Mark Zuckerberg has been publishing a series of articles ddressing the most important issues facing Facebook. This is the second in the series. Here are a few selected extracts
The team responsible for setting these policies is global -- based in more than 10 offices across six countries to reflect the different cultural norms of our community. Many of them have devoted their careers to issues like child safety, hate
speech, and terrorism, including as human rights lawyers or criminal prosecutors.
Our policy process involves regularly getting input from outside experts and organizations to ensure we understand the different perspectives that exist on free expression and safety, as well as the impacts of our policies on different
communities globally. Every few weeks, the team runs a meeting to discuss potential changes to our policies based on new research or data. For each change the team gets outside input -- and we've also invited academics and journalists to join
this meeting to understand this process. Starting today, we will also publish minutes of these meetings to increase transparency and accountability.
The team responsible for enforcing these policies is made up of around 30,000 people, including content reviewers who speak almost every language widely used in the world. We have offices in many time zones to ensure we can respond to reports
quickly. We invest heavily in training and support for every person and team. In total, they review more than two million pieces of content every day. We issue a transparency report with a more detailed breakdown of the content we take down.
For most of our history, the content review process has been very reactive and manual -- with people reporting content they have found problematic, and then our team reviewing that content. This approach has enabled us to remove a lot of harmful
content, but it has major limits in that we can't remove harmful content before people see it, or that people do not report.
Accuracy is also an important issue. Our reviewers work hard to enforce our policies, but many of the judgements require nuance and exceptions. For example, our Community Standards prohibit most nudity, but we make an exception for imagery that
is historically significant. We don't allow the sale of regulated goods like firearms, but it can be hard to distinguish those from images of paintball or toy guns. As you get into hate speech and bullying, linguistic nuances get even harder --
like understanding when someone is condemning a racial slur as opposed to using it to attack others. On top of these issues, while computers are consistent at highly repetitive tasks, people are not always as consistent in their judgements.
The vast majority of mistakes we make are due to errors enforcing the nuances of our policies rather than disagreements about what those policies should actually be. Today, depending on the type of content, our review teams make the wrong call in
more than 1 out of every 10 cases.
Proactively Identifying Harmful Content
The single most important improvement in enforcing our policies is using artificial intelligence to proactively report potentially problematic content to our team of reviewers, and in some cases to take action on the content automatically as
This approach helps us identify and remove a much larger percent of the harmful content -- and we can often remove it faster, before anyone even sees it rather than waiting until it has been reported.
Moving from reactive to proactive handling of content at scale has only started to become possible recently because of advances in artificial intelligence -- and because of the multi-billion dollar annual investments we can now fund. To be clear,
the state of the art in AI is still not sufficient to handle these challenges on its own. So we use computers for what they're good at -- making basic judgements on large amounts of content quickly -- and we rely on people for making more complex
and nuanced judgements that require deeper expertise.
In training our AI systems, we've generally prioritized proactively detecting content related to the most real world harm. For example, we prioritized removing terrorist content -- and now 99% of the terrorist content we remove is flagged by our
systems before anyone on our services reports it to us. We currently have a team of more than 200 people working on counter-terrorism specifically.
Some categories of harmful content are easier for AI to identify, and in others it takes more time to train our systems. For example, visual problems, like identifying nudity, are often easier than nuanced linguistic challenges, like hate speech.
Our systems already proactively identify 96% of the nudity we take down, up from just close to zero a few years ago. We are also making progress on hate speech, now with 52% identified proactively. This work will require further advances in
technology as well as hiring more language experts to get to the levels we need.
In the past year, we have prioritized identifying people and content related to spreading hate in countries with crises like Myanmar. We were too slow to get started here, but in the third quarter of 2018, we proactively identified about 63% of
the hate speech we removed in Myanmar, up from just 13% in the last quarter of 2017. This is the result of investments we've made in both technology and people. By the end of this year, we will have at least 100 Burmese language experts reviewing
Discouraging Borderline Content
One of the biggest issues social networks face is that, when left unchecked, people will engage disproportionately with more sensationalist and provocative content. This is not a new phenomenon. It is widespread on cable news today and has been a
staple of tabloids for more than a century. At scale it can undermine the quality of public discourse and lead to polarization. In our case, it can also degrade the quality of our services.
ur research suggests that no matter where we draw the lines for what is allowed, as a piece of content gets close to that line, people will engage with it more on average -- even when they tell us afterwards they don't like the content.
