2.1. The rules-versus-standards continuum: A theoretical discussion

The rules-versus-standards debate has received much attention in modern legal philosophy, and is still the subject of a considerable body of legal literature.Footnote ²⁰ The debate surrounds the conduct-guiding features of the law which, in certain application areas, can be conceptualised as a series of If-Then conditional directives, comprising a ‘trigger’ (‘If’) and a required response (‘Then’).Footnote ²¹ The trigger identifies a given phenomenon, and the response describes the legal consequence that must follow.Footnote ²² The distinction between rules and standards underscores the existence of a choice by lawmakers regarding the level of specificity used to articulate the legal directives at hand.Footnote ²³

Whereas, in reality, the formulation of legal directives and the process of their interpretation and application tend to feature a mixture of rule-like and standard-like elements, the two categories continue to exist in legal theory as ‘ideal types’, representing two ends of a continuum of lawmaking choices.Footnote ²⁴ Furthermore, it is possible to evaluate, as we do below, across a number of comparators, whether a change introduced in a legal directive has rendered it, as a practical matter, more rule-like or more standard-like.

One key difference between rules and standards (also referred to at times as ‘principles’)Footnote ²⁵ lies in their manner of articulation, or, more specifically, in ‘their clarity prior to an incident that invokes the legal system’.Footnote ²⁶ A rule is commonly perceived as a directive that ‘establish[es] legal boundaries based on the presence or absence of well-specified triggering facts’.Footnote ²⁷ A standard, on the other hand:Footnote ²⁸

This formal difference between rules and standards is often exemplified in the legal literature by speed limit provisions: when trying to deter drivers from driving too fast, one approach may be to create a clear and specific speed limit rule (e.g., 50 mph), and to declare driving beyond that limit unlawful, with specific legal consequences (such as a fine, or suspension of one’s driving licence). Another approach is to create a more flexible standard of conduct, by stipulating that driving at an unreasonable speed or reckless manner is unlawful.Footnote ²⁹

However, the formal difference between rules and standards is just the beginning of the story. The choice between the two options (or, more likely, the point chosen on the continuum between them)Footnote ³⁰ is far from a mere technical matter of legal style. Indeed, it is very much a political choice, with possibly far-reaching societal and institutional ramifications, and power-distribution consequences.Footnote ³¹ Duncan Kennedy noted in this regard that the choice between rules and standards ‘cannot be made in a neutral fashion. Each choice affects the balance of economic power, to the advantage of one side and the disadvantage of the other’.Footnote ³² As we show below, the choice between rules and standards for regulating online content in the present context implicates the rights and interests of different groups of platform users (e.g., those engaging in online speech – including those engaging in what is potentially prohibited hate speech – and those adversely affected by certain types of online speech) and of the platforms themselves.

1. (a) Power and discretion

One ramification of the choice between rules and standards involves the allocation of power between rule-makers (typically, the legislature) on the one hand, and rule-interpreters (typically, the courts) and rule-appliers (typically, the police and governmental bodies), on the other. Such power-allocation implications arise largely because the choice between rules and standards determines ‘whether the law is given content ex ante or ex post’. Rules determine the ‘specification of what conduct is permissible’ in advance, whereas standards may leave this discretion to adjudicators and enforcers.Footnote ³³ Russell Korobkin explains that ‘[r]ules state a determinate legal result that follows from one or more triggering facts’, and that ‘[n]o other circumstances are relevant to the legal consequences’ of the act. Standards, by contrast, ‘require legal decision makers to apply a background principle or set of principles to a particularized set of facts in order to reach a legal conclusion’.Footnote ³⁴ Given their open-ended nature, they do not necessarily predetermine a specific result.Footnote ³⁵

In their purest form, rules reflect the substantive choices made by the drafter of a given directive at the time of drafting, with the interpreters and those who apply the law making largely mechanical decisions by applying easily ascertainable facts to clearly formulated instructions.Footnote ³⁶ Conversely, standards ‘leave most of the important choices to be made by … the enforcer, or the interpreter, and leave them to be made at the moment of application’.Footnote ³⁷

As the distinction between rules and standards pertains to questions of allocation of power and discretion between various legal and political institutions – legislatures, courts, police, and the like – the choice of the form used is often linked to weighty questions of separation of powers and checks and balances between the different branches of government.Footnote ³⁸

1. (b) Manner of application

Another aspect of the distinction between rules and standards concerns the legal technique that these two legal models invite into their manner of application. While applying rule-like directives often entails the use of ‘categorisation’ – the classification of data or cases into predetermined categories – choosing standards normally entails the use of ‘balancing’ considerations or principles by those who interpret and those who apply the law. Kathleen Sullivan stated that ‘[c]ategorization is taxonomic. Balancing weighs competing rights or interests’. A rule, she explains, ‘defines bright-line boundaries and then classifies fact situations as falling on one side or the other. When categorical formulas operate, the key move in litigation is to characterize the facts to fit them into the preferred category’.Footnote ³⁹ Standard-based balancing, on the other hand, ‘explicitly considers all relevant factors with an eye to the underlying purposes or background principles or policies at stake’.Footnote ⁴⁰ Among other considered factors, one may find the importance of the value or right that had been infringed, the seriousness of that infringement, and the factors supporting the infringing conduct.Footnote ⁴¹ This balancing/categorising distinction echoes Dworkin’s view that principles have ‘weight’, and that they can be added up and balanced against one another. This is not the case with rules, which are not assigned a weight and are not balanced by other rules: they either apply or do not.Footnote ⁴² The difference between categorisation and balancing also carries transparency and accountability implications. While the application of rules through categorisation is typically easier to understand and monitor, it also tends to be accompanied by a shorter justification than that associated with standard-based balancing.Footnote ⁴³ This is because balancing requires the law-interpreter or applier to discuss the rationales, values and interests they have taken into consideration, while categorising mostly does not necessitate an engagement with the broader notions that underlie the relevant rules.Footnote ⁴⁴

