Written supplementary follow up evidence submitted by Meta (Facebook) (OSB0224)



November 12, 2021


Damian Collins MP

Joint Committee on Draft Online Safety Bill,

House of Lords, London SW1A 0PW



Dear Damian,


Thank you again for the recent opportunity for Antigone and I to give evidence to your Committee.


This letter provides some further information about our company in response to the questions Members asked during our session, and by email. It is worth noting that several of the questions asked here are based on misunderstandings of the recently leaked documents; we have tried to clarify these in this letter.


The Committee has a copy of our written submission regarding the Bill, and we would of course be happy to discuss any particular points in it or our specific recommendations to help you understand our perspective on the draft legislation. Our teams are already in touch, and as discussed separately with the clerks will follow up with Committee members directly on some of the other statements that were made.


Kind regards,


Chris Yiu

Director of Public Policy for Northern Europe







Questions from Committee Members


  1. In response to Darren Jones – How do we define ‘tiers’ relating to degree of oversight and speed of response to concerns about safety, and what action do they lead to? i.e. Tier 1 ‘war'; Tier 2 ‘operational rooms’; tier 3? Which tier is the UK in?


The teams at our company dedicated to keeping our community safe have always prioritized enforcing our policies and keeping the impact of violations low, at all times and in all countries where we offer our services. For many years, we’ve focused on building and scaling our policy enforcement operation: growing and training global teams of reviewers, refining policies and processes, and developing technology to help us find and take action on violations at scale.


As part of this work, our teams have developed an industry-leading process of reviewing and prioritizing which countries have the highest risk of offline harm and violence every six months. We make these determinations in line with the UN Guiding Principles on Business and Human Rights and following a review of societal harms, how much Facebook’s products impact these harms and critical events on the ground. We do not make the determinations regarding individual countries public, and we keep all designations under review. However, I wanted to set out some further details about how this system works and the reasons why we believe it is crucial to our work of protecting our users.


Facebook supports peoples’ right to express themselves freely, regardless of where they are in the world. Freedom of expression is a foundational human right, and enables many other rights. But we know that platforms for free expression, information, and opinion can also be abused to spread hate and misinformation - a challenge made even worse in places where there is a heightened risk of conflict and violence.


This requires developing both short-term solutions that we can implement when crises arise, but also having a long-term strategy to help keep people safe. Since 2018, we’ve had dedicated teams spanning product, engineering, policy, research and operations to better understand and address the way social media is used in countries experiencing conflict.


Many of these individuals have experience working on conflict, human rights, and humanitarian issues, as well as addressing areas like misinformation, hate speech and polarization; have lived or worked in the countries we’ve identified as highest risk; and speak relevant languages. They are part of the over 40,000 people we have working on safety and security, including global content review teams in over 20 sites around the world reviewing content in over 70 languages.


In the last two years, we've hired more people with language, country and topic expertise. For example, we’ve increased the number of team members with work experience in Myanmar and Ethiopia to include former humanitarian aid workers, crisis responders, and policy specialists. And we’ve hired more people who can review content in Amharic, Oromo, Tigrinya, Somali, and Burmese. Adding more language expertise has been a key focus area for us. This year alone, we've hired content reviewers in 12 new languages including Haitian Creole, Kirundi, Tswana, and Kinyarwanda.


As mentioned above, based on this work our teams have developed an industry-leading process of reviewing and prioritizing which countries have the highest risk of offline harm and violence every six months. We make these determinations in line with the UN Guiding Principles on Business and Human Rights and following a review of societal harms, how much Facebook’s products impact these harms and critical events on the ground.


In those countries that we deem most at risk, we develop long-term plans to respond to potential offline harms before they happen. We take a multi-pronged approach to prevent online content from contributing to offline harm, keep our community safe, and to protect freedom of expression. Our approach includes:

       Understanding and engaging with local contexts and communities

       Developing and evaluating policies to prohibit harmful content

       Improving our technology and enforcement to keep our community safe.


We have made public several recent examples of our longer-term work including our preparations for elections in Myanmar, Ethiopia, and India.


In a crisis, we will determine what kind of support and teams we need to dedicate to a particular country or language, and for how long we need to keep them in place. This might include deploying our Integrity Product Operations Centers (IPOCs) model to monitor and respond to threats in real time; seeking to ensure our integrity systems and resources are robust and ready where there may be ongoing risk of political unrest; or building temporary product levers ahead of a protest or a culturally sensitive event - all while ensuring that we have teams ready to support unplanned events, such responding to the coup in Myanmar.


