Written evidence submitted by Meta (SMH0037)
Response to the Science, Innovation and Technology Committee inquiry into social media, misinformation and harmful algorithms
0.1. Meta welcomes the opportunity to respond to the Science Innovation and Technology Committee’s inquiry into social media, misinformation, and the role of algorithms.
0.2. Every day, billions of people use Facebook, Instagram, Messenger and Threads to share their experiences, connect with friends and family, and build communities.
0.3. Keeping the people that use our services safe is our top priority.
0.4. For Meta to be successful, our users need to have a safe and enjoyable experience on our services and the businesses, both large and small, who advertise with us must feel their brand reputation is secure. Harmful content including hate speech and threats of violence fundamentally undermines that and in doing so risks undermining our core business model. This is why we have around 40,000 people working on safety and security, with more than $20 billion invested in teams and technology in this area since 2016, including $5 billion in the last year alone.
0.5. Like many across the UK and around the world, we were shocked by the horrific attack in Southport and the outbreak of rioting and violence which followed it over the summer. That is why we worked quickly to establish a dedicated team to work round the clock to help identify and remove content that violated our rules, to complement our algorithmic systems which use highly effective AI to find and remove violating posts 24/7.
0.6. It is also the reason Meta has long been a supporter of the Online Safety Act as an example of the type of systems-based, outcome-focused regulation that we believe can be most effective at keeping people safe while protecting their rights to free expression. We are working closely with Ofcom to help them bring it into force over the next year.
0.7. While the Online Safety Act took some time to pass, Meta has not waited for this legislation to take action. In addition to the investment set out above we have twenty years’ experience in tackling these issues through establishing policies, building tools and technologies, and producing guides and resources, all in partnership with experts both within and outside our company.
0.8. On misinformation specifically, we want everyone who comes to our platforms to find high quality information that is meaningful and valuable to them. Inevitably we attempt to strike a balance of giving people the ability to express themselves while preventing people from using our services to carry out harm. We have no incentive to have harmful or low quality content on our platforms.
0.9. In this submission, we provide answers to your specific questions. We also set out our approach to combatting content with the potential to cause harm, including in crisis moments such as the UK riots and outline how our algorithms are designed to reduce the spread of misinformation and harmful content.
1.1. Meta’s overriding incentive is to ensure the people who use our services have a positive and enjoyable experience - our business model relies entirely on this. Our users tell us in surveys that they don’t want to see harmful content, companies that advertise on our services do not want their adverts to appear next to it, and we do not want this content on our platforms.
1.2. The existence of this content fundamentally undermines our core value proposition to the people who use our services - to provide a safe enjoyable way of connecting with people. As a result we have invested significantly over many years to address a range of harm types and have done so in advance of the Online Safety Act coming into force.
1.3. Since 2016 we have quadrupled the size of the teams working on safety and security to about 40,000 people, this includes around 15,000 content reviewers, who are located across the globe, in every major timezone. We have invested more than $20 billion in teams and technology in this area and this investment is only increasing -$5 billion was spent in the last year alone.
1.4. These investments go towards developing and enforcing our publicly available global set of Community Standards and Advertising Standards. These are the rules which govern what is and is not allowed on our platforms and which give us the basis for taking action and removing content where we need to. In developing these policies, we consult hundreds of civil society organisations and academics from around the world. When applying these policies, we seek to find a balance between safety and allowing users to have a voice.
1.5. Of relevance to this inquiry, these policies include strict rules on:
● Hate speech: We do not allow hate speech on Facebook and Instagram. We define hate speech as violent or dehumanising speech, statements of inferiority, calls for exclusion or segregation based on protected characteristics or slurs. In Q3 2024, we removed over 95% of the hate speech before someone reported it – up from 24% in 2017.
● Violence and incitement: We aim to prevent potential offline harm that may be related to content on Facebook and Instagram. We remove content, disable accounts and work with law enforcement when we believe there is a genuine risk of physical harm or direct threats to public safety. In Q3 2024, we removed over 98% of the violence and incitement before people reported it.