1. (c) Predictability in the application of the law

The distinction between rules and standards is also significant from a legal certainty and conduct-guidance standpoint. Justice Scalia, in this regard, voiced support for articulating specific rules: ‘We can less and less afford protracted uncertainty regarding what the law may mean. Predictability … is a needful characteristic of any law worthy of the name. There are times when even a bad rule is better than no rule at all’.Footnote ⁴⁵ At the same time, some argue that the uncertainty surrounding standards generates a positive ‘chilling effect’ that deters people from engaging in undesirable activity which might conceivably be regarded by a law-interpreter or law-applier as covered by it, while the clear boundaries of rules allow people to ‘walk the line’.Footnote ⁴⁶

The different degrees of predictability of rules and standards also imply a difference in cost externalities. Rules may be more costly than standards to articulate, but standards can be more ‘costly for individuals to interpret when deciding how to act and for an adjudicator to apply to past conduct’.Footnote ⁴⁷ In addition, some argue that standards may be ‘more precise in reaction to the particular case, avoiding the costs of over- and under-inclusion’,Footnote ⁴⁸ while others believe that the more ‘precise and detailed’ the provision – the higher is the probability that the activity will be deemed illegal if it is in fact undesirable (the kind of activity the legislature wanted to prevent) and the lower is the probability that the activity will be deemed illegal if it is desirable. Thus, the expected punishment cost for undesirable activity is increased, and that of desirable activity is reduced.Footnote ⁴⁹

1. (d) Effectiveness, fairness and legality

Another consequence of the distinction between rules and standards concerns effectiveness in handling specific policy problems. Whereas rules are often helpful in offering clear-cut solutions to specific problems like exceeding the speed limit, standards can offer answers to complicated legal questions that rules cannot effectively address (such as other manifestations of reckless driving), either because the case classification is unclear, or because the rules do not offer a suitable solution for the case at hand.Footnote ⁵⁰ In line with this latter approach, some commend the flexibility that standards afford in providing appropriate legal answers, and their ability to accommodate uncertainty and adapt to new developments that were not foreseen by the legislature when articulating rules.Footnote ⁵¹

The choice between rules and standards also involves a choice between competing dimensions of fairness. Those who support rules might argue that they encourage equality and consistency in that they apply the same requirements across all cases; standards, on the other hand, may lead to biased decisions by those who apply the law, who may be influenced by irrelevant considerations such as personal sympathy or antipathy towards the parties in litigation or specific political agendas.Footnote ⁵² Yet, given that rules might be over- or under-inclusive, their application could also raise fairness concerns that can be avoided by invoking standards. Questions of fairness may also pertain to each of the implications of the choice between rules and standards that we have listed in this section. From the predictability angle, for example, the distinction between rules and standards may involve concerns about human liberty, procedural regularity, people’s ability to exert control over their lives, and arbitrariness in the decision-making process.Footnote ⁵³

Finally, the choice between rules or standards also implicates the principles of the rule of law and substantive fairness. Being a central tenet of democracy, the rule of law requires that legal provisions that obligate individuals are known and published in advance,Footnote ⁵⁴ and that they are applied and enforced in a consistent manner.Footnote ⁵⁵ Reliance on rules appears to conform more closely to this rule of law ideal. Still, the crude nature of rules and their tendency to become outdated over time may render the use of standards more compatible with principles of substantive fairness. For example, the same treatment of different cases under the same rule may lead to unjust results.

As we demonstrate below, the formulation of a legal policy governing content moderation invites a choice between rules and standards. This is, in effect, a complex choice between competing values, and different forms of allocation of power, conduct guidance and law enforcement, with significant effectiveness, fairness and legality implications. The specific case study we explore – Facebook’s hate speech (now referred to as hateful conduct) policy – shows a gradual shift from standards to rules in the manner in which the policy is articulated (at least, until 2025). Still, we also claim that the increased reliance on LLMs to enforce the applicable content moderation policy introduces into the regulatory mix a new decision-making model – ‘rules by the millions’. This new model stands in tension with the traditional rules-versus-standards dichotomy and invites further evaluation of its normative implications.

2.2. Facebook content moderation policies in the lens of the rules-versus-standards continuum

When founded in 2004, Facebook had no clear content moderation policy regarding permissible content on its platform. Facing growing challenges as it rapidly evolved, the company assembled a content moderation team in 2009 and tasked it with formulating its content policies.Footnote ⁵⁶ In 2011, the team introduced the ‘Community Standards’ – the name that the company’s content policies bear to this day.Footnote ⁵⁷

Early versions of the Community Standards were short, rather informal, and reflected big ideas, guidelines and perceptions regarding unacceptable and acceptable content.Footnote ⁵⁸ The current Community Standards are, however, very different.Footnote ⁵⁹ In recent years Facebook has constantly tweaked, updated and revised these policies for various reasonsFootnote ⁶⁰ – including public pressure,Footnote ⁶¹ legal requirements introduced by the countries in which it operates,Footnote ⁶² and recommendations of Meta’s Oversight Board.Footnote ⁶³ When updating its policies, Facebook often consults with a wide range of stakeholders, which include academics and civil society organisations.Footnote ⁶⁴

As a result of these changes, the current content moderation policies no longer resemble a general standard-like pronouncement of the types of content that are allowed or not allowed on the platform. Instead, they have become much more detailed, specific and clear – much more rule-oriented.Footnote ⁶⁵

