Dear Lord Puttnam,
I wanted to follow up from my recent appearance before the committee.
I was pleased when I appeared at the hearing to be able to update the committee that significant progress had been made during the course of your hearings on the issue of releasing more data to the academic community for research into elections.
But this announcement represents a genuine step change. Researchers have never before been able to understand population-level URL sharing and exposure dynamics -- in other words how information and disinformation is propagated across Facebook
The committee had a number of questions about how the decisions had been taken to release the data in the form we did. The most authoritative sources for this information is the following article which explains this:
The challenge we faced in releasing this data is that by sharing databases, even anonymised, it is theoretically possible for sophisticated actors to identify individuals. Given the sensitive nature of the data involved - often relating to their engagement with political content - it should be clear why this must be kept entirely safe.
Social Science One - based at Harvard University and our partners in the release of this data - do make clear they disagree with our interpretation of our privacy responsibilities under the law to our users. But they also make clear that they
“would not be on the hook if our legal interpretation did not win the day in court or with regulators.“
As regards providing more background to the committee on the legal advice relating to our engagement with regulators on this matter I hope you will understand that our legal assessments are subject to legal privilege and remain confidential and our discussions with regulators are conducted on a confidential basis.
However - as I undertook to the committee - we have reflected on what information we can provide to the committee that would reassure you of our determination throughout this process to provide as comprehensive a data set as possible.
We engage extensively with regulators and data protection authorities around the world on numerous issues, and are bound by the very high standards set by the GDPR, California's CCPA and of course the Federal Trade Commission consent order recently imposed on FB
It should be clear to the committee that our caution in these areas is grounded in a good understanding of regulators’ expectations. As I said to the committee it is of course entirely within the prerogative of regulators - and would be welcome - to provide even more comfort in this area to researchers and companies.
If the committee is looking for signs of our good faith intent to provide as much data as we can to the academic community then I would ask the committee to consider the fact that we invested over $11m dollars, and significant engineering resource in addition to this, to contribute - working with universities and others - to the development of the new Differential privacy approach.
This process operates by adding enough random noise to data such that there are mathematical guarantees (a closed form proof) of individuals’ protection from re-identification.
As such, the results of analysis are the same whether or not a given individual is included in the data, and therefore preventing this kind of identification.
The purpose of this is precisely to provide as much relevant information to researchers as is possible without compromising the privacy of our users.
As I told the committee the database we have released enables the academic community - despite these necessary safeguards - to gain:
Developing tools like differential privacy is not a one-time gain. It is available to the world as a tool for sharing such datasets more easily, so this process is one we anticipate continuing, as regulators, companies and researchers develop their understanding of these issues in a post GDPR and CCPA world.
I am confident that the committee will agree that this demonstrates our wholehearted commitment to transparency in this area.
A few other questions were raised by the committee where I undertook to follow up;
The committee also asked for more information on how the Oversight board’s decisions would impact on our content policies. Since the hearing, on the 12th May, we announced the establishment of the board and more details on it’s powers and role can now be found here: https://about.fb.com/news/2020/05/welcoming-the-oversight-board/
We have made clear that
‘For our part, Facebook will implement the board’s decisions unless doing so could violate the law, and will respond constructively and in good faith to policy guidance put forth by the board.’
The Oversight Board site is here: https://www.oversightboard.com/news/announcing-the-first-members-of-the-oversight-board/
As the board makes clear:
‘The Board will take final and binding decisions on whether specific content should be allowed or removed from Facebook and Instagram’
‘The Board will review whether content is consistent with Facebook and Instagram’s policies and values, as well as a commitment to upholding freedom of expression within the framework of international norms of human rights. We will make decisions based on these principles, and the impact on users and society, without regard to Facebook’s economic, political or reputational interests. Facebook must implement our decisions, unