since reluctantly withdrawn the soundalike voice (Robins-Early, Reference Robins-Early2024). Further, in Starbucks v Meta Platforms (2025), conservative activist Robby Starbuck successfully claimed defamation against Meta for disseminating false statements about his participation in the capital riot of January 2021, among other false statements, via its chatbot Meta AI, resulting in a settlement (see Starbuck v Meta Platforms, Inc., 2025; Volokh, Reference Volokh2025a). Starbuck has followed up with another complaint of defamation against Google, alleging that the company spread ‘outrageously false’ lies through its AI products (Volokh, Reference Volokh2025b).

Misinformation and disinformation in trade or commerce may also be targeted under consumer protection laws such as the Federal Trade Commission Act (FTCA) in the US (specifically, its § 5 prohibition on ‘unfair or deceptive acts or practices’), the Australian Consumer Law (ACL) of 2010 in Australia (including its s 18 prohibition on misleading or deceptive conduct), and the EU Unfair Commercial Practices Directive (UCPD) of 2005 in Europe (see Federal Trade Commission, 2024; Harding, Paterson and Bant, Reference Harding, Paterson and Bant2022; Duivendoorde, Reference Duivenvoorde2025). Fraud may also be claimed in egregious cases.Footnote ⁵ Already, there are examples. For instance, following the Cambridge Analytica scandal, in 2019 Facebook was fined a record $5b by the Federal Trade Commission (FTC) for deceiving users about their ability to keep personal information private (FTC, 2019a), and Cambridge Analytica was prohibited from making further misrepresentations (FTC 2019b). A securities fraud claim launched by investors against Mark Zuckerberg and others has also been settled (Associated Press, 2025). The FTC further has responsibility for the US Children’s Online Privacy Protection Act (COPPA) 1998, requiring parental consent for children (under 13 years) to access commercial websites and online services. As such, COPPA in a sense protects children from online misinformation and disinformation, although its remit is much broader. The same may be said of the Australian social media ban for children (under 16 years) in the Online Safety Amendment (Social Media Minimum Age) Act 2024, in force from December 2025.

Data protection laws establishing standards for the processing of personal data may also be relied on to address the generation and spread of misinformation and disinformation within their compasses.Footnote ⁶ The EU General Data Protection Regulation (GDPR) 2016 could well prove to be especially powerful, given its design as a comprehensive modern regime (see Yeung and Bygrave, Reference Yeung and Bygrave2021). Its impact can be seen already in certain cases, including a recent case against Google where the European Court of Justice (CJEU) held that the right of erasure/right to be forgotten in article 17 GDPR may be drawn on to require dereferencing by search engines of ‘manifestly inaccurate’ personal data in breach of article 5(1)(d), under specified conditions (TU and RE v Google LLC, 2022).Footnote ⁷ But other regimes may also provide important protections, either directly through their accuracy requirements for data processing or indirectly through other standards applied to data processing of personal data. To give just one example, in Illinois, a class action lawsuit launched under the Biometric Information Privacy Act 2008 (BIPA) in relation to Clearview AI’s generation and dissemination of profiles harvested from biometric face images on the internet, for use by Clearview’s commercial customers including in law enforcement, was recently settled (In Re Clearview AI Consumer Privacy Litigation, 2025). Although the non-consensual processing of individuals’ biometric data was the main legal foundation of the BIPA claims, the court noted as ‘not wholly without merit’ the concern expressed by the American Civil Liberties Union (ACLU) that ‘numerous studies have shown that face recognition technology misidentifies Black people and other people of colour at higher rates than white people’ leading to ‘higher incidence of wrongful arrest’ (In Re Clearview AI Consumer Privacy Litigation, 2025: 32).Footnote ⁸

Also worth noting are electoral laws which may be drawn on to deal with current issues of misinformation and disinformation in election contexts. For instance, in Australia, in the lead-up to the May 2025 federal election, the Australian Electoral Commission established under the Commonwealth Electoral Act 1918 (Cth) launched a ‘Stop and Consider’ campaign urging Australian voters to pause and reflect on ‘information being distributed which is seeking to influence your vote’, to ask questions such as whether it is from ‘a reliable or recognisable source’, and whether it could ‘be a scam’, and stressed the AEC’s role in providing reliable information on the electoral process and in investigating complaints (AEC, 2025; see Grantham, Reference Grantham2025). In other countries, more concrete steps have been taken under existing political advertising laws. For instance, in the US the Federal Communications Commission (FCC) which has responsibility for telephone and broadcast communication, determined that the US Telephone Consumer Protection Act (TCPA) of 1991 extends to telephone calls that use or include AI-generated voices, and in September 2024 fined a political consultant $6 million over robocalls that mimicked President Joe Biden’s voice, urging New Hampshire voters not to vote in that state’s Democratic primary (FCC, 2024a).

The developments discussed earlier suggest that older laws (including some laws that are not all that old and are still being interpreted and updated) will continue to be looked to in this context, as in others. At the very least, they highlight the flexibility of established legal frameworks and doctrines that may be called on to address misinformation and disinformation. In some, perhaps even many, instances they provide a way to address these, while still providing a balance with free speech and innovation.

We now turn to newer laws, which have been designed (at least in part) to deal with misinformation and disinformation, and their associated harms.

1.3 Newer Laws

As a preliminary comment, we note that, while most jurisdictions view misinformation and disinformation and their harms as problems that requires a definite legal solution, there is little consistency in terms of the scope and range of the new laws that are emerging (including whether they should be framed in broader or narrower terms, and whether their focus should be primarily on misinformation and/or disinformation, or rather on harm, or on some combination of these) and their balances with freedom of speech and innovation. Table 1 presents examples of legislative initiatives in selected jurisdictions over the last five to six years. We then move on to consider some of the types of harms that these laws seek to address (in either specific terms or as part of a broader sweep), focussing on hate speech, sexually explicit deepfakes, political interference, and scams,Footnote ⁹ although we still note considerable variations between jurisdictions as to whether, when, and how these should be regulated.

Table 1Examples of New Laws

Table 1Long description

This table compares legal and regulatory frameworks across Singapore, Australia, the European Union, the United Kingdom, and the United States. It contains four columns: jurisdiction, laws, scope, and measures.

Each jurisdiction has several rows listing individual statutes.

• In respect of Singapore, laws addressing false information, harmful online content, and scams are listed.

• In respect of Australia, laws concerning online safety, sexually explicit deepfakes, and scam prevention are listed.

• In respect of the European Union, instruments covering illegal or harmful online content, AI system regulation including deepfakes, and political advertising transparency are listed.

• In respect of the United Kingdom, laws on harmful online content and upcoming offences related to sexually explicit deepfakes are listed.

• In respect of the United States, a federal act targeting non-consensual intimate imagery, including deepfakes, is listed.