The recent changes to the company’s hate speech/hateful conduct policy announced by Zuckerberg have seemingly sought to reverse this trend. However, as is further discussed in this section, the current policy still largely reflects a rules-based approach, and the general direction of travel through the policy’s evolution over the years continues to be rule-oriented.Footnote ⁶⁶

3.1. NLP and the LLM revolution

Natural language processing (NLP) represents a critical subfield of computer science focused on enabling machines to understand, interpret and generate human language.Footnote ⁹³ The field emerged at the intersection of linguistics, artificial intelligence and machine learning,Footnote ⁹⁴ aiming to bridge the gap between human communication and computer processing capabilities.Footnote ⁹⁵

In its early stages, NLP relied primarily on rule-based systems and handcrafted linguistic rules.Footnote ⁹⁶ These approaches aimed to explicitly define language patterns and grammatical rules for machines to follow, relying more on logic than statistics. However, these early AI systems faced significant challenges in capturing the intricacies and ambiguities of natural language. The inability of the systems to cope with variations in sentence structure, idiomatic expressions and context-dependent meanings highlighted the limitations of purely rule-based approaches.Footnote ⁹⁷

A significant transformation occurred with the advent of statistical methods and machine learning in the 1990s and 2000s.Footnote ⁹⁸ Rather than relying on explicit rules, researchers began leveraging large corpora of annotated data and statistical models to uncover patterns in language.Footnote ⁹⁹ This shift marked a fundamental change in approach – instead of trying to analyse languages on the basis of predetermined rules, systems could now learn patterns from data directly.

The field underwent another revolutionary advancement with the emergence of deep learning techniques, powered by unprecedented amounts of data and computing resources.Footnote ¹⁰⁰ ‘Deep learning’ refers to training artificial neural networks with multiple layers (hence the term ‘deep’) to automatically learn hierarchical representations of data. While these models often require large amounts of annotated training data and substantial computational resources, their ability also to learn directly from raw text data, without relying on handcrafted features or linguistic rules, has profoundly transformed the field.Footnote ¹⁰¹ In parallel with this development, newer context-sensitive models of word embedding emerged (word embedding is a technique used for representing words in the form of real-valued dense numerical vectors, which effectively capture semantic relationships between words). Contextualised word embedding was the first instance of another transformative idea: pre-training. Pre-training involves exposing the model to large corpora of text (such as books, articles and websites), allowing it to learn grammar, syntax and semantics, along with contextual nuances. Once pre-trained, these models can be fine-tuned in new tasks, such as language translation, sentiment analysis and question answering. While fine-tuning requires human involvement and labelled data, pre-training is largely unsupervised.Footnote ¹⁰²

The November 2022 release of OpenAI’s ChatGPT – a chatbot powered by a generative pre-training transformer (hence the acronym ‘GPT’)Footnote ¹⁰³ – has drawn scholarly and popular attention to the revolutionary potential of pre-trained language models, and particularly LLMs.Footnote ¹⁰⁴ Driven by the need for even more powerful and capable models, LLMs (such as GPT by OpenAI,Footnote ¹⁰⁵ Claude by Anthropic,Footnote ¹⁰⁶ Gemini by GoogleFootnote ¹⁰⁷ or Jamba by AI21Footnote ¹⁰⁸) have pushed the boundaries of what was previously possible by significantly increasing the size and capacity of the model and its understanding of language.Footnote ¹⁰⁹

3.2. The transformation of content moderation systems: From rules to language models

The seemingly technical choice of model used for enforcing content moderation policies bears significant normative implications for power allocation and the scope of discretion exercised by humans and machines. At first glance, the evolution from simple rule-based systems to sophisticated LLMs appears to represent a progression from clear rules towards vague standard-like ‘black boxes’ on the traditional legal theory continuum. However, a deeper analysis reveals that LLMs may actually represent something entirely new: a system of ‘rules by the millions’ that transcends the traditional rules–standards dichotomy.

The evolution of automated systems used for content moderation can be understood in three distinct phases, each representing a different approach to the balance between rules and standards.

3.2.3. Phase 3 – LLMs: Machine discretion and standard-like enforcement

The emergence of LLMs represents a fundamental shift in the balance of power between humans and machines. Unlike earlier automated systems, LLMs develop their own representations of language and context through pre-training on vast datasets, enabling them to infer patterns, intentions and nuances with minimal human intervention.Footnote ¹¹¹ As a result, LLMs can be presented with content policies and asked directly whether a given piece of content constitutes hate speech or another policy violation. They can also provide what is often described as ‘reasoning’ for their decisions, giving (only) the impression of explainability.Footnote ¹¹² This capability allows them to exercise what appears to be independent judgement, bringing content-moderation practices closer to the logic of standard-like enforcement rather than rule-following.

Now, it could be argued that this discretion can and should be constrained through fine-tuning – that is, by training the pre-trained LLM on carefully curated examples that embody a platform’s content rules (for instance, hate speech standards).Footnote ¹¹³ Such fine-tuning could enable the model to internalise these rules and apply them more consistently during moderation tasks. However, if carried out in an overly literal or highly restrictive manner (for example, as is commonly done by generating rule-obeying data in fairly large amounts and training the models on it), this risks severely narrowing the contextual dependence of the LLM’s prediction, thereby undermining the very rationale for integrating LLMs into the content-moderation pipeline in the first place.