This is a basic incentive problem that we can address by penalizing borderline content so it gets less distribution and engagement. By making the distribution curve look like the graph below where distribution declines as content gets more
sensational, people are disincentivized from creating provocative content that is as close to the line as possible.
The category we're most focused on is click-bait and misinformation. People consistently tell us these types of content make our services worse -- even though they engage with them. As I mentioned above, the most effective way to stop the spread
of misinformation is to remove the fake accounts that generate it. The next most effective strategy is reducing its distribution and virality.
Interestingly, our research has found that this natural pattern of borderline content getting more engagement applies not only to news but to almost every category of content. For example, photos close to the line of nudity, like with revealing
clothing or sexually suggestive positions, got more engagement on average before we changed the distribution curve to discourage this. The same goes for posts that don't come within our definition of hate speech but are still offensive.
This pattern may apply to the groups people join and pages they follow as well. This is especially important to address because while social networks in general expose people to more diverse views, and while groups in general encourage inclusion
and acceptance, divisive groups and pages can still fuel polarization. To manage this, we need to apply these distribution changes not only to feed ranking but to all of our recommendation systems for things you should join.
One common reaction is that rather than reducing distribution, we should simply move the line defining what is acceptable. In some cases this is worth considering, but it's important to remember that won't address the underlying incentive
problem, which is often the bigger issue. This engagement pattern seems to exist no matter where we draw the lines, so we need to change this incentive and not just remove content.
Building an Appeals Process
Any system that operates at scale will make errors, so how we handle those errors is important. This matters both for ensuring we're not mistakenly stifling people's voices or failing to keep people safe, and also for building a sense of
legitimacy in the way we handle enforcement and community governance.
We began rolling out our content appeals process this year. We started by allowing you to appeal decisions that resulted in your content being taken down. Next we're working to expand this so you can appeal any decision on a report you filed as
well. We're also working to provide more transparency into how policies were either violated or not.
The closed-door trilogue efforts to finalise the EU Copyright Directive continue. The Presidency of the Council, currently held by Austria, has now circulated among the EU member state governments a new proposal for a compromise between the
differing drafts currently on the table for the controversial Articles 11 and 13.
Under this latest proposal, both upload filters and the link tax would be here to stay -- with some changes for the better, and others for the worse.
Let's recall: In its final position, the European Parliament had tried its utmost to avoid specifically mentioning upload filters, in order to avoid the massive public criticism of that measure. The text they ended up with, however, was even
worse: It would make online platforms inescapably liable for any and all copyright infringement by their users, no matter what action they take. Not even the strictest upload filter in the world could possibly hope to catch 100% of unlicensed
This is what prompted YouTube's latest lobbying efforts in favour of upload filters and against the EP's proposal of inescapable liability. Many have mistaken this as lobbying against Article 13 as a whole -- it is not. In Monday's Financial
Times, YouTube spelled out that they would be quite happy with a law that forces everyone else to build (or, presumably, license from them) what they already have in place: Upload filters like Content ID.
In this latest draft, the Council Presidency sides with YouTube, going back to rather explicitly prescribing upload filters. The Council proposes two alternative options on how to phrase that requirement, but they match in effect:
Platforms are liable for all copyright infringements committed by their users, EXCEPT if they
cooperate with rightholders
by implementing effective and proportionate steps to prevent works they've been informed about from ever going online determining which steps those are must take into account suitable and effective technologies
Under this text, wherever upload filters are possible, they must be implemented: All your uploads will require prior approval by error-prone copyright bots .
On the good side, the Council Presidency seems open to adopting the Parliament's exception for platforms run by small and micro businesses . It also takes on board the EP's better-worded exception for open source code sharing platforms like
On the bad side, Council rejects Parliament's efforts for a stronger complaint mechanism requiring reviews by humans and an independent conflict resolution body. Instead it takes on board the EP's insistence that licenses taken out by a platform
don't even have to necessarily cover uses of these works by the users of that platform. So, for example, even if YouTube takes out a license to show a movie trailer, that license could still prevent you as an individual YouTuber from using that
trailer in your own uploads.
Article 11 Link tax
On the link tax, the Council is mostly sticking to its position: It wants the requirement to license even short snippets of news articles to last for one year after an article's publication, rather than five, as the Parliament proposed.
In a positive development, the Council Presidency adopts the EP's clarification that at least the facts included in news articles as such should not be protected. So a journalist would be allowed to report on what they read in another news
article, in their own words.