Within each row, the “Scope” column summarizes what the law covers (e.g., falsehoods, minors’ safety, scams), while the “Measures” column describes the enforcement mechanisms such as takedowns, blocking orders, transparency requirements, risk mitigation duties, or criminal penalties.

1.4 Summary

This section has considered some of the new legislation addressing misinformation and disinformation and their associated harms. As we have seen, there are significant differences between these jurisdictions, with the EU adopting a much more expansive approach in terms of the types of conduct and harms being addressed than the US, and a third group (UK, Australia and Singapore) coming in between. Further, the variations between the jurisdictions surveyed reflect not just different conceptions of what harms should be addressed, and how they should be addressed, but also different balances found with (among other things) freedom of speech and innovation. In particular, the narrow scope of legal regulation of misinformation and disinformation in the US compared to the broader, more open-ended and proactive legal regulation in (most) of the other jurisdictions surveyed, and especially the EU, suggest some very different balances between freedom of speech and innovations in these jurisdictions. This conclusion is reinforced by the recent publication of the US (Republican led) House of Representatives Staff Report characterising the EU DSA as a ‘powerful censorship law’ (US Judiciary Committee Staff Report, 2025: 2, 15).Footnote ²¹ While this may be regarded as an extreme statement of the US position on freedom of speech and innovation, there can be little doubt that (as Chesterman puts it) the US ‘is somewhat of an outlier in this area, with strong First Amendment protections covering speech that would be unlawful in many other jurisdictions’ (Chesterman, Reference Chesterman2025: 947, footnote 83).

Such divisions discussed earlier suggest that much of the work in addressing the harms of misinformation and disinformation, which anyway often operate across national and regional boundaries with little regard to law, will (continue to) fall to platforms and their practices, including their uses of technologies, and to broader social measures. Indeed, in our view, one of the strengths of the EU DSA, as explored further in the following section, is its acknowledgement of the value of these alternative modalities.

2.1 Content Moderation

As our previous section has shown, despite the extensive and highly variable amount of legal regulation focussed on addressing harmful misinformation and disinformation in many jurisdictions, which may be underpinned by different conceptions of rights, there are normative and practical limits to legal intervention. Further, even where there are laws in place, the scale and scope of online content distribution has meant that jurisdictions are increasingly turning to self-regulatory or co-regulatory frameworks that necessarily require the assistance and input of platforms. As a result, it is impossible to discuss the current state of misinformation and disinformation regulation without considering the complex and often-technical system of content moderation that shapes our contemporary online environment. Indeed, one thing that is particularly notable about content moderation today is the way that AI is being employed to detect misinformation and disinformation and suppress its distribution. We outline these developments in this section and highlight an alternative narrative around AI, where the technology is cast as (at least part of) a solution to the misinformation and disinformation problem as opposed to the cause. While we are equally sceptical of claims that AI can solve all of these challenges, a claim often circulated by self-interested platforms, our examples allow us to assess the extent to which developments around AI can help address a significant policy issue.

The section also examines key normative questions about the challenges associated with private regulation of speech, in particular deploying technical systems, sometimes in conjunction with human fact-checkers. Of course, the absence of law does not entail the absence of regulation or intervention. And nor does the presence of law necessarily determine the amount of regulation that goes on. It is quite commonly being done privately on platforms anyway, with much less accountability and oversight – although Meta’s Oversight Board presents a notable exception here. While this argument about lack of accountability and insight has been made many times (Balkin, Reference Balkin2018; Gillespie, Reference 56Gillespie2018b), we have chosen to highlight Michael Davis, a scholar and former public servant who was involved in the development of Australia’s self-regulatory framework around misinformation and disinformation. He notes that without establishing wider regulatory infrastructures, content moderation is left to platforms who become ‘the arbiters of truth’ (Davis, Reference Davis2024). We build on this useful critique and suggest that in many cases, these tensions around regulation point to the need for a more sophisticated infrastructure around platforms, involving efforts such as researcher data access and platform observation.

The section proceeds as follows. First, we introduce and define what we mean by content moderation. In the process, we sketch out how AI and other technologies such as blockchain are being deployed by platforms. Second, looking forward, we discuss experimental and scholarly work directed at the capacity of these technologies to detect or address harmful misinformation and disinformation, either by themselves or in conjunction with human fact-checkers – using examples of detecting deepfakes and automated fact-checking to further explore the interventions. Finally, we end with a reflection and development of our argument, addressing what we set out at the beginning of this Element, as well as those views outlined in this section.

2.2 Definitions, Tools and Techniques

The term content moderation refers to a range of practices that digital platforms employ to manage the constant stream of messages, pictures, videos, articles and other forms of content that are published online along with any interactions that occur. One of the most productive and expansive definitions of the term is provided by Gillespie et al. (Reference Gillespie, Aufderheide and Carmi2020) who describe it as:

From the earlier description, the centrality of private power is made clear. The practice of content moderation emerges from and is shaped by the terms and conditions of platforms as well as any associated guidelines (although legislation can inform these rules). There is also a brief nod to the significant emotional and physical labour that moderation can involve, given its reliance on human beings as well as technical systems (Gillespie, Reference Gillespie2018a; Roberts, Reference Roberts2019). While promotion of safety, or alleviation of harm, emerges as an obvious lens through which to view moderation, it is deployed for a range of reasons including managing the spread of misinformation and disinformation. A range of techniques are used to moderate content, and it is worthwhile briefly noting these before turning our focus to the (seemingly) cutting-edge misinformation and disinformation detection technologies promoted as part of the wider AI moment.

Even though the entire goal of content moderation is to anticipate issues and remove problematic content before it is even published on the platform, the large volumes of content posted online every day inevitably continues to cause concern amongst audiences. In response, people are encouraged to ‘flag’ content that concerns them, which refers to the act of a person using platform tools to report content through the ‘predetermined rubric of a platform’s […] community guidelines’ (Crawford and Gillespie, Reference Crawford and Gillespie2016: 411). The practice on its own raises significant concerns as people can use the tool to silence or harass ‘users that are already vulnerable to platforms’ censorship, such as sex workers, LGBTQIA+ and sex-positive users, activists and journalists’ (Are, Reference 48Are2024) – to highlight just a few affected demographics.