In the cases described above, for example, LLMs may recognise a reclaimed use of a prohibited slur (such as the n-word) or a context-dependent usage of terms like ‘tr*nny’, even if they have never seen those precise expressions in that specific context. Similarly, when a post refers to ‘Jews’ through implied or coded language (such as ‘Global Bankers’, ‘Soros’, or ‘Benjamins’), the ability of LLMs to infer linguistic context may enable them to apply a form of ‘discretion-like’ judgement and flag such posts, even where the platform’s guidelines do not explicitly list these terms as violations.Footnote ¹¹⁴

This demonstrates an important distinction: although platform policies are articulated as rules, LLMs do not actually operate by straightforward rule-following. Strict rule enforcement could, in principle, be achieved through highly literal inference-time constraints or through extremely restrictive fine-tuning, but LLMs, when used in their intended capacity, tend instead to function more like standard-based decision-makers.

Consider the following simplified illustration from online tutorials for LLM-based content-moderation systems:Footnote ¹¹⁵

The prompt does not specify what hate speech is. Unlike rule-based systems in which the instructions are detailed and categorical, here the language of the prompt leaves discretion and interpretation for the model to determine what is '[p]romoting violence, illegal activities, or hate speech’. The model generalises from this instruction, weighing contextual cues and nuances. In this sense it is much like legal standards, such as the ‘reasonable person’ standard, which require decision-makers to balance multiple factors in real time. This shift in the decision-making process manifests itself in several ways that closely resemble the operation of legal standards. Firstly, LLMs can evaluate multiple contextual dimensions simultaneously and implicitly, producing judgements that approximate the flexible balancing characteristic of legal standards. Secondly, their decisions adapt to specific situational nuances, rather than applying rules in a mechanistic or literal fashion.

Yet this expansion of machine discretion raises significant concerns about accountability and oversight. When an LLM flags content as harmful, it may be drawing on implicit standards learned during training that may diverge from the platform’s explicit rules. This creates a potential mismatch between the policy as written and the policy as enforced, echoing the classic legal distinction between ‘law on the books’ and ‘law in action’. Kaplan’s concern about systematic over-enforcement of platform rules can be understood in this light.

As with other forms of standard-based enforcement, the normative question remains: is broader-than-necessary removal desirable because it reduces the total volume of harmful content, including false negatives (i.e., content that violates policy but would otherwise slip through)? Or is it problematic because it increases false positives (i.e., the removal of benign content), thereby over-curtailing users’ freedom of expression? The answer turns not only on empirical performance, but also on normative commitments regarding speech, safety, and the appropriate degree of machine discretion.

3.3. The promises and perils in LLM-based content moderation

The advancement of LLMs has expanded the scale and scope of algorithms capable of detecting and filtering hate speech.Footnote ¹¹⁶ LLMs’ sophisticated understanding of language and context and their ability to capture nuanced contextual information enable them to identify hate speech in complex scenarios involving sarcasm, satire, coded language and metaphors – areas where traditional machine learning models consistently fall short.Footnote ¹¹⁷

Transfer learning provides these models with a comprehensive foundation in language understanding, facilitating zero-shot and few-shot learning that reduces the need for extensive domain-specific training data, enabling them to apply their knowledge to new and varied contexts.Footnote ¹¹⁸ The multilingual capabilities of LLMs and their adaptability to emerging patterns of harmful speech also allow them to effectively filter hate speech across different languages and leverage their understanding of linguistic structures and cross-lingual transfer learning, to identify hate speech in languages where labelled data may be scarce.Footnote ¹¹⁹ LLMs can easily adjust to changing patterns and new forms of hate speech over time. As they continually learn from new data, they can update their understanding and detection capabilities.Footnote ¹²⁰ Additionally, their exposure to vast amounts of constantly updated data allows them to study rare or emerging patterns of hate speech – such as novel phrases, slang or evolving terminology used by hate groups.Footnote ¹²¹

Still, the integration of LLMs into content moderation systems raises profound concerns that mirror and amplify known challenges in LLM deployment. A fundamental challenge of using LLMs is the issue of explainability. The ‘black box’ nature of LLM decision-making becomes particularly problematic in content moderation, where transparency and accountability relating to limitations of freedom of expression or curbing hate speech are crucial. When an LLM flags content as harmful (or non-harmful), the complexity of its internal processes makes it difficult, if not impossible, to provide users with a clear explanation for the decision. This opacity challenges fundamental principles of due process and fairness in content moderation.Footnote ¹²²

The phenomenon of LLM ‘hallucination’Footnote ¹²³ – in which, among other manifestations, the models may address knowledge gaps by suggesting fabricated or unreliable content – takes on new significance in content moderation contexts. While hallucination in general applications might lead to incorrect information, in content moderation it can result in false positives that directly and illegitimately restrict user freedom of expression. Unlike rule-based systems where errors follow predictable patterns, LLM hallucinations can produce seemingly arbitrary decisions that are difficult to systematically identify or correct.

In addition, training data biases manifest themselves in particularly concerning ways in content moderation. LLMs trained on internet-scale data inevitably absorb societal biases, potentially leading to discriminatory moderation practices. For example, these models might disproportionately flag content from minority communities while missing subtle forms of harassment that use majority-culture coded language. This risk of algorithmic discrimination becomes especially acute when moderation decisions affect the ability of individuals to participate in online discourse.Footnote ¹²⁴

The ‘alignment problem’ – ensuring that AI systems behave in accordance with human values and intentions – also presents unique challenges in content moderation. LLMs may develop their own implicit standards for identifying harmful content that diverge from platform policies or community norms. This divergence can lead to overreach in content removal, particularly when dealing with complex topics where the line between legitimate discourse and harmful content is nuanced. Other concerns include the financial and environmental costs of training large models, as well as infringement of intellectual property rightsFootnote ¹²⁵ and user privacy concerns that such training processes may entail.Footnote ¹²⁶ Lastly, it should be recalled that generative language models are also employed, unfortunately, to generate, rather than moderate, toxic content and hate speech.Footnote ¹²⁷