Council fails to clearly exclude hyperlinks -- even those that aren't accompanied by snippets from the article. It's not uncommon for the URLs of news articles themselves to include the article's headline. While the Council wants to exclude
insubstantial parts of articles from requiring a license, it's not certain that headlines count as insubstantial. (The Council's clause allowing acts of hyperlinking when they do not constitute communication to the public would not apply to such
cases, since reproducing the headline would in fact constitute such a communication to the public.)
The Council continues to want the right to only apply to EU-based news sources -- which could in effect mean fewer links and listings in search engines, social networks and aggregators for European sites, putting them at a global disadvantage.
However, it also proposes spelling out that news sites may give out free licenses if they so choose -- contrary to the Parliament, which stated that listing an article in a search engine should not be considered sufficient payment for reproducing
snippets from it.
The French President, Emmanuel Macron has announced a plan to effectively embed French state censors with Facebook to learn more about how to better censor the platform. He announced a six-month partnership with Facebook aimed at figuring out how
the European country should police hate speech on the social network.
As part of the cooperation both sides plan to meet regularly between now and May, when the European election is due to be held. They will focus on how the French government and Facebook can work together to censor content deemed 'harmful'.
It's a pilot program of a more structured engagement with the French government so that both sides can better understand the other's challenges in dealing with the issue of hate speech online. The program will allow a team of regulators, chosen
by the Elysee, to familiarize [itself] with the tools and processes set up by Facebook to fight against hate speech. The working group will not be based in one location but will travel to different Facebook facilities around the world, with
likely visits to Dublin and California. The purpose of this program is to enable regulators to better understand Facebook's tools and policies to combat hate speech and, for Facebook, to better understand the needs of regulators.
In order to prepare Rainbow 6 Siege for expansion into China, Ubisoft announced that it will be making some global censor cuts to the game's visuals to remove gore and references to sex and gambling.
In a blog post, Ubisoft explained:
A Single, Global Version
We want to explain why these changes are coming to the global version of the game, as opposed to branching and maintaining two parallel builds. We want to streamline our production time to increase efficiency
By maintaining a single build, we are able to reduce the duplication of work on the development side. This will allow us to be more agile as a development team, and address issues more quickly.
Ubisoft provided examples of their censorship:
Icons featuring knives become fists
Icons featuring skulls are replaced
Skulls in artwork are fleshed out into faces
Images of slot machines are removed
Blood spatters are removed from a Chinese landscape painting
Microsoft has just inflicted a new 'code of conduct' that prohibits customers communicating nudity, bestiality, pornography, offensive language, graphic violence and criminal activity, whilst allowing Microsoft to steal the money in your
If users are found to have shared, or be in possession of, these types of content, Microsoft can suspend or ban the particular user and remove funds or balance on the associated account.
It also appears that Microsoft reserves the right to view user content to investigate violations to these terms. This means it has access to your message history and shared files (including on OneDrive, another Microsoft property) if it thinks
you've been sharing prohibited material.
Unsurprisingly, few users are happy that Microsoft is willing to delve through their personal data.
Microsoft has not made it clear if it will automatically detect and censor prohibited content or if it will reply on a reporting system. On top of that, Microsoft hasn't clearly defined its vague terms. Nobody is clear on what the limit on
offensive language is.
Facebook has files a patent that describes a method of using the devices of Facebook app users to identify various wireless signals from the devices of other users.
It explains how Facebook could use those signals to measure exactly how close the two devices are to one another and for how long, and analyses that data to infer whether it is likely that the two users have met. The patent also suggests the app
could record how often devices are close to one another, the duration and time of meetings, and can even use its gyroscope and accelerometer to analyse movement patterns, for example whether the two users may be going for a jog, smooching or
catching a bus together.
Facebook's algorithm would use this data to analyse how likely it is that the two users have met, even if they're not friends on Facebook and have no other connections to one another. This might be based on the pattern of inferred meetings, such
as whether the two devices are close to one another for an hour every Thursday, and an algorithm would determine whether the two users meeting was sufficiently significant to recommend them to each other and/or friends of friends.
I don't suppose that Facebook can claim this patent though as police and the security services have no doubt been using this technique for years.
The Satanic Temple in Salem, Massachusetts is suing Netflix and producers Warner Brothers over a statue of the goat-headed deity Baphomet that appears in the TV series Chilling Adventures of Sabrina .