Another strategy is labelling, which sees potentially controversial or inaccurate content identified as such. In some cases, this may be a fact-check (more on this aspect later), but labels can also include ‘a click-through barrier that provides a warning, a content sensitivity alert, or the provision of additional contextual information’ (Morrow et al., Reference Morrow, Swire-Thompson and Montgomery Polny2022: 1365). One example of the latter is the addition of a ‘forwarded’ label on WhatsApp messages, in an effort to add further context around viral misinformation or spam (Morrow et al., Reference Morrow, Swire-Thompson and Montgomery Polny2022). X (formerly Twitter) also may label content that is ‘significantly and deceptively altered, manipulated, or “fabricated”’ (Authenticity, 2025). Other digital platforms such as Meta, TikTok and YouTube require users to label AI-generated and manipulated content. There are also growing plans to incorporate content credentials (i.e., metadata that are cryptographically attached to digital content to provide verifiable trails of the content’s origins and edits) on their platforms (Soon and Quek, Reference Soon and Quek2024). C2PA is the most visible initiative with a range of key players, from Adobe and Microsoft to BBC and Sony. Other efforts include Open AI ensuring that images generated by their DALL-E 3 model contain metadata. These efforts go beyond classic concerns associated with content moderation and highlight an emerging problem space around ensuring the provenance of content and identifying the presence of synthetic media (Soon and Quek, Reference Soon and Quek2024). While labelling for different purposes has become more prominent across a variety of platforms, the overall efficacy of the intervention is still uncertain (Morrow et al., Reference Morrow, Swire-Thompson and Montgomery Polny2022; Peters, Reference Peters2023). Still, it presents a potentially useful first step to addressing harms.

Labels do not happen in isolation, with various platforms often seeking to additionally reduce the visibility of labelled content amongst the wider audience. While removal had long been seen as the preferred outcome from content moderation efforts, there are increasing concerns about the potential for ‘backfire effects’ (Nyhan and Reifler, Reference Nyhan and Reifler2010) where motivated users posting and sharing misinformation become emboldened because of enforcement efforts, which to the targeted individual can feel like censorship. In response to this, platforms are starting to restrict the circulation of suspect content (often called ‘borderline content’) without removing it from the platform (Gillespie, Reference Gillespie2022; Zeng and Kaye, Reference Zeng and Kaye2022; Are, Reference 48Are2024). The removal of, or restriction of access to, content identified to be false, altered, unverified or harmful, or user accounts propagating such content by digital platforms such as TikTok, Meta and YouTube is arguably done in pursuit of the goal of responsive protection (Soon and Quek, Reference Soon and Quek2024). Once again, these efforts do not always solely focus on issues such as misinformation and disinformation. For example, the approach has allegedly been used by TikTok to restrict the visibility of certain types of content related to the Black Lives Matter movement, developmental disorders (such as autism) and unattractive individuals (Zeng and Kaye, Reference Zeng and Kaye2022). Along with these efforts, platforms may stop people from engaging with selected content, or creators from monetising their content by, for example, placing advertisements against their posts. With creators rarely told about these interventions, the term ‘shadow-banning’ (or reduction) has become a commonly used term to describe this form of moderation (Gillespie, Reference Gillespie2022; Are, Reference 48Are2024).

Indeed, once we get past interventions like flagging that involve human effort, technical systems and interventions that are commonly referred to as AI starts to enter the picture. Machine learning systems are regularly used to support the shadow-banning (or reduction) methods described previously. The effort involves the use of natural language processing, which relies on supervised training, clean datasets and structured systems to produce predictions. We outline these details to point out the differences between these carefully constructed models, as well as the open and adaptive nature of generative systems like ChatGPT. As Robert Gorwa and colleagues go on to explain, these systems can be used to ‘predict whether text may constitute hate speech, and based on that score, flag it for human review’ (Gorwa, Binns and Katzenbach, Reference Gorwa, Binns and Katzenbach2020). In a similar fashion, misinformation and disinformation can be identified by using machine learning to extrapolate from a massive database of ratings produced by human classifiers (Gillespie, Reference Gillespie2022).

In general, content moderation has always involved the use of technical systems to a greater or lesser extent. The introduction of machine learning has simply resulted in different technical capabilities being made available to both platforms and the human moderators, who are still deeply involved in the overall enterprise. Given this track record, it is perhaps no surprise that one of the most human-intensive misinformation and disinformation interventions – fact-checking, which requires working fact-checkers to develop detailed assessments of incorrect information – is starting to engage with AI technologies. Various platforms collaborate with fact-checking organisations to better identify and correct misinformation and disinformation. Google and YouTube have provided funding to the International Fact-Checking Network to ‘increase the impact of fact-checking journalism’, and Meta famously introduced a third-party fact-checking programme, which has funded a significant number of fact-checking institutions currently in operation (Watt, Montaña-Niño and Riedlinger, Reference 68Watt, Montaña-Niño and Riedlinger2025). The most notable deployment was around COVID-19, which saw concerns about the quality of information circulating online become particularly pressing. Conspiracies about how the virus is transmitted, as well as the quality and efficacy of the vaccines, spread like wildfire – in response, platforms were pressured to ensure that official communications were prioritised. Extra initiatives were taken by platforms, demonstrating their stronger commitment to combat misinformation and disinformation over the pandemic, as well as their support of the vaccination programmes (Tan, Reference Tan2022a, Reference Tan2022b). For example, both platforms X and Google worked with public health authorities like WHO and the respective health ministries in many countries to review health information on the pandemic and to ensure that credible information is accessible.

Looking forward, there is an increasing perception in technology quarters that some elements of fact-checking – although perhaps not the whole role – may be able to be automated using AI. This raises the following questions: How close are we to a world where ‘corrections’ are automated, and what are the ethical limitations of moving to this approach?

2.3 From Fact-Checkers to Fact-Bots?

Fact-checking originally started in the early 2000s as a way of correcting the record in political debate, challenging the wordsmithing and obfuscations of statements and claims made by politicians. These efforts were originally located on dedicated websites like FactCheck.org, before gradually spreading across the wider institution of journalism. Various established newsrooms launched their own fact-checking operations, often for election periods; new fact-checking organisations started to appear across the world, with a particular boom period noted in the early to mid-2010s (Graves, Reference Graves2018; Vinhas and Bastos, Reference Vinhas and Bastos2023). Funding and support for the practice boomed following the election of Donald Trump in 2016. Not only did Trump make a habit of verbally disparaging the press and making wild and inaccurate claims, but as noted earlier, there was evidence of Russian interference in that US election. As part of this period of growth and transition, fact-checking became embedded in the wider content moderation process across various platforms (Vinhas and Bastos, Reference Vinhas and Bastos2023).

It is worth repeating that fact-checking is not (and has never been) wholly free from automation, AI or other technical advances. To pick one leading example from the technology sector, Meta’s third-party fact-checking programme, which has been switched to a ‘community notes’ system in the US,Footnote ²² embeds fact-checkers in systems and processes. The standard process involves human fact-checkers accessing a stream of content that Meta’s system has identified as misinformation and disinformation. Accredited fact-checking organisations then selected pieces of content to fact-check, and label the content accordingly as ‘false, altered or partly false’ or ‘Missing context’ (Meta, 2025). As an interesting aside, the organisations earn revenue from Meta for the completed tasks, which has turned the technology giant into one of the major funders of fact-checking more generally (Watt, Montaña-Niño and Riedlinger, Reference 68Watt, Montaña-Niño and Riedlinger2025).