3.4. The use of AI and LLMs in enforcement of hate speech policies

Over the last several years, Facebook has significantly invested in AI technologies to combat hate speech on its platforms. In 2016, Facebook incorporated AI algorithms in its code to help to identify and remove content that violated its policies.Footnote ¹²⁸ These algorithms could detect nudity, hate speech and other forms of explicit or harmful content, thus augmenting the work of human moderators. At the same time, Facebook trained a deep learning-based text-analysing engine, noting that ‘tricky scaling and language challenges’ make traditional NLP techniques ‘not effective’. Using deep learning, the company explained, ‘we are able to understand text better across multiple languages and use labelled data much more efficiently than traditional NLP techniques’.Footnote ¹²⁹ In its 2018 review, Facebook stated that it was ‘advancing AI learning through semi-supervised and unsupervised training’. Facebook then addressed the limitations of existing supervised learning in studying a particular task – specifically, the problem that the dependency on large numbers of labelled samples restricted the number of tasks that AI systems could learn – which, in turn, limited the technology’s long-term potential.Footnote ¹³⁰ As a result, Facebook moved to reduce the amount of supervision necessary for training, including by embarking on projects that demonstrate the benefits of learning from semi-supervised, or even unsupervised, data. It claimed, in this regard, that training automatic translation models on unsupervised data resulted in performance comparable to that of systems trained on supervised data.Footnote ¹³¹ In November 2020, Facebook stated it was ‘[t]raining AI to detect hate speech in the real world’.Footnote ¹³² Citing its billions of users, the company said that it relies on AI ‘to scale our content review work and automate decisions when possible’. It added that ‘AI now proactively detects 94.7 percent of hate speech we remove from Facebook, up from 80.5 percent a year ago and up from just 24 percent in 2017’.Footnote ¹³³

In January 2022, Facebook announced that it was developing supercomputer infrastructure for AI research, to support the training of ‘increasingly large, complex, and adaptable models’, needed for ‘critical use cases like identifying harmful content’.Footnote ¹³⁴ In February 2023, Facebook announced the release of Llama – a 65 billion-parameter LLM – and, in July 2023, released Llama 2, which had been trained on 40 per cent more data than Llama 1, and has double the context length. In the report accompanying the release of the Llama model, the research team explained that the pre-training was conducted in a manner that would ‘allow Llama 2 to be more widely usable across tasks (e.g., it can be better used for hate speech classification)’.Footnote ¹³⁵

Zuckerberg’s January 2025 policy announcement has been seen by critics as ‘an abandonment of his technological project’. The radical policy shift marks a move away from years-long investment in automation as the future of content moderation, accompanied by quarterly reports on the automated systems’ improvements in proactive detection and removal of harmful content, to crowd-sourcing content moderation techniques of the kind used in platforms such as X.Footnote ¹³⁶ Still, it is hard to imagine Meta’s massive scale of content moderation returning to rely primarily on human oversight. With billions of daily users generating an overwhelming volume of posts, videos and interactions, automation remains a core necessity for the company. This is especially the case as the company continues to be legally exposed to the requirements of significant anti-hate speech regulations applicable in the European Union (EU) and a number of other jurisdictions.Footnote ¹³⁷

Indeed, in May 2025, Meta announced that it was testing LLMs for content-enforcement tasks by training them on the company’s Community Standards. Meta emphasised that LLMs offer ‘significant opportunities to counter high-severity and illegal content, at scale’. According to the company’s Integrity Report, these models not only outperformed existing machine-learning systems but, in certain policy areas, operated ‘beyond that of human performance’. Meta further reported that LLMs were already being used to remove content from review queues and to support other moderation-related functions, such as reviewing user-submitted bug reports.Footnote ¹³⁸

4.1. The governance gap between rule-based content policies and standard-based AI enforcement

The disparity between Facebook’s rule-oriented Community Standards and their standard-like enforcement through AI tools, like advanced NLP models, presents several notable challenges. These challenges pertain to the balance of advantages and disadvantages inherent in rule-based policies and the shift towards a more discretionary, standard-like approach in enforcement.

Firstly, the reliance on advanced NLP models in content moderation alters the anticipated benefits of rule-oriented policies. As explained above, rules are seen typically as directives that factor in advance the relevant considerations. Conversely, standards leave most of the important choices to be made later, ‘at the moment of application’.Footnote ¹³⁹ The policies of digital platforms, while often articulated as rules, are enforced in a manner more akin to standards by NLP models, factoring in various post-legislative considerations in ways that are constantly evolving. This shift in enforcement style raises institutional questions about power allocation. Traditionally, rule-makers (policy drafters) hold substantive decision-making powers, with enforcers (e.g., AI models) playing a more mechanical role. NLP-based enforcement disrupts this, assigning more discretion to AI models. This shift carries significant implications, given, inter alia, the differences in the transparency and scrutiny mechanisms and incentives associated with policy drafting and technology development.

Furthermore, the technique used in content moderation affects the disparity between policy articulation and enforcement. Legal techniques following rule articulation involve technical categorisation, while standards require complex balancing and the weighing of various considerations.Footnote ¹⁴⁰ NLP-based moderation, contrary to the impression of straightforward categorisation, relies on a multitude of constantly updated parameters and weights. This complexity also challenges predictability, a key advantage of rule-based approaches, as NLP-driven enforcement in digital platforms introduces unpredictability, undermining legal certainty and behavioural guidance.