● Dangerous Organizations: We do not allow any organisations or individuals that proclaim a violent mission or are engaged in violence to have a presence on Facebook and Instagram. This includes organisations or individuals involved in far right or terrorist activity, organised hate or other dangerous organisations. We also remove content that expresses support or praise for groups, leaders or individuals involved in these activities. We remain vigilant in learning about and combatting new ways people may try to abuse our platforms. We work with external partners to get the latest intelligence about adversarial behaviour across the internet, and we commission independent research from academics and experts. In Q3 2024, of the violating content we actioned for dangerous organizations we removed over 99% before people reported it.
1.6. Where something violates our rules we will take various forms of action depending on the circumstances, including but not limited to applying warning labels, removal of content and banning of accounts. Each quarter we publish the Community Standards Enforcement Report to track our progress and demonstrate our continued commitment to making our services safe and inclusive.
Question 1a: How do social media companies and search engines use algorithms to rank content, how does this reflect their business models, and how does it play into the spread of misinformation, disinformation and harmful content?
1.7. Far from spreading misinformation, disinformation and harmful content, algorithmic classifiers are an essential part of our ability to combat these kinds of content. Algorithmic ranking is an essential component of how we show users the posts that are relevant to them, and remove content which can be harmful.
1.8. We use algorithms via state-of-the-art machine learning systems to ensure people have the best possible experience on our platforms. At any given moment there could be a few thousand posts that could potentially appear in that person’s feed, so we offer people a personalised experience where content is tailored to an individual’s interests. This is possible through the use of algorithmic tools, including the Feed ranking system, which consists of multiple layers of AI that we apply in order to predict the content that’s most relevant and meaningful for each user. We give people the option of only seeing chronological content if they do not want algorithmic recommendations and we have also introduced a feature where people are able to reset their recommendations, which we detail further below. For us, what’s most important is that people feel like their time on Facebook, Instagram and Threads is time well spent.
1.9. As described above, algorithmic systems play an important role in highlighting and elevating the content that a user is most likely to find meaningful and engaging. But engaging content that drives long-term value for our users is not the same as addictive, provocative sensational content. As a result, in 2018 we changed our ranking for Feed to prioritise meaningful social interactions. The change led to a decrease of 50 million hours a day worth of time spent on Facebook.
1.10. Similarly, although political content only makes up about 3% of what people see on Facebook, users tell us they want less of this type of content. That’s why over time, we began refining our approach on Facebook to reduce the amount of political content people see in Feed and recommendations. We extended this approach to Instagram and Threads, where we have aimed to avoid making recommendations that could be about politics or political issues, including within In-Feed Recommendations.
1.11. The impact of these changes were recently referenced in an article by BBC Journalist Marianna Spring who ran an experiment using fictional characters to explore how social media platforms could influence undecided voters in the US election. She concluded that as a result of our political ranking changes, the content on Facebook and Instagram remained largely apolitical.
1.12. We're always listening to our users and evaluating our approach for changes we can make in response to their feedback that will improve their experience.
Using technology to find, remove and reduce harmful content
1.13. We have pioneered the use of machine learning to detect and remove harmful content more efficiently, given the vast volume of content posted on our services. Although these tools are not perfect, our Community Standards Enforcement Report shows that year on year there is continuous improvement with this technology delivering highly successful outcomes when it comes to finding and removing harmful content.
1.14. While we have more work to do, we are making significant progress.
● Harmful content remains a very small portion of content users see on our platforms.
● Prevalence of hate speech - i.e. the amount of hate speech people actually see - is very low, estimated to be around 0.02% in Q3 2023. In other words, for every 10,000 content views, we estimate just two will contain hate speech.
● Hate speech content removal has increased over 15X on Facebook and Instagram since we first began reporting it.