We know that we face a number of challenges with this work and it is a complex and often adversarial space - there is no one size fits all solution. Many of these offline issues have existed for decades or longer, and media services have a long history of being abused by those seeking to assert or maintain power or incite violence.


That said, we know our work to keep our global community safe will never be finished and it requires ongoing vigilance and investments. That’s what we’ve done for many years and we will continue doing it going forward.


  1. In response to Suzanne Webb – Can we have an organisation chart showing how the safety and integrity teams report into the FB Board and to Mark Zuckerberg specifically?
  1. In response to Suzanne Webb and Darren Jones, how does the Product Safety team report to the Audit and Risk Oversight Comm of FB Board?


As set out above, Facebook employs over 40,000 people to work on safety and security, and these teams are spread across almost every part of our company (including Engineering, Product Management, Operations, Business, Policy, and Legal functions). The Integrity team, led by Guy Rosen (VP, Integrity), sits within the Central Product Services team, led by Javier Olivan (VP, Central Product Services) which spans efforts across our family of apps. The Safety Policy team, led by Antigone Davis (Director, Global Head of Safety) sits within the Content Policy team, led by Monika Bickert (VP, Content Policy), which is itself within the global policy team led by Joel Kaplan (VP, Global Policy). 


All of our functions ultimately report to the Chief Executive Officer. However, a diagram of reporting lines cannot and does not reflect the cross-functional collaboration, shared goals, and information exchanges which are hallmarks of our corporate culture, nor would it reflect leadership priorities in the manner the Committee appears to infer.


The Audit and Risk Oversight Committee (“AROC”) of the Facebook, Inc (now Meta Platforms) Board has several responsibilities, including oversight of the company’s risks related to social responsibility. As set out in its Charter,[1] AROC reviews the company’s assessment of the major ways in which our  services can be used to facilitate harm or undermine public safety or the public interest, as well as the steps the company has taken to monitor, mitigate, and prevent such abuse.   AROC conducts these reviews together with management at least annually, and generally are briefed by Guy Rosen (VP, Integrity) and Monika Bickert (VP, Content Policy) on Community Safety and Security issues twice a year.


In response to Suzanne Webb and Darren Jones: How would expressions of concern from employees reach a level where action could be considered and taken if necessary?

  1. In response to Debbie Abrahams – what is the guidance given to employees about how to raise safety concerns about what they are seeing or working with?


We have a Whistleblower and Complaint Policy, and a channel by which employees can raise any concern. This Policy and channel are highlighted in our Code of Conduct,[2] internal training courses, and internal communications tools. People can raise concerns with their (or another manager),  with the Legal or Compliance teams, an Employee or Contingent worker Relations Partner, or Internal Audit. They may also raise a concern anonymously through our Speak Up website and phone hotline. 



  1. In response to Damian Collins - when did you stop recommending civic groups? ‘Before 6 January’, but what was the exact date?


Long before the US election period began last year, we expected that the 2020 election would be one of the most contentious in history — and that was before we even knew it would be conducted in the midst of a pandemic. We began our planning for the 2020 election two years in advance. We built our strategy to run all the way through Inauguration Day in 2021, knowing that there was a high likelihood that the election results would be contested. So we planned specifically for that scenario.


We activated multiple different measures throughout the course of 2019 and 2020, and kept many of these in place until Inauguration Day or beyond. You can read about more than a dozen of the interventions that we made during that period in our recent blog post, called “Our Comprehensive Approach to Protecting the US 2020 Elections Through Inauguration Day”, which includes the measures to stop recommending civic groups. In the United States, measures to restrict recommending civic groups were introduced well in advance of the November 2020 election. Similar measures were rolled out globally in March 2021.


  1. In response to Baroness Kidron – Why are you able to remove copyright and IP infringements faster than harmful content?


We are often asked whether we prioritise copyright violations over harmful content like bullying or hate speech. We don’t. Both are against our rules, but in some cases our automated systems can be more effective in catching and removing copyright violations.


Our tool for helping rights holders to enforce their copyrights on our platforms is called Rights Manager. Rights Manager works by copyright holders directly uploading exact copies of their works, which we can then scan for to find and remove unauthorised copies. We use similar technology to find and remove known imagery of child exploitation, or terrorist material, where our proactive removal rates have for many years been extremely high.