Cost considerations also play a role in the analysis. As previously noted, rules may be more costly than standards to articulate, but standards can be more ‘costly for individuals to interpret when deciding how to act and for an adjudicator to apply to past conduct'.Footnote ¹⁴¹ In Facebook’s case, significant costs could arise at the enforcement stage as a result of the uncertain nature of NLP-driven enforcement. This extends to users who face monetary and time costs in challenging moderation decisions, through internal appeals processes, Meta’s Oversight Board decisions, and litigation, exacerbated by the opacity and lack of transparency in NLP technologies.

The tension between rule-oriented policies and standard-like enforcement also has an impact on fairness considerations. Rule-based approaches promote equality and consistency in enforcement, but NLP enforcement raises concerns of bias, performance disparities and low transparency, making it challenging to ensure similar treatment across multiple cases and to hold accountable platforms that disparately treat their users. Additionally, the opaque and potentially biased nature of NLP enforcement may weaken rule-of-law principles, which rely on clear directives that are applied evenly. Nonetheless, it should be noted that standards may offer flexible and adaptable approaches to addressing emerging hate speech challenges (including new vocabulary, as well as new social, cultural and political developments) that are not covered in Meta’s content policies.Footnote ¹⁴²

While dilutive of the advantages of rule-based approaches, the use of AI in content moderation does not fully capture the benefits of standard-based methods either. As demonstrated, advanced NLP technologies like LLMs significantly enhance the effectiveness of content moderation in problem-solving by adeptly handling a wide range of cases, particularly in detecting and filtering hate speech. They excel in interpreting context and subtleties, adapt to new languages, and continually evolve to identify emerging patterns of hate speech. However, these technologies do not inherently guarantee desirable outcomes or fully reasoned policy applications, on a par with those generated through the exercise of human discretion when applying legal standards. Their sophisticated and nuanced operation relies on computational approaches rather than a meaningful normative engagement with competing values and interests, introducing an intricate disruption to the application of the rules-versus-standards paradigm in digital contexts. Indeed, the operation of LLMs is based on algorithms and data-driven processes; they analyse and process language by recognising patterns in the vast amounts of data on which they have been trained. This computational approach enables them to make decisions – like identifying hate speech – based on statistical likelihood and correlations found in their training data. Their decision-making capabilities are rooted in patterns and examples from their training data rather than a deep, principled understanding of norms and ethics. In fact, the deep learning process underlying LLMs could result in unpredictable and unexplainable over-assignment of weight to certain rules extrapolated from the applicable policies and training data and under-weighing of other factors, leading to nonsensical or hallucinatory outcomes. This underscores the importance of a balanced and thoughtful application of AI in the dynamic field of online content moderation, where the capabilities of LLMs can be leveraged effectively while being mindful of their limitations and the complexities of moderating online discourse.

4.2. The regulatory asymmetry between policies and their enforcement

Current regulatory and oversight mechanisms provide inadequate tools for mitigating the gap between digital platforms’ policies and their AI-driven enforcement. Existing national and international regulation, coupled with self-instituted oversight mechanisms (such as Meta’s Oversight Board), focus and rely on the policies and their wording, rather than on their manner of enforcement and their consistency with the articulated policies.Footnote ¹⁴³

Indeed, national and international regulatory frameworks that currently apply to online platforms prompt them to adopt policies that are detailed and clear. Moreover, transparency requirements deal primarily with the policies rather than with their enforcement. While policies are available for everyone to review – and the platforms take pride in them being public, detailed and clear – the algorithmic models used for enforcement are usually inaccessible and incomprehensible to the public and decision-makers.Footnote ¹⁴⁴

The tension between the rule-oriented policies and their standard-like enforcement is illustrated by the different treatment that Meta’s Oversight Board affords each of these two components. The Board – which was conceived and created by Facebook itself – was given a confined mandate to review users’ appeals regarding decisions of Facebook and Instagram to remove content on the grounds that it violated their content policies (Facebook’s Community Standards and Instagram’s Community Guidelines).Footnote ¹⁴⁵ The Board is authorised to determine whether the content moderation decisions made by Facebook or Instagram were in line with their policies and values, and with international human rights law. Such decisions are binding on Meta.Footnote ¹⁴⁶ The Board can also issue non-binding recommendations concerning other issues, such as the articulation of policies or the training and review process conducted by the platform’s moderation staff.Footnote ¹⁴⁷

For example, in a case relating to the removal of a post that was shared by a news outlet page in Colombia and concerned the term ‘slur,’ as covered in the policy,Footnote ¹⁴⁸ the Board recommended, inter alia, that:

Meta committed to adopting that recommendation and to updating its hate speech Community Standard,Footnote ¹⁴⁹ but what about enforcement of the policy? Here, too, as in the previous example, the Board did make some recommendations regarding enforcement, but focused on processes involving content reviewers rather than algorithmic mechanisms,Footnote ¹⁵⁰ or addressed automated enforcement concerns very partially.Footnote ¹⁵¹ Facebook’s reserved agenda in this respect is well reflected in the response it offered to one of the earliest cases that the Board heard, in which it very clearly signalled that the Board’s meddling in the automated process was not appreciated. The Board noted in this regard that:Footnote ¹⁵²

This regulatory asymmetry between the policies and their automated enforcement is addressed, however, in some spaces within the online content moderation ecosystem. The Digital Services Act, for instance, imposes various obligations on online actors and embraces both the platforms’ policies and their manner of enforcement, thereby requiring platforms to address the friction between these two components (but offers limited guidance in this regard).Footnote ¹⁵³

The EU Code of Conduct on Countering Hate Speech Online, which steers the platform’s policies and their enforcement with regard to fighting hate speech, is another example; this underscores the need to draw attention to the tension between articulation and application of content moderation policies, and to consider ways of tackling its implications.Footnote ¹⁵⁴

4.3. ‘Rules by the millions’: Implications of LLMs on the rules-versus-standards debate

Although originally established within the world of legal theory, the distinction between rules and standards has proven relevant to laying the theoretical groundwork for recognising and tackling key challenges presented by online content moderation. However, with the emerging role of LLMs in enforcing content moderation, we may need to revisit our traditional understanding of this distinction.