1.15. We have also begun testing Large Language Models (LLMs) by training them on our Community Standards to help determine whether a piece of content violates our policies. These initial tests suggest the LLMs can perform better than existing machine learning models. We're also using LLMs to remove content from our content moderation review queues in certain circumstances when we're highly confident it doesn't violate our policies. This frees up capacity for our reviewers to focus on content that's more likely to break our rules, as they can now concentrate on evaluating posts that are assigned to them in the review queue
1.16. We take a different approach to content that our systems predict likely break our Community Standards but have not been confirmed as violations, as well as content that doesn’t break our rules but is considered problematic or low-quality, such as clickbait and engagement bait. We use algorithms to reduce distribution in Feed, as outlined in our Content Distribution Guidelines. These reductions, also called demotions, are rooted in our commitment to responding to user feedback, incentivising creators to invest in high-quality content, and fostering a safer community.
Using technology and partnering with fact-checkers to tackle misinformation
1.17. Using algorithms to reduce problematic content is a key part of our approach to addressing misinformation. We remove misinformation where it is likely to directly contribute to the risk of imminent physical harm. We also remove false claims that are likely to directly contribute to interference with the functioning of political processes.
1.18. For all other misinformation, in the past year we’ve worked with a global network of more than 100 independent fact-checking organisations, who review content in over 60 languages on Facebook, Instagram and Threads (since April). This is the largest fact-checking network of any platform and in the UK includes Full Fact, Reuters, Logically and Fact Check Northern Ireland. To ensure high standards and accuracy, all of our third-party fact-checking partners are certified through the non-partisan International Fact-Checking Network (IFCN) and follow IFCN’s Code of Principles.
1.19. When a fact checker rates something as false, our systems are set up to use technology to reduce its distribution so fewer people see it and add a warning label with more information. We also notify the person who posted it and anyone else who previously shared it, and reduce the distribution of Pages, Groups, and domains who repeatedly share misinformation.
1.20. Since the field of misinformation is always changing, we have teams dedicated to constantly assessing our approach to ensure it’s effective, and are constantly updating and improving our approach as the ecosystem changes.
1.21. We also provide resources to increase media and digital literacy so that people can decide what to read, trust and share themselves. We require people to disclose, using our AI disclosure tool, whenever they post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered.
Using technology to counter disinformation
1.22. When it comes to disinformation (understood as the deliberate intent to mislead or manipulate) and influence operations, our approach is to find and stop 'inauthentic behaviour'. This includes coordinated campaigns and adversarial actors that often also engage in the spreading of disinformation. Our inauthentic behaviour policy is intended to protect the security of user accounts and our services, and create a space where people can trust the people and communities they interact with. To deal with inauthentic behaviour we have tools that leverage machine learning and Artificial Intelligence to identify suspicious content using our A, B, C (Actor, Behaviour, Content) detection model. This comprehensive approach ensures that we can flex and evolve our systems as bad actors adapt their behaviours.
1.23. We’ve found that one of the best ways to fight this behavior is by disrupting the incentives structure behind it. Therefore, we’ve built teams and systems to detect and enforce against inauthentic behavior tactics such as clickbait, fraud and inauthentic spam accounts. We take a hard line against fake accounts and block millions each day, most of them at the time of creation. In 2023, we disabled more than 2.6 billion of them, 99% of these were removed before users reported them to us.
1.24. More specifically, we address influence operations, the most complex forms of inauthentic behavior, via our coordinated inauthentic behavior (CIB) policy, which we define as coordinated efforts to manipulate public debate for a strategic goal, in which fake accounts are central to the operation. When we investigate and remove CIB operations, we focus on behavior rather than content—in this sense, it does not matter who is behind them or what they post. We have removed over 200 networks of coordinated inauthentic behavior since 2017, including 20 this year alone, from locations such as Russia, Iran, and China. Despite our efforts, people continue to look for new ways to mislead people, which is why we continue to take steps to make it harder for them to do so and constantly improve our detection and enforcement systems. When we find and remove this activity, we identify the tactics used and we build tools into our platforms to make those tactics more difficult at scale.
Transparency and control
1.25. Ensuring that people have the ability to control their own experience online is at the heart of what we do. This means giving them both the ability to personalise their social media feeds, through our customisation tools and also to take steps to ensure their own safety. It is a myth that algorithms leave people powerless over the content they see.