Nevertheless, as with many other sectors, questions are being raised about whether human labour can be augmented (or even replaced) with AI. Fact-checking – like all journalism – takes a significant amount of time. Phone calls need to be made, information must be sourced, and details must be confirmed and re-confirmed before publication. The question is whether automation and AI can help do this within a shorter time and at lower costs. Many platforms are relying on algorithmic moderation and promoting authoritative sources, rather than deleting misinformation which could reflect alternative viewpoints or constitute legitimate dissent. As a result, these platforms are making certain disputable or contestable content less visible, rather than removing it completely, hence allowing for different perspectives to be shared while reducing the viral potential of what could be harmful falsehoods (see Cademix, 2025).

Discussions around efficiencies are a common feature of debates around AI and the labour market, and it is no surprise that we find them circulating around fact-checking (OECD, 2021). For one, these tendencies align with the natural philosophical orientation of technology firms raised in the ashes of the ‘Californian ideology’ (Barbrook and Cameron, Reference Barbrook and Cameron1996) – one centred around optimism and technological determinism – to embrace technical solutions over socio-technical interventions. However, the sort of policy discourses we set out in the opening of this Element have also added to the perceived sense of crisis and the need for automation. If there is a misinformation and disinformation crisis on our hands, and Generative AI is going to lead to even more error-strewn or conspiratorial information circulating online, then an argument can be easily made that these issues are no longer able to be managed by purely human interventions. These fact-checking specific debates also fit into a much a wider narrative around content moderation and scale (Gillespie, Reference Gillespie2020) which both invokes and demands AI and automation as some sort of solution to managing the significant amount of content online. Early evidence, however, reveals that AI is being used increasingly as a way of augmenting fact-checking work, rather than replacing it entirely.

Critical work is still being done in universities, with researchers actively experimenting with technology to better understand how AI can assist fact-checkers. In a literature review on what they call automated fact-checking (AFC), Dierickx and colleagues (Reference Dierickx, Lindén and Opdahl2023) identify four different areas where AFC may be valuable: finding claims, detecting already fact-checked claims, evidence retrieval and verification. Guo and colleagues (Reference Guo, Schlichtkrull and Vlachos2022) offer a broadly similar framework but go further in outlining the specific technologies involved. For example, they describe the growing use of neural networks to help identify whether a particular claim is worth checking. The ability to incorporate additional context around the claim is incredibly beneficial, when compared to previous classification efforts which had to rely on fixed features, such as up-votes or named political entities such as parties or politicians. Other efforts have seen LLMs increasingly used to identify the veracity of certain claims. However, because AI systems are dependent on the datasets upon which they are trained – a phrase which has become something of a truism in recent years – the ability of AI to detect false information may be limited.

For the success of these systems, what is considered to be misinformation and disinformation to be censored must be agreed on. One significant shortcoming is the over-blocking of lawful content by AI through its over-inclusivity, henceforth potentially negatively affecting the freedom of expression through censorship of reliable content (Kertysova, Reference Kertysova2018). In the same vein, the systems also appear to be inaccurate, with Das and colleagues (Reference Das, Liu, Kovatchev and Lease2023) noting that ‘even the state-of-the-art natural language processing models perform poorly on […] benchmarks’, referring to the capacity of models to accurately detect claims worth checking.

Although the deployment of automation including especially AI in the context of fact-checking is still in its infancy, there is a growing effort to use the technology to assist both content moderation directly, as well as support downstream efforts in fact-checking organisations. As might be expected, the other area where AI is being deployed with a view to supporting content moderation in the future is in relation to deepfakes, where a similar tale of experimental academic research and gradual industry engagement is occurring. Both of these examples are discussed further in the following section.

2.4 Technical and Ethical Limitations

As mentioned earlier, AI systems are technically dependent on the datasets upon which they are trained, and its capabilities in detecting false information can be limited. Hence, it becomes imperative for the training dataset: to be large enough so that it can be adequately reflective for the AI to differentiate between false information and truth; to be updated regularly to counter new campaigns; and to contain text accurately labelled as false content (Marsoof et al., Reference Marsoof, Luco, Tan and Joty2023). Ascertaining which content is false can be hard, since online content may contain a mix of true, false, as well as true but misleading claims. Further, there may be opinions, speculations and predictions not intended by the author to be assertions of fact. These subtle differences both between what is true and what is false or misleading, as well as what are opinions and facts can only be identified with linguistic competence and background knowledge of more complex topics (Brasoveanu and Andonie, Reference 49Brasoveanu and Andonie2021), which trained AI may not have at hand. Prior studies have also demonstrated that reliably labelled datasets of false information are hard to come by (Asr and Toboada, Reference Asr and Taboada2019). Therefore, notwithstanding the promise of automated techniques, algorithmic moderation has the potential to exacerbate, rather than relieve, several key problems with content policy.

There are several tensions that need to be resolved. For example, when the removal of content by a ‘fallible’ human moderator is associated with scientific impartiality, it allows online platforms to both justify replacing people with automated systems as well as keep their decisions non-negotiable and non-transparent (Crawford, Reference Crawford2016). This worsens the problem that content moderation is infamously seen to be an opaque and secretive process (Gillespie, Reference 56Gillespie2018b; Flew, Suzor and Martin, Reference Flew, Suzor and Martin2019). The absence of a human-in-the-loop in fully automated decision-making systems is a dangerous progression. To elaborate on what has been said earlier, because machine learning is notoriously poor at making difficult context-dependent decisions, automated systems could both over-block important forms of expression, or under-block such content (including harmful content), thereby not protecting informational integrity either way. Having people in the loop moderates this risk and limitation of AI. As algorithmic moderation becomes more seamlessly integrated into regular users’ daily online experience, the reality behind the use of automated decision making in content moderation must continue to be questioned, so that online platforms remain accountable to the public and engaged with crucial content policy decisions (Gorwa, Binns and Katzenbach, Reference Gorwa, Binns and Katzenbach2020).

Unfortunately, some AI systems like large language models are unexplainable (or at least very difficult to explain), involving sophisticated layers of architecture that are almost impossible to analyse and explain – they are hence frequently referred to as ‘black boxes’ as there is little insight into how they are coded, what datasets they are trained on, how they identify co-relations and make decisions, as well as how accurate and reliable they are (Pasquale, Reference Pasquale2025a). If users do not understand why their content is removed due to the lack of transparency, accountability and explainability in platforms’ content moderation policies, their trust in online platforms may be eroded, together with their ability to challenge takedowns, thereby impacting on these users’ rights to online speech (Fischman-Afori, Reference Fischman-Afori and Schovsbo2023; Marsoof et al., Reference Marsoof, Luco, Tan and Joty2023). It has been suggested that some implementations of algorithmic moderation techniques threaten to: decrease decisional transparency, through making non-transparent practices even more difficult to understand; complicate outstanding issues of justice (with respect to biases and how certain viewpoints, groups or types of speech are privileged); as well as obscure the complex politics that underlie the practices of contemporary content platform moderation (Gorwa, Binns and Katzenbach, Reference Gorwa, Binns and Katzenbach2020). Given these critiques it has been suggested that the use of content-filtering AI should always be subject to mandatory human review (i.e. a human-in-the-loop), particularly in high-risk and low-accuracy scenarios (Marsoof et al., Reference Marsoof, Luco, Tan and Joty2023). And, to further improve accuracy, datasets relied on to train such AI should contain articles which are individually labelled by people with expertise on the topic of these articles (Asr and Toboada, Reference Asr and Taboada2019).