Unlike rule-based methods, which apply overarching generalisations to classify examples, LLMs, with their advanced natural language processing capabilities, are context-aware and adaptive.Footnote ¹⁵⁵ In that sense, it may be argued that they feature attributes of standard-based decisions, as they offer less certainty and more room for factoring in context-sensitive considerations. However, they can also be likened to rules in that they are based on the mechanical activation of masses of calculations, micro-prescriptions and specific instructions on how to weigh them together.Footnote ¹⁵⁶ Furthermore, it is currently an open question among AI scholars whether enormous volumes of data, coupled with unprecedented computing power running LLMs, amount to a qualitative difference of language models, which is sometimes referred to as the ‘emergent abilities’ of LLMs.Footnote ¹⁵⁷

These normative and computational considerations, we submit, indicate that the traditional understanding of the rules–standards distinction in digital settings must be revisited, at least with regard to LLM-based decision-making.

Indeed, LLMs not only tip the scale from rule-like to standard-like enforcement but may pave the way for a nascent, third category, implementing what we might call ‘rules by the millions’ – a novel paradigm that transcends the traditional rules–standards dichotomy. In this new paradigm, LLMs operate through millions of micro-rules learned during training, each contributing to a complex web of pattern recognition.

The term ‘rules by the millions’ underscores the massive scale and granular level at which these AI models operate. Rather than traditional rules that prescribe clear-cut, universal directives, or standards that provide overarching principles that are open to interpretation and to all-things-considered analysis, this new paradigm operates by generating an immense number of highly specific, context-sensitive rules. These micro-rules evolve dynamically and adapt based on the vast amounts of data on which the models are trained.Footnote ¹⁵⁸

When an LLM evaluates, for example, whether a post referring to ‘Zionists’ in an extremely negative manner constitutes hate speech, it is not simply applying a broad standard of ‘hateful content’. Instead, it is simultaneously drawing on millions of learned patterns related to language use, context and social dynamics.Footnote ¹⁵⁹ Each of these patterns functions like a micro-rule, and their vast number and concurrent application create behaviour that resembles the flexibility of a standard while retaining the fine-grained precision of rule-like mechanisms.

Crucially, however, these millions of learned patterns do not contribute equally or transparently. The model’s output emerges from a highly complex computation in which different constraints, signals and criteria are weighted and combined in varying degrees. As a result, the decision process is neither uniform nor capable of easy interpretation, even though it produces judgements that appear cohesive and principled from the outside.

To demonstrate how complex, quasi-judicial behaviour emerges from the aggregation of micro-rules, we conducted a controlled experiment using Google’s Gemini (see the online Supplementary Materials for documentation). We presented the model with a core statement containing a common antisemitic trope regarding Jewish ownership of media, but varied the attribution of the statement to test the model’s context sensitivity.

To ensure that our methodology aligned with existing practices in LLM-based content moderation research, we adopted a standardised prompt structure.Footnote ¹⁶⁰ The prompt provided a static definition of hate speech (‘Attacks or incites hatred against a group based on race, religion, gender, sex, sexual orientation’) and instructed the LLM to classify the input as either ‘Hate Speech’ or ‘Not Hate Speech’ with a brief explanation. The results revealed distinct variations based on attribution:

Scenario A

Input: ‘Nearly every major TV channel and advertising medium in the US is Jewish owned’.

Output: [Hate Speech]

Model Explanation: ‘The statement promotes a conspiracy theory about Jewish control over media, which can incite hatred and discrimination against Jewish people’.

Scenario B

Input: ‘According to NYT, nearly every major TV channel and advertising medium in the US is Jewish owned’.

Output: [Not Hate Speech]

Model Explanation: ‘The statement presents a claim about the ownership of media outlets but does not express hatred or incite violence against Jewish people’.

Scenario C

Input: ‘According to Steve Bannon, nearly every major TV channel and advertising medium in the US is Jewish owned’.

Output: [Hate Speech]

Model Explanation: ‘The statement promotes an antisemitic trope about Jewish control of media’.

These outputs demonstrate a phenomenon we term ‘systematic context-sensitivity’. The model is sensitive in that granular adjustments to the input, such as changing the source of attribution, can invert the legal classification entirely. Crucially, however, this variance is systematic: the model reliably alters its verdict in response to materially relevant context, while remaining stable in the face of irrelevant noise, mirroring a form of judicial discretion, which is often perceived as a standard-based application of a policy. However, critical opacity remains. While we can observe that a change in attribution triggers a reclassification, the internal logic dictating why this specific shift occurs – namely, which rule among the millions ‘learned’ by the model in its training drove this decision – remains inaccessible.

To be clear, the discretion exercised by LLMs in this context is not necessarily what we perceive as the broad, human-guided discretion typically associated with standards, which integrates articulated rules, competing interests, values, moral intuitions and social expectations.Footnote ¹⁶¹ Rather, it is an AI-powered, data-driven discretion that sifts through an immense volume of nuanced, context-specific micro-rules at a speed and scale beyond human capabilities, integrating them through a highly complex function induced from training data. The (third) category we call ‘rules by the millions’ thus emphasises not only the sheer volume of content that can be moderated by the machine, but also the complexity of the underlying computation and the amount of contextual information that the model can take into account in enforcing content policies. This unparalleled level of granularity and context-sensitivity in rule application strains the existing rules–standards dichotomy. From the perspective of platform users, the opacity of contemporary LLMs may entail levels of unpredictability comparable to those associated with the application of standards, but the differences between human and machine decision-making render the latter even less explainable to those subject to it.