1.26. People who use Facebook and Instagram can tailor the content that appears in their Feeds in multiple ways, from easily choosing to snooze, unfollow, or block unwanted accounts to showing their Feeds from select accounts or in different orders on Facebook, including chronologically. They can control the ads and other content they are served and amend or remove the information we are using to tailor their experience. This month we’ve also started testing the ability for everyone on Instagram to reset their recommendations in Explore, Reels and Feed when they want a fresh start. We’ve also introduced the ability for users to toggle how much content they want to see from categories that we typically reduce the distribution of by using our sensitive content controls. Finally, on Instagram we’ve introduced controls for political content.
1.27. We are also transparent about how our feed algorithms work. Those transparency initiatives range from videos explaining the ranking process to detailed blog posts in the Meta Transparency centre explaining significant ranking changes. People can use our ‘Why Am I Seeing This?’ tool to get more information about why any given post or ad was shown to them and implement controls to see more or less of this type of content or block posts or ads from particular accounts. We’ve also released 22 system cards for Facebook and Instagram which give information about how our AI systems rank content, some of the predictions each system makes to determine what content might be most relevant to people, as well as the controls you can use to help customize your experience. They cover Feed, Stories, Reels and other surfaces where people go to find content from the accounts or people they follow. You can find out more and see visuals here.
Question 1b: What role do generative artificial intelligence (AI) and large language models (LLMs) play in the creation and spread of misinformation, disinformation and harmful content?
1.28. Emerging technologies including generative AI and large language models present both challenges and opportunities in addressing misinformation, disinformation and harmful content. As outlined above, at Meta, we have long used AI to enhance the accuracy and speed of identifying and removing harmful content. However, we remain vigilant about the challenges Generative AI presents.
1.29. At the start of this year, many people were warning of the potential impact of generative AI on the upcoming elections, including the risk of widespread deepfakes and AI-enabled disinformation campaigns. While there were instances of confirmed or suspected use of AI in this way, the volumes remained low and our existing policies and processes proved sufficient to reduce the risk around generative AI content. Across major global elections, ratings on AI content related to elections, politics and social topics represented less than 1% of all fact-checked misinformation.
1.30. Like any technology, AI can be used for bad as well as good. While it may amplify the volume and convincingness of misinformation, we believe that its benefits in supercharging our detection efforts will ultimately outweigh these risks. By leveraging AI to enhance our detection capabilities, we can focus on identifying not just the content itself, but also the actors and behaviours behind it, as outlined in the A, B, C model. This approach will allow us to stay ahead of malicious actors, even as they attempt to exploit AI for their own purposes.
1.31. To address these challenges all of Meta’s existing policies - such as our Community Standards and Advertising Standards apply to content generated by AI and we take action against any content violations.
1.32. As set out in more detail later in the paragraphs below, we have also taken a number of steps to ensure where content is created with AI that people viewing that content can easily see its origin. For example we are already applying “AI Info” labels to photorealistic images created using our Meta AI feature. However it's important we are able to do this with content created with other companies’ tools too.
Labelling AI Generated Content
1.33. For content that doesn’t violate our policies, we still believe it’s important for people to know when photorealistic content they’re seeing has been created using AI. That is why we already label photorealistic images created using Meta AI.
1.34. We’ve also been working with industry partners to align on common technical standards that signal when a piece of content has been created using AI. Being able to detect these signals will make it possible for us to label AI-generated images that users post to Facebook, Instagram and Threads. We’re building this capability now, and in the coming months we’ll start applying labels in all languages supported by each app.
1.35. While companies are starting to include signals in their image generators, they haven’t started including them in AI tools that generate audio and video at the same scale, so we can’t yet detect those signals and label this content from other companies. While the industry works towards this capability, we’re adding a feature for people to disclose when they share AI-generated video or audio so we can add a label to it and we may apply penalties if they fail to do so.
1.36. If we determine that digitally created or altered image, video or audio content creates a particularly high risk of materially deceiving the public on a matter of importance, we may add a more prominent label, so people have more information and context.
1.37. For misinformation specifically, AI-generated content is eligible to be fact-checked by our independent fact-checking partners and we label debunked content so people have accurate information when they encounter similar content across the internet. One of the rating options is “Altered”, which includes, “Faked, manipulated or transformed audio, video, or photos.” When it is rated as such, we label it and down-rank it in feed, so fewer people see it.