Digital platforms are making some efforts to improve the accuracy of AI systems. For instance, platforms like YouTube are partnering with external evaluators and using their inputs as training data for machine learning systems to build models that review large volumes of videos to detect harmful misinformation and disinformation. OpenAI has been working on developing detection classifiers to evaluate the likelihood that the content is AI generated (Soon and Quek, Reference Soon and Quek2024). Additionally, given the challenges of AI-based systems in understanding subtler forms of human expression and contextual cues beyond explicit content, such as sarcasm and irony that can give content further meaning (Santos, Reference Santos and Carrilho2023), the approach of semi-supervised learning – involving people in the text analysis process to develop AI tools for the automated detection of misinformation and disinformation – may ameliorate these difficulties and further help to improve the accuracy of such tools. Similarly, involving people in the actual detection of misinformation and disinformation may circumvent the challenges faced with respect to the accuracy of textual analysis programmes in identifying misinformation and disinformation, due to the difficulty of parsing complex and sometimes conflicting meanings from text (Marsden and Meyer, Reference Marsden and Meyer2019).

Additionally, while no longer seen as especially viable, blockchain technology has also been considered as a method of verifying the authenticity of digital content. Blockchains can provide a transparent and immutable record of authenticity. Examples of initiatives that use blockchain to combat misinformation and disinformation include the News Provenance Project, Fake Check and Democracy Notary. The New York Times, for instance, conducted a project to provide provenance metadata around news, using blockchain technology to track the dissemination and provide contextual information to readers of online news (Santos, Reference Santos and Carrilho2023). Separately, there are initiatives such as the DeepTrust Alliance which verifies the provenance and authenticity of digital content (through metadata), as well as WeVerify’s Deepfake Detector which maintains a public database of known fakes and uses blockchain to scrutinise social media and web content for disinformation and fabricated content (Harrison and Leopold, Reference Harrison and Leopold2021). As Fátima Santos concludes after an exhaustive study, combining blockchain technologies with AI’s capacity to analyse large volumes of content in real time can result in a more efficient system against misinformation and disinformation (Santos, Reference Santos and Carrilho2023). At the same time, she does not suggest that these tools can provide a complete solution, arguing rather that ‘the fight against disinformation requires a multifaceted approach that involves not only the use of AI and other technological tools but also human verification’.

2.5 Value and Limits of Technical Regulation

There are ongoing normative debates about the role that private entities should play in the regulation of private speech (Klonick, Reference Klonick2017; Balkin, Reference Balkin2018; Gillespie, Reference Gillespie2018a; Santos, Reference Santos and Carrilho2023). As Davis and Molitorisz put it (Reference Davis and Molitorisz2025: 3), ‘platforms are largely to control the dissemination of content on their services, and this lies at the heart of their market offering […] and their moderation of content often exhibits arbitrariness, inconsistency and deference to power rather than rights’. While we have been emphasising the tools available to platforms to address misinformation and disinformation in this section, our comments highlight the limits of leaving the choice whether to do so, or not, to private actors – whose practices may vary sharply, shaped by corporate goals of maximising user engagement and profit or political changes, and whose ideas about what and how misinformation and disinformation should be regulated may also be very different from those of their own communities.Footnote ²³ Indeed, even where decisions made by platforms seek to address misinformation and disinformation, technical tools may still be insufficient, especially if the end goal is alleviation of harm and/or vindication of some idea of rights. They may also be controversial. For instance, the use of AI and technological tools in fact-checking and content moderation processes often face technical and ethical challenges.

The question we now turn to is whether, given our arguments that reliance on law, and even platforms and their technologies will not be enough to address the harmful effects of misinformation and disinformation in balance with freedom of speech and innovation, social interventions can assist.

3.1 Beyond Law and Technology

In this section we outline the important role of tools and approaches which are predominantly social. We are mindful not just of the harms which are already being identified with misinformation and disinformation, including those being targeted by various laws (as discussed in Section 1) or highlighted in the moderation of platforms (as discussed in Section 2), but the need to build general social resilience so as better to head off future harms. The idea that tackling misinformation and disinformation for the future requires ‘a comprehensive or total defence system, in which every individual and organisation should play a role, including as checks and balances in the overall information ecosystem’ is not a new idea (see, e.g., OECD 2024: 74). But what is needed is a fuller understanding of the predominantly social tools and approaches that might be adopted within that system.

It is important also to set out exactly what we mean here. Some legal frameworks have avenues for ordinary people to activate, or even participate in, certain processes. For example, under the EU DSA (discussed in Section 1) people are empowered to flag illegal content on platforms under (article 16), have recourse to an internal complaint handling system that providers of platforms must provide (article 20) and can escalate unresolved complaints to an independent settlement body (article 21). In addition, there is scope for civil society groups such as fact-checkers or other NGOs to be identified as trusted flaggers with their notices ‘given priority […] processed and decided upon without undue delay’ (article 22(1)). And vetted researchers may be granted access to conduct research ‘that contributes to the detection, identification and understanding of systemic risks’ (article 40(4)). Platforms, which are required to conduct risk assessment and mitigation, are also encouraged to include input from communities (and the European Commission conducts its own consultation: see Jahangir, Reference Jahangir2025a). In this respect, EU law is very different from the law in the US, for instance. But such approaches offer a useful way of incorporating wider society into legal responses, ensuring that a range of actors are involved in platform governance.

However, there are some harms that are not easily accounted for by legal frameworks or technological solutions. One challenge that some scholars have begun pointing to already is the ‘foundational’ problem of people (individually and collectively) not knowing when to believe and when not to believe in the truth of information being circulated and the resultant lack of trust in all information (see Jungherr, Reference Jungherr2024). For instance, writing in 2019, Robert Chesney and Danielle Citron posited that the perverse product of such an environment where this phenomenon becomes widespread is the ‘liar’s dividend’ – here, the public dismisses genuine allegations of wrongdoing because the basis of truth has become muddled with that of falsity (Chesney and Citron, Reference Chesney and Citron2019). More recently, Chesterman argues that the harm here may not (just) be that people believe in lies, but that they cease to believe in anything at all (Chesterman, Reference Chesterman2025). Frank Pasquale suggests further that people may become demoralised, and hence too cynical or disengaged to participate in democracy (Pasquale, Reference Pasquale and Diurni2025b; cf. also Philip, Reference Philip2020; Kreps and Kriner, Reference Kreps and Kriner2023). At present this nebulous ‘harm’ arguably finds little support in current ethical standards and declarations of rights (although see Pasquale, Reference Pasquale and Diurni2025b, advocating for a ‘maximalist’ conception of rights). However, it is one that we believe will ultimately demand considered attention from policymakers and society more generally.