For many of the other types of harmful content on the platform, and which we discussed with the Committee, it is much harder to distinguish between legitimate speech and violating posts. Hate speech or bullying almost always needs a human reviewer to consider the context; for example if it is being used to attack, we’d remove it, but if someone is condemning it, we’d allow it. That is why the published figures in our transparency report show that we continue to rely more on people flagging harmful posts to us in order to remove them.


However, we have made considerable progress in many of these areas—for example, our proactive removal rate for bullying and harassment has increased from 12% to 54% in just the last 12 months. We firmly believe that it should not be the responsibility of the person experiencing abuse to have to report it, which is why we are committed to investing heavily in advanced technology to continue to find more harmful posts proactively, and we are on track to spend $5bn dollars on safety and security in 2021 alone.


  1. In response to Lord Knight - 
    1. How do the community standards and terms and conditions apply to paid-for adverts?
    2. Who is responsible for ensuring that they meet those rules (and are safe)?
    3. How are those rules enforced in respect of advertising?


The first line of our ads policies on what is prohibited in ads states that “Adverts must not violate our Community Standards”. Everyone who uses Facebook or Instagram must comply with our Terms of Service, our Community Standards (called Guidelines on Instagram). Our Community Standards are a baseline set of rules that guide what is and isn’t allowed on our services. These apply to all types of content on Facebook and Instagram, including ads.


Our Advertising Policies also place additional requirements on advertisers in order to help further protect people from poor experiences — especially because ads may be delivered to people in their News Feed from Pages or accounts they don’t follow.


All ads go through our ad review system before they go live. Our ad review system relies primarily on automated technology to apply our Advertising Policies to the millions of ads that are run across our apps. However, we do use human reviewers to improve and train our automated systems, and in some cases, to manually review some ads. Based on the review the ad may go live or be rejected.


Our terms and policies require that advertisers follow all applicable laws and regulations,[3] and we partner with many consumer protection bodies and regulators globally to help us better detect and remove content that violates our policies or local law once they report it to us.


The UK has one of the longest-established advertising regulatory regimes, considered by many globally to be best-in-class at protecting consumers. Online advertisers, including those advertising on Facebook and Instagram, are not only subject to our advertising rules and our community standards but are also subject to the strict Code(s) of Advertising Practice administered by the Advertising Standards Authority and every advertiser is responsible for adhering to that Code. The advertiser remains responsible for ensuring the content, placement and targeting of their ads complies with the UK Advertising Code(s) and our ad policies.


  1. In response to Damian Collins –
    1. Your transparency figures only show (a) how much hate you identified and (b) how much of the hate you identified was removed with AI. It does not explain how much hate is actually on the platform. Do you assess how much is on the platform to establish how effective your tools are?


Yes, and this estimate is provided in our transparency reports, which actually show three types of figures for hate speech: (a) the number of posts removed, (b) how much of that we removed proactively before it was reported to us, and (c) our estimate of the prevalence of hate speech on the platform. In Q3 2021 we estimate that 3 out of every 10,000 views of content on the platform were of content that violated our hate speech standards. This figure has reduced by more than 50% since we began reporting it in November 2020.


Prevalence is the metric that we have developed to guide our work in reducing harmful content on Facebook and Instagram. Reducing the prevalence of harmful posts is the exact metric that we give to our safety, product, and engineering teams internally. Prevalence is focused on estimating views of harmful content, not just counting how much harmful content exists, because views are how harmful content actually causes harm—one harmful video that is viewed a million times causes far more harm than 1,000 harmful videos that are only viewed once or not at all.


We have set out the details of how we estimate and measure prevalence on our transparency website, here.



    1. Of all the hate speech you take down, what proportion is taken down by AI? (in all cases, both UK and global data):


In Q3 of 2021, the proportion of hate speech proactively found and actioned by us globally was 97%. This is a significant increase from 52% three years earlier.


Please note that ‘found and actioned by us’ does not mean taken down by AI alone. Hate speech is nuanced and changeable and despite advances in our technologies, very often human review is required to confirm that a piece of content violates our policies. Please see our answer to question 8d below.


We do not currently provide country-specific figures in our Community Standards enforcement report, because of the complexities in defining and estimating this. For example, if a piece of hate speech is posted by a person in the EU within a Facebook Group that is administered by people in the US, but most of the views of that content are from users in the UK, and an Australian user reports it leading to it being removed, it is not clear which country such a takedown should be attributed to. However we are committed to continuing to provide more granular data about our work to enforce our Community Standards, and making public additional information like that you have requested is a focus of this ongoing work.