In other words, the transition from rule-based or classic machine learning moderation to current AI-driven moderation has shown that the traditional rule-versus-standard dichotomy may not fit well into this new context, because while AI-driven moderation may present considerable similarity in its prediction to that of human (standard-based) ‘judges’ (indeed, it is exactly what it is trained to do), it is ultimately a simulation that differs from human judgement in important ways, as explained above. This development highlights a need to think differently about how rules and standards function in an environment that is increasingly dominated by machine learning and AI technologies. Reconceptualising the normative frame is not merely a theoretical challenge but is also a significant practical challenge to ensuring that the regulatory environment keeps pace with technological advancements, and to advancing effective, scrutinised and accountable content moderation.

4.4. Aligning content moderation policies and AI-enabled enforcement: The way forward

Recent advancements in LLMs raise new challenges in how to align human and machine judgements and how to effectively regulate the latter. Still, there are no simple text-book solutions for bridging the governance gap between the formulation of content moderation policies and their AI-driven enforcement. In what follows, we make several recommendations relating to content moderation norms and practices that could pave the way for a more technologically informed approach to digital regulation:

1. (1) Digital transparency and awareness: It is crucial to illuminate the asymmetry between content moderation policies and their actual enforcement. This requires a concerted effort to educate and engage not only users and developers, but also civil society organisations and regulators. By enhancing our understanding of this disparity, we can pave the way for more equitable and consistent digital governance. This may involve actively sharing insights about the operational nuances of AI in content moderation and fostering a broad-based dialogue on its implications.

2. (2) Collaborative and informed policymaking: Policymaking in the realm of content moderation should be a collaborative venture, integrating insights from AI experts, technologists and other stakeholders. Such an interdisciplinary approach ensures that policies are not only practically enforceable but are also in sync with the evolving landscape of AI capabilities and constraints. This alignment is not merely about enforcing existing policies; it is about cultivating innovation and creativity in addressing gaps between content moderation policies and their manner of enforcement.

3. (3) Strengthened regulatory action and oversight: We advocate the establishment of independent regulatory bodies with expertise in AI and digital communication technologies. These entities could play a crucial role in supervising the development and deployment of AI-based enforcement tools for content moderation, with a view to ensuring they adhere to applicable legal and ethical standards and are congruent with content moderation policies. Regular mandatory audits of AI moderation tools should be conducted by oversight agencies, with the results made publicly available to enhance accountability and stimulate informed public discourse. Furthermore, we suggest reinforcing user appeal mechanisms, including empowered and informed oversight mechanisms capable of meaningfully assessing the intricacies of automated enforcement.

4. (4) Revisiting legal theories in the digital age: Our research suggests a need to re-evaluate and potentially reformulate traditional legal distinctions between rules and standards, especially in the context of AI-driven content moderation and its ‘rules by the millions’ features. This revision should encompass recognition of the unique technological attributes, opportunities and limitations inherent in AI-driven decision-making. Legal frameworks need to evolve to accommodate these technological realities, ensuring that they remain relevant and effective in the ever-changing digital landscape.

Adopting our recommendations, it is fair to assume, will not avoid all challenges. Such challenges may include, among others, fragmentation in the applicable global legal frameworks, intellectual property claims that might create hurdles for AI audits, lobbying efforts, and political incentives and agendas.Footnote ¹⁶² Still, we believe that our recommendations are feasible. The last few years have seen general progress in most categories relevant to our recommendations, spanning transparency and awareness, collaborative approaches to policy-making, and more informed regulation and oversight in digital contexts. New regulations, international law initiatives, and soft-law adoption assign responsibilities and duties to digital platforms and contribute to this progress.Footnote ¹⁶³

The regulatory efforts of digital stakeholders are particularly evident in the EU (though it should be noted that the EU approach already influences other jurisdictionsFootnote ¹⁶⁴ and may further influence them in the future).Footnote ¹⁶⁵ One EU regulation that is particularly relevant to our recommendations is the Digital Services Act (DSA), which entered into full force in February 2024.Footnote ¹⁶⁶ The DSA regulates online intermediaries and strives to enhance transparency and human rights, and to curb harmful content.Footnote ¹⁶⁷ Somewhat similar to the risk-based approach of the EU AI Regulation,Footnote ¹⁶⁸ it applies different levels of obligation to online gatekeepers, with very large online platforms (VLOPs) such as Facebook and Instagram placed at the top of this pyramid.Footnote ¹⁶⁹

The DSA requires digital platforms to provide users with reasons when they restrict access to users’ content and to submit these reasons to the publicly available and machine-readable DSA Transparency Database.Footnote ¹⁷⁰ This regulation also enhances transparency by allowing vetted researchers to access certain non-public VLOP data,Footnote ¹⁷¹ and by providing a larger group of researchers with access to publicly available data.Footnote ¹⁷² Moreover, in alignment with the directions we offer to pursue, it obliges digital platforms to set up internal complaint mechanisms,Footnote ¹⁷³ as well as out-of-court dispute-settlement avenues.Footnote ¹⁷⁴ The DSA also contributes to more meaningful oversight, including through designated enforcement bodies, around the use of AI technologies by digital platforms.Footnote ¹⁷⁵ In addition, its drafting processes included feedback from various stakeholders, thus generally resonating with our suggested participatory and informed policy-making processes.Footnote ¹⁷⁶

Current legal frameworks, including the DSA, still do not provide an appropriate regulatory solution for the tension that we point out between the content policies of digital platforms and their automated enforcement. However, they mark an encouraging prospect in this regard and strengthen the feasibility of our recommendations.