Below we detail more systemic social approaches to combatting misinformation and disinformation, including extending the role for trusted third parties to monitor and address harms, building digital literacy, and making space for continuing further research (including on the scale of the problem, the immediate and more fundamental harms that need to be addressed, and measures that may effectively address these without undermining freedom of speech). In our proposals we identify two features that we think are important guiding principles: first, a move away from the idea that contesting truth and falsity is the best way to deal with misinformation and disinformation instead centring on harm as the justification for intervention; and, second, devising an approach which is genuinely social and geared to communities rather than just relying on experts and elites to lead the way.

3.2 Devising Social Solutions

As we have already seen, third parties have emerged to battle misinformation and disinformation. Changes include the rapid growth in the fact-checking industry, as a way for digital platforms to help outsource their content moderation challenges, and journalists to reclaim their role as arbiters and authorities of establishing the truth of current events (Graves, Reference Graves2018; Cotter, DeCook and Kanthawala, Reference Cotter, DeCook and Kanthawala2022). Alongside professional fact-checkers, community volunteers have been deployed by platforms X and Meta and introduced by YouTube. These volunteers may make a useful contribution to reliable information as shown by the experience of Wikipedia (Tkacz, Reference Tkacz2014).Footnote ²⁴ Scholars have raised concerns that a community notes system could miss falsehoods and amplify hateful content (see Gilbert, Reference Gilbert2025), but there is some evidence that community notes are viewed as more trustworthy than expert-tagged misinformation flags (Drolsbach, Solovev and Prollochs, Reference Drolsbach, Solovev and Prollochs2024). Moreover, John Wihbey points to another benefit of community notes functioning as ‘“bridging” mechanisms that funnel disparate people together into more deliberative and cooperative tasks of producing counterspeech together’ (Wihbey, Reference Wihbey2025b: 75).

A body of evidence is also emerging to suggest that community notes work best on platforms in tandem with professional fact-checking. Through this partnership notes can appear faster on misleading posts, henceforth allowing misinformation and disinformation to be addressed before it spreads further (Nyariki, Reference Nyariki2025). While fact-checking has continued to struggle with combatting the spread of misinformation and disinformation on platforms, these collaborations suggest a promising way forward.

There are potential limits to the effectiveness of community notes though. Recent research has found that people who look to social media and commercial broadcasting (i.e., television or radio) for information, score lower on a civic values scale, than those who turn to newspapers, online news aggregators and non-commercial television. The scale measures the levels of trust in media institutions and the government, as well as people’s openness to considering perspectives that challenge their own (Andrejevic, Reference Andrejevic2025).Footnote ²⁵ As a result, while community notes may be seen as trustworthy by the public and even help fact-checkers, their effectiveness could be limited by the civic dispositions of the media audiences who are likely to be involved in the effort.

What then of platforms delegating responsibilities regarding content moderation to trusted entities? We have noted earlier the role of trusted professional fact-checkers supplementing (and acting as a check on) the activities of community actors. But also consider here the role of trusted entities which serve as dispute resolvers when platform content moderation decisions are challenged. Kate Klonick (Reference Klonick2017) points to Meta (previously Facebook) as having taken meaningful steps towards redressing issues through the development and launch of its Oversight Board, which enables user participation and offers something close to procedural fairness (Klonick, Reference Klonick2020). Further, the Oversight Board’s reliance on the international human rights standards of the United Nations’ International Covenant on Civil and Political Rights (1966) offers the promise of an international set of norms about fundamental rights (rather than privileging, for instance, US ideas about the overriding importance of free speech or European ideas about the need to vigorously regulate hate speech).

Could more be done to integrate the work of the Oversight Board’s experts with the broader community? As Evelyn Douek points out, the Oversight Board’s ‘material impacts’ on online speech are hard to track – and in high-profile, especially hard cases, ‘the Board has too often ducked its responsibility to act as decider of last resort and avoided giving a definitive answer altogether’ (Douek, Reference Douek2024: 379–80). On the other hand, we might wonder whether the Oversight Board’s indecisiveness is a harbinger of hybrid models that bring together experts and communities. In critiquing Oversight Board’s indeterminacy, Douek gives the example of the Board’s COVID-19 misinformation decision in response to Meta’s request for advice in 2022 about ‘whether Meta’s current measures to address COVID-19 misinformation under [its] harmful health misinformation policy continues to be appropriate’ considering the changing pandemic conditions. Its response was to advise Meta to continue its current policy but also begin a process to reassess the health-related claims Meta currently removes, through stake holder engagement. Douek questions here ‘wasn’t that exactly what Meta was asking the Board to do?’ (Reference Douek2024: 432). Yet, as Douek points out, given that COVID-19 misinformation is one of the most politically controversial topics in content moderation, we might see the justification for trying to bring in a range of knowledge and opinions to address the issue, with trusted experts.

A similar response might be offered to proposals from Andrew Kenyon for what he calls greater ‘democracy of expression’ (Kenyon, Reference Kenyon2021; and see also Kenyon, Reference Kenyon, Krotoszynski, Koltay and Garden2024), emphasising diversity of speech as a key to countering misinformation and disinformation. As Ian Cram argues, by excluding all but the rational, consensus-seeking and civilised speech, little contentious discourse will be let in (Cram, Reference Cram2019: 140). Kenyon further argues that ‘trusted public media’ should be able to be relied on to counter potential misinformation and disinformation with other reliable information, giving the examples of German and French public media (Kenyon, Reference Kenyon2021: 241). However, we might wonder whether public media are necessarily well placed to carry out this role if operating independently of communities they serve.Footnote ²⁶ Could diverse private media and community media – if found sufficiently credible by community standards – help to make up the difference?

On the one hand, both media play a particularly critical role when it comes to the amplification of misinformation and disinformation. Work by Axel Bruns and colleagues show how conspiracy claims associated with the spread of COVID-19 were amplified by mainstream media reportage (Bruns, Harrington and Hurcombe, Reference Bruns, Harrington, Hurcombe, Lewis, Govender and Holland2021). These views were largely shared within existing conspiracist networks until celebrities started to make similar claims, resulting in news and entertainment outlets reporting on the controversial opinions of the rich and famous. Reporting was led by ‘entertainment news outlets, and entertainment journalists within larger news organisations’, with stories adopting ‘an amused or dismissive tone’ rather than challenging the claims being made. On the other hand, statistical studies lead by Sacha Altay (Altay, Nielsen and Fletcher, Reference Altay, Nielsen and Fletcher2024) show that, in general, mainstream media news consumption correlates with a stronger awareness of false or contestable claims, and a lower susceptibility to embracing these by news consumers. In our view, these studies can sit in productive tension with one another. While we might view ongoing engagement with news as a useful balance to false beliefs (Altay, Nielsen and Fletcher, Reference Altay, Nielsen and Fletcher2024), media amplification of misinformation and disinformation still matters because we cannot expect all individuals to be regular, engaged news consumers. In that context, responsible reporting becomes key as a socially oriented solution, necessarily involving ongoing engagement with experts and communities if conspiracist claims are to be effectively challenged.