2.1. We, like many across the UK and the world, were deeply shocked by the tragic events in Southport and the subsequent riots. Our thoughts remain with the victims and all those affected by these incidents.
2.2. In response to the riots, we activated our Crisis Policy Protocol and established a dedicated task force to address the situation. This team worked tirelessly to identify and remove content that violated our policies, collaborating closely with UK fact-checking partners, government, and law enforcement agencies.
Using algorithms to take action on violating content during riots
2.3. Our technology played a crucial role in mitigating misinformation and harmful content during the riots.
2.4. Our independent third-party fact-checking partners in the UK, including Reuters Fact Check and Full Fact, debunked trending false claims about the riots. This included false claims about the name of the suspect in Southport, their nationality and status as a refugee, the alleged two-tier policing of protesters, the arrest of Tommy Robinson, and the claim amplified by Elon Musk that rioters would be detained on the Falklands. When fact-checkers rated this content as misinformation, a clear label was applied and content was shown lower in people’s feeds.
2.5. We recognise the importance of speed in moments like this, so we took steps to make it easier for UK fact-checkers to find and rate content related to the riots, using keyword detection to group related content in one place.
2.6. Independent third-party fact-checkers published 15 fact-check articles and directly rated more than 150 pieces of content. We used technology to automatically match their ratings to further pieces of content.
2.7. Additionally, we used technology to find and remove content that violated our community standards.
● During the unrest, we removed 24,000 posts from Facebook and Instagram in the UK for breaking our rules on violence and incitement—and our systems found 98% of those posts proactively / before they were reported to us.
● We removed 12K posts under our hate speech rules, and found 94% ourselves
● We removed 2.7K posts under our rules on hate organisations: 92% ourselves
2.8. Through these measures, our technology and protocols played a positive role in reducing the spread of harmful content during the crisis, distinguishing our platforms from others where such content was more prevalent.
Crisis Protocol Implementation
2.9. Recognising the gravity of the situation, we swiftly established an internal working group as part of our crisis policy protocol (CPP) to manage the crisis. This protocol guides our response to events of mass violence, allowing us to adapt our content review processes to the heightened risks.
2.10. The incident was classified as an event of multiple victim violence under our Dangerous Organisations and Individuals policy. Our content reviewers were instructed to remove any glorification, support, or representation of the event and its perpetrator from our platforms.
2.11. Within hours of the attack, we were in close contact with the UK government and the police. Our law enforcement engagement team worked throughout the weekend to support Police in the UK, including on the investigation into the stabbings. Also working with counter terror police and with Greater Manchester police to provide data on accounts they believe could be advertising protests that could lead to disorder.
2.12. Meta platforms did not experience the same level of riot organization or coordination than other platforms.
2.13. We have also banned key members of the online far right, including Tommy Robinson, the English Defense League and Andrew Tate, since 2019. As outlined above, as part of our dangerous organisations policy, they cannot have a presence and any glorification, support, or representation of them is removed when we find it.
2.14. As well as removing content from our own platforms, we had to take action to block activity from platforms with less stringent controls and stop it infiltrating users' feeds. This included blocking links to 22 Telegram riot coordination groups on our platforms.
3.1. The introduction of the Online Safety Act represents a significant step forward in addressing the complex challenges posed by online harm. Meta has long supported the development of this framework, sharing the UK Government's objective of making the internet safer while protecting its vast social and economic benefits.
3.2. Over the last several years, we have worked with both the UK Government and Parliament to support the development of the Act. We have been vocal that new and innovative regulatory frameworks must strike a complex balance between safety and people's rights, such as freedom of speech and providers need to take their share of the responsibility to support how to strike this complex balance.
3.3. We have welcomed Ofcom's approach throughout the consultation process to develop this regulatory framework by focusing on how to implement the safety duties efficiently, as well as the clear objective to develop guidance and Codes of Practice based on the principles of proportionality and collaboration. We have been supporting them to get ready for this exercise for the last three years and are continuing to assist them by responding to their consultations.