The temple is claiming that Netflix and Warners are violating the copyright and trademark of the temple's own Baphomet statue, which it built several years ago.
Historically, the androgynous deity has been depicted with a goat's head on a female body, but The Satanic Temple created this statue with Baphomet having a male chest an idea that was picked up by Netflix.
The Temple is seeking damages of at least $50 million for copyright infringement, trademark violation and injury to business reputation. In the Sabrina storyline, the use of the statue as the central focal point of the school associated with
evil, cannibalism and possibly murder is injurious to TST's business, the Temple says in its suit.
Speaking at the Web Summit conference in Lisbon, Tim Berners-Lee, inventor of the World Wide Web, has launched a campaign to persuade governments, companies and individuals to sign a Contract for the Web with a set of principles intended to
defend a free and open internet.
Contract for the Web CORE PRINCIPLES
The web was designed to bring people together and make knowledge freely available. Everyone has a role to play to ensure the web serves humanity. By committing to the following principles, governments, companies and citizens around the world can
help protect the open web as a public good and a basic right for everyone.
Ensure everyone can connect to the internet so that anyone, no matter who they are or where they live, can participate actively online.
Keep all of the internet available, all of the time so that no one is denied their right to full internet access.
Respect people's fundamental right to privacy so everyone can use the internet freely, safely and without fear.
Make the internet affordable and accessible to everyone so that no one is excluded from using and shaping the web.
Respect consumers' privacy and personal data so people are in control of their lives online.
Develop technologies that support the best in humanity and challenge the worst so the web really is a public good that puts people first.
Be creators and collaborators on the web so the web has rich and relevant content for everyone.
Build strong communities that respect civil discourse and human dignity so that everyone feels safe and welcome online.
Fight for the web so the web remains open and a global public resource for people everywhere, now and in the future.
We commit to uphold these principles and to engage in a deliberative process to build a full "Contract for the Web", which will set out the roles and responsibilities of governments, companies and citizens. The challenges facing the web
today are daunting and affect us in all our lives, not just when we are online. But if we work together and each of us takes responsibility for our actions, we can protect a web that truly is for everyone.
Gab, the social media site that prides itself as being uncensored, has been forced offline by its service providers after it became clear that the alleged Pittsburgh shooter Robert Bowers had a history of anti-semitic postings on the site.
Formed in August 2016 after Twitter began cracking down on hate speech on its social network, Gab describes itself as a free speech website and nothing more. But the platform has proved popular among the alt-right and far right, including the man
accused of opening fire on a synagogue in Pennsylvania on Saturday, killing 11.
In the hours following the attack, when the suspect's postings were discovered on the site, Gab's corporate partners abandoned it one by one. PayPal and Stripe, two of the company's payment providers, dropped it, arguing that it breached policies
around hate speech.
Cloud-hosting company Joyent also withdrew service on Sunday, giving Gab 24 hours notice of its suspension, as did GoDaddy, the site's domain registrar, which provides the Gab.com address. Both companies said the site had breached their terms of
Gab responded in a statement:
We have been systematically no-platformed by App Stores, multiple hosting providers, and several payment processors, the company said in a statement posted to its site. We have been smeared by the mainstream media for defending free expression
and individual liberty for all people and for working with law enforcement to ensure that justice is served for the horrible atrocity committed in Pittsburgh.
Gab is back online following censorship in the wake of the anti-Semitic shooting at a Pittsburgh synagogue. The social network had been banned by its hosting provider Joyent and domain registrar GoDaddy, and blacklisted by other services such as
PayPal , Stripe and Shopify.
Now, Gab has come back online and has found a new hosting provider in Epik. According to a blog post published on November 3rd, Epik CEO Robert Monster spoke out against the idea of digital censorship and decided to provide hosting privileges to
Gab because he looks forward to partnering with a young, and once brash, CEO who is courageously doing something that looks useful.
Prior to Google's bosses being called in to answer for its policy to silence conservative voices, it has filed a statement to court saying that even if it does discriminate on the basis of political viewpoints. It said:
Not only would it be wrong to compel a private company to guarantee free speech in the way that government censorship is forbidden by the Constitution, but it would also have disastrous practical consequences.
Google argued that the First Amendment appropriately limits the government's ability to censor speech, but applying those limitations to private online platforms would undermine important content regulation. If they are bound by the same First
Amendment rules that apply to the government, YouTube and other service providers would lose much of their ability to protect their users against offensive or objectionable content -- including pornography, hate speech, personal attacks, and