One option to support this kind of work might be to consider greater engagement with community-oriented mechanisms that have historically helped to build institutional trust, such as newspaper readers’ editors or ombudsmen. These were roles that either stood in for, advocated for or engaged with the audience, creating a link between the newspaper and the wider community (Ferrucci, Reference Ferrucci2019). Such positions have been gradually in decline, with reader comments, social media and external media commentary viewed as suitable replacements. While these developments have all been valuable in their own way, they do not offer formal accountability – as a result, news outlets have not really availed themselves of the opportunity to both learn from and respond to community concerns, particularly around reporting conspiracies or covering content in a sensationalist or controversial fashion.Footnote ²⁷

Of course, in our discussion of the news media – and particularly private media – we do not seek to diminish the baseline level of journalistic standards that most newsrooms adhere to, or the value of press codes more broadly. These codes can help to establish a reliable media ecosystem and generate an expectation of reliable speech and information. The problem is when institutional positions and self-regulatory instruments get used as proxies for trust rather than reflecting actual behaviour. Questions can be raised about what media do with their legitimacy. How do they handle misinformation and disinformation when it comes to their own reporting practices (as noted by Bruns and colleagues earlier)? Does sensationalism further threaten audience trust? As already discussed, loss of trust can separate the media from the audience quite dramatically. With people in some countries no longer actively consuming news content, maintaining this relationship becomes increasingly critical (Villi et al., Reference Villi, Aharonib and Tenenboim-Weinblatt2022). Consequently, the news media are presented with a wider challenge that requires them to look beyond established institutional regulatory patterns and reflect on the critical role they play in a disrupted information ecosystem.

Our earlier examples have shown the limits of institutional capacity and highlight the value of focussing on a rich diversity of civic practices that can contribute to a healthier media ecosystem. Instead of relying (just) on regulatory proxies or formal determinations, the most effective interventions appear to be community led and socially supported. However, this can only occur if audiences have media literacy (see, for instance, Coe, 2023), a topic we now turn to.

3.3 Improving Media and Algorithmic Literacy

Looking to community involvement in solutions, as well as relying on the abilities – and readiness – of audiences to recognise, and presumably then disbelieve, misinformation and disinformation when they come across it requires a degree of media literacy. In addition to efforts at fact-checking and providing alternative sources of information, some institutions are turning to pre-bunking, which involves building up people’s media literacy skills (Ecker et al., Reference Ecker, Lewandowsky and Coo2022) to increase their resilience to misinformation and disinformation.

There is moderate evidence for the effectiveness of media literacy training (Altay, Reference Altay2022) and some gamification strategies (Glas et al., Reference Glas, van Vught and Fluitsma2023; Birrer and Just, Reference Birrer and Just2024), which suggests that wider education strategies may be more beneficial, when anticipating any significant influx of incorrect information emerging from AI-generated or distributed content. There is much work to do here in some countries, with research into Australian educational settings showing that teachers are not well supported to embed media literacy into the curriculum (Corser, Dezuanni and Notley, Reference Corser, Dezuanni and Notley2022). However, recent work of the eSafety Commissioner on developing a ‘Best Practice Framework’ (eSafety Commissioner, 2024) may help to address that deficit. Moreover, progress is being made more generally, for instance in Europe (although with variation: see Pranaityte et al., Reference Pranaityte, Miller, Mesquita and Murphy2024), in some US states like New Jersey that mandate the teaching of media literacy from K–12 (Burney, Reference Burney2023), and with a widespread review in the UK suggesting a similar interventionFootnote ²⁸ (Badshah, Reference Badshah2024).

It is also important to look beyond the traditional focus on content. A growing number of researchers are highlighting the importance of algorithmic literacy, noting it is increasingly necessary to understand the technical systems and contexts through which content is circulated and promoted. Leyla Dogruel and colleagues (Reference Dogruel, Masur and Joeckel2022) provide a useful working definition, describing it as:

They go on to explore literacy rigorously through a validated scale, but literacy can also be spotted more generally, with certain people – such as professional content creators – developing robust ‘folk theories’ about how certain algorithms work (Dogruel, Reference Dogruel2021). This area is still developing though and perhaps in the same space that media literacy was a few decades ago, with approaches and interventions still at a very nascent stage of development (Oeldorf-Hirsch and Neubaum, Reference Oeldorf-Hirsch and Neubaum2025). That being said, algorithmic literacy is a critical educative step, that may help people understand not just how to spot problematic content, but why they came across it in the first place.

Private companies are taking incredibly small steps in this regard. Meta now provides a prompt on its feeds to explain to people why they are seeing certain advertisements. However, research has shown that the feature tends to suggest that targeting occurs through basic category matching rather than by the ‘dynamic, associative machine learning process’ (Burgess et al., Reference Burgess, Carah, Angus, Obeid and Andrejevic2024) that drives Meta’s infrastructure. The focus on individual targeting also ignores developments in advertising, where an embrace of Generative AI allows advertisements to iterate over time, engaging users effectively through changing in response to user preferences and behaviours (Carah et al., Reference Carah, Hayden and Brown2024). Similar efforts have been undertaken by TikTok, which provides equally vague insight into its content recommendation processes (Main, Reference Main2022). As a result, algorithmic literacy may only be enhanced through empowering individuals and communities to engage with these systems on their terms.

Regardless of whether these reforms eventuate, we believe there is a need to prioritise both media and algorithmic literacy. Leaving aside the challenges around the educational arena, adults also have limited knowledge with respect to media. A 2024 survey conducted by Tanya Notley and colleagues (Reference Notley, Chambers, Park and Dezuanni2024) in Australia have found that less than half of Australian adults are confident that they can check if information found online is true, or if a website can be trusted. More notably, only 33% of Australians feel that they know what media literacy even means (Notley et al., Reference Notley, Chambers, Park and Dezuanni2024). Similar studies conducted in 2023 by Ofcom in the UK reveal that since 2015 around 40% of respondents over 16 continue to state that they do not critically assess search results, with the overall percentage largely holding steady since then (OFCOM, 2023: 22). Around half the survey population also regularly do not identify or recognise the first few links as sponsored. These studies show that while the use of digital technology is widespread, there are still notable skills gaps around assessing and evaluating information. As noted earlier, this does not mean that people are liable to believe everything they read, but rather that they are clearly uncertain about navigating an increasingly complex information ecosystem.