3.4. Only once Ofcom has issued final versions of the guidance and Codes of Practice, and after an implementation period, will the obligations apply. As this is a large, novel, and complicated piece of legislation, implementation will take time, but Meta has not waited for this process to complete before taking action - indeed many of the measures and controls set out above have been in place for years to help ensure the safety and security of the people who use our platforms.
3.5. Meta strongly welcomes the Online Safety Act's systems-focused framework, which requires companies to use proportionate systems and processes to address the risks of harmful content and sets high standards for transparency. We also believe the systems and processes we have in place for tackling harmful content and dealing with crisis scenarios are exactly what the Online Safety Act requires platforms to have in place under the safety duties it imposes.
3.6. At this stage prior to implementation, we would caution against the introduction of any new requirements until we are better able to assess the effectiveness of the regime. The OSA already provides adequate levers in the regulatory toolkit to enable Ofcom to get the information it requires to further its objectives in ensuring online safety for UK users.
3.7. We have been supportive of the Act as we believe the framework largely avoids many of the mistakes seen in some other jurisdictions where policymakers have tried to too tightly prescribe what must be done in relation to certain specific harms or focus down on individual instances of harmful content online. Our experience is that the more prescriptive the regulatory requirements and the greater the focus on individual pieces of content, the slower and more complex the decision-making for content reviews.
Question 3a: What more should be done to combat potentially harmful social media and AI content?
3.8. We believe that regulation by itself isn’t the answer to all harms on the internet. Digital and media literacy play a crucial role in helping people understand how to use the internet safely.
3.9. We welcomed the clause in the Online Safety Act relating to media literacy, which gives Ofcom a duty to promote media literacy through educational activities and guidance (with powers for the Secretary of State to give directions on this guidance). In Meta’s experience, these activities can help to address harm online, and we look forward to seeing Ofcom’s work set out in more detail, including how services might work with Ofcom in meeting their respective duties.
3.10. Over the years, Meta has worked closely with the UK Government on its Online Media Literacy Strategy, and we will continue to work with the Government and our partners to support programmes that educate and empower internet users across the UK to manage their online safety.
Question 3b: What role do Ofcom, and the National Security Online Information Team play in preventing the spread of harmful and false content online
3.11. We have engaged with both Ofcom and Government on the steps we take when addressing harmful content, including in pertinent moments where information sharing is integral. For example, during both the UK riots and the UK General Election, were regularly engaged in industry-wide conversations convened by the National Security Online Information Team to discuss cross-platform trends. These collaborations allowed us to share information and best practices, which helped to mitigate the spread of harmful content.
3.12. We look forward to working with Ofcom’s new Advisory Committee on Misinformation and Disinformation once it has been established.
3.13. In our experience, a key part of our approach to combating misinformation is our third-party fact-checking (3PFC) program. However, sometimes these issues can be subjective and not clear cut so information sharing is important to get to an agreed approach. One example that illustrates this challenge is the case of UK riots, where it took three days for the real name of the perpetrator to be released and subsequently fact-checked. If this information had been released earlier, it could have been downranked sooner, potentially reducing its spread. Our view is that there is a space for greater voluntary collaboration between stakeholders to mitigate these issues in future.
4.1. The spread of misinformation, disinformation, and harmful content online is a multifaceted issue that requires a multifaceted response.
4.2. As detailed above, Meta has taken measures to prevent the spread of misinformation and harmful content on our platforms. We also believe that Ofcom, as the UK online safety regulator, has an important role to play in promoting accountability among platforms to ensure the systems and processes are robust and help to create a safer online environment for all users.
4.3. However, beyond a certain point individuals who knowingly spread harmful content must also take accountability for their actions online. Whilst Tech companies can and should put barriers in place to help prevent illegality occurring on the platform, illegality online should be treated in the same way as illegality offline, therefore Government and Law Enforcement has a role to play to disincentivise this behaviour.
4.4. Ultimately, preventing the spread of harmful and false content online requires a collaborative effort between government agencies, regulators, technology companies, and civil society organisations. We are committed to working together to address these complex issues and ensure that our platforms are safe and trustworthy for all users.
18 December 2024