    1. Of all of the hate speech on the platform, what proportion is taken down by AI?
    2. Of all of the hate speech on the platform, what proportion is taken down?


Our policies against hate speech apply equally across our platforms, and every item we find in violation of our policies is taken down. However, no system to prevent or find and remove violating content is or can be perfect. As described above, our transparency efforts are focused on estimating the prevalence of harmful content on the platform, which focuses on understanding views of harmful content, not how much content is on the platform as a whole.


This is our focus because we want to determine the extent to which any harmful content affected people on Facebook or Instagram. A piece of violating content could be published once but seen 1,000 times, one million times or not at all. Measuring views of violating content rather than the amount of violating content published better reflects the impact on the community.


A small prevalence number can still correspond to a large amount of impact on our services, due to the large number of overall views of content on our services, so our teams and technology are focused on reducing the distribution of content that is likely to be violating and likely to receive views, and reviewing and removing such content where it does violate. A substantial amount of content on our platform is never viewed at all, and retroactively searching through unviewed posts would take away resources from finding and preventing harmful views of content that is more likely to be viewed.



    1. You claim in your transparency reports that 97% of all hate is taken down with AI. You also claim that only 0.05% of all views are for hateful content. However, the Facebook Files published by WSJ report that only 3-5% of all hate on Facebook is taken down with AI. How do you explain this discrepancy?


This is an example of a misunderstanding of the information in the leaked documents. The claim that we only remove 3-5% of hate speech cites internal research which refers only to hate speech which is found by AI and then automatically removed with no further input. This does not account for the variety of ways that we use technology to combat the extremely nuanced issue of hate speech.


We have a very high threshold for when we automatically delete content (meaning without human review) for hate speech so that we don’t make a mistake in removing acceptable content. Importantly, we also use AI to proactively detect suspected hate speech and then reduce its distribution and visibility while routing it to our teams of content reviewers to investigate further. These multifaceted approaches to tackling harmful content help manage the complexities inherent in managing online speech, in line with the cautious and proportionate approach suggested in the Draft Bill.



Are the figures in the transparency report checked externally or by other teams within Facebook (e.g. Facebook AI research)?


The figures in our transparency report are carefully developed and checked by a wide range of teams across Facebook. We periodically measure our actions with a separate, independent system that measures content actions. On a regular basis, we check these various independent metrics, which are intended to identify large errors in our accounting.


For example we conduct a range of random spot checks to verify the accuracy of our measurement systems in near real time. This includes checking various outcomes that happen later in our system to double-check upstream outcomes. For example, we confirm that content that is appealed is also logged as content that has been actioned, as content must be actioned in order to be appealed. Many of these checks are intended to identify large errors such as content that is appealed but was never removed.


However, we fully appreciate that additional, external validation is an important component of driving public trust in our data. To ensure that our methods are transparent and based on sound principles, we seek out analysis and input from subject matter experts on areas such as whether the metrics we provide are informative.


In order to ensure that our approach to measuring content enforcement was meaningful and accurate, we worked with the Data Transparency Advisory Group (DTAG), an external group of international academic experts in measurement, statistics, criminology and governance. In May 2019, they provided their independent, public assessment of whether the metrics we share in the Community Standards enforcement report provide accurate and meaningful measures of how we enforce our policies, as well as the challenges we face in this work, and what we do to address them.


Overall, they found our metrics to be reasonable ways of measuring violations and in line with best practices. They also provided a number of recommendations for how we can continue to be more transparent about our work, which we discussed in detail and continue to explore.


We know the credibility of our systems should be earned, not assumed. That is why, in 2020 we took a further step towards validating that our metrics are measured and reported correctly. We contracted with an external auditor, Ernst & Young, to conduct an independent audit of our metrics and we look forward to them publishing their assessments once completed in Spring 2022.


17 November 2021






[1] https://investor.fb.com/leadership-and-governance/audit-committee-charter/default.aspx

[2] Slide 8, https://s21.q4cdn.com/399680738/files/doc_downloads/governance_documents/2021/06/FB-Code-of-Conduct.pdf

[3] Terms of Service at https://www.facebook.com/legal/self_service_ads_terms.  and within our Ad policies at 12.2: https://www.facebook.com/policies/ads