That being said, while literacy efforts provide general support and protection for the population at large, there is evidence to suggest that these initiatives do not help the small proportion of people who actively seek out misinformation or come across it. A key reason for this lack of efficacy is that in specific contexts (such as politics), people are not motivated by accuracy but are instead seeking to accept information that confirms pre-existing beliefs (Nyhan, Reference 62Nyhan2020). Further insights into the complexity surrounding underlying motivations to access this sort of content were revealed in recent study on conspiracy theories in Australia and New Zealand. Matthew Marques and colleagues (Reference Marques, Ling and Williams2022) found that a belief in conspiracy theories correlated with relational and existential factors (such as less trust in government and others, as respective examples). However, there was no correlation with critical thinking, which meant that people were not turning to conspiracies to make sense of the world. Corrections or other forms of fact-checking are therefore likely to have a limited impact on these core groups of politically motivated misinformation consumers and conspiracy theorists, who may also be most prone to contribute to harms (Stockemer and Bordeleau, Reference Stockemer and Bordeleau2024).

In addition to the limited inoculative effect of improving literacy for high-risk populations, more expansive critiques can be made. As the earlier overview reveals, much of literacy discussions are largely focussed on skills, with education programmes in schools easily cast as capacity building interventions. It can be easy to embrace this sort of approach, which promises some sort of broad causal link between news literacy and the spread of misinformation and disinformation. However, notable media literacy practitioner and theorist Paul Mihailidis cautions against this, arguing for initiatives that foster a ‘civic intentionality’ (Mihailidis, Reference Mihailidis2018). Instead of stepping away from media and assessing it, he calls for ‘interventions […] designed to bring people together in support of solving social problems, reinventing spaces for meaningful engagement [and] creating positive dialog’ (p. 159). These efforts align with our call for a less instrumentalist approach to misinformation and disinformation.

In addition to enabling a range of people (including children, young people and adults of different backgrounds and political persuasions) to better participate and engage in social life, and support a healthier media ecosystem, these social solutions may offer opportunities for conspiracy theorists to reconnect with the mainstream. At very least, there is a case for reaching out to them if only to appreciate what the mainstream is up against and find ways to make their practices more transparent and counteract their influence. For instance, Allan Britton (Reference Britton2025) argues that media literacy programmes can foster diverse viewpoints and mitigate against echo chambers that reinforce conspiratorial beliefs. Of course, the risk is that the publics may come to embrace conspiratorial thinking through these processes. Perhaps, the best we can hope for is that ultimately the social conclusions they draw will come to be seen as wrong with the benefit of experience.

3.4 Making Space for Research

Bernhard Rieder and Jeanette Hoffman (Reference Rieder and Hofmann2020) note that platforms are inherent dynamic environments and argue that such ‘volatility … requires continuous observation’. Currently, many worries about misinformation and disinformation are anticipatory and speculative. Examples that do exist are often idiosyncratic examples with little context or knowledge provided about wider causal impacts. Even in the case of well-understood harms that relevant communities might agree need to be addressed, at least to some extent, in balance with freedom of speech and innovation, there can be differences at the margins. For example, political interference in elections might be a general social concern but while there is evidence of misinformation and disinformation circulating around elections, we do not always have a sense of its audience, overall impact or the source.

Observation won’t help us with all these problems, but it will take us some of the way there, allowing researchers to assess and analyse critical datasets and better understand the current ecosystem (cf. Wihbey, Reference Wihbey2025b: 75). Likewise, particular cases where elections were obviously subject to extreme pressure from external actors (as with the Romanian elections of 2024 and 2025 discussed in Section 1.3.3) provide useful case studies for researchers’ efforts as well as pointing to the practical need for ongoing sustained research into measures needed to avoid such events in the future. Such efforts are likely to sit productively alongside the sorts of monitoring done by state intelligence of external actors seeking to disrupt electoral processes.

As noted earlier, law can help here (see generally Lessig, Reference Lessig1999, Reference Lessig2006; Tan, Reference Tan2018). For instance, while article 16 of the EU DSA provides for trusted flaggers to be given priority by platforms within their designated areas of expertise (van de Kerkhof, Reference van de Kerkhof2025), article 40 provides for access by vetted researchers who conduct research ‘that contributes to the detection, identification and understanding of systemic risks’.Footnote ²⁹ However what we are arguing for in this section is participation by researchers at a level going beyond, and in ways that, laws may not easily (or need to) accommodate.

3.5 Enhancing Democracy

Social measures and tools can also act as a counter to the anti-democratic use of technologies (as well as laws). As we have seen whether one automates different elements of a fact-check or detects deepfakes using machine learning, the technical components rely on natural language processing, statistical modelling and probabilistic labelling to assess textual claims, or convolutional neural networks that can analyse audio and video signals. However, the outcomes from these technologies can be used to justify a normative model of communication and democracy that presumes ‘reason, rationality and truth’ to quote Johan Farkas and Jannick Schou (Reference Farkas and Schou2019: 5). Indeed, a narrative that we are contesting in this Element adheres to this presumption, gesturing to an idealised online environment before social media and Generative AI, where truth and falsities were (thought to be) easily identified. Farkas and Schou critique this ongoing focus on truth and in doing so, offer a diagnosis for democracy, arguing that beyond ‘facts, reason and evidence’, democracies are also about ‘giving voice to different political projects and groups’ (Reference Farkas and Schou2019: 7).

Our concerns about democracy are more pragmatic, but no less normative. We suggest that concerns about truth and falsity often sit outside of the embodied, emotional, multi-layered world of actual conspiratorial thinking and disenfranchised people that consume and spread misinformation and disinformation. We are not making claims of relativism here but rather point to ongoing issues associated with seeking to automate truth and veracity, when we are engaging with a deeply emotional terrain where correctness and correction may not even be the right paradigms from which to engage. While we have argued that community input can act as a democratic counter to the content moderation of platforms and technologies, along with (other) elites, a broader perspective would ask whether truth should even be our north star when seeking to address the policy problems at hand. As we are seeing with a range of legal interventions, harm is increasingly being viewed as a more productive mechanism through which to engage with misinformation and disinformation as a paradigm (as discussed in Section 1).

Could democratisation be taken further? A productive proposal comes from Lauren Naudts and colleagues (Reference Naudts, Helberger, Veale and Sax2025), who call for co-design initiatives as well as spaces for individuals and community groups to give feedback during a recommender system’s development, and for these actors, along with experts and regulators, to assess the performance of the system after the fact, and offers space for people to actively contribute to the development of recommendation systems. The logic of the ‘right to constructive optimization that promotes people’s effective enjoyment of fundamental rights and civic values’ that Naudts and colleagues (Reference Naudts, Helberger, Veale and Sax2025) propose might be extended further to embrace participation by individuals and communities in the design and development of systems adopted to address misinformation and disinformation. Further research will ideally assist in helping to flesh out and respond to such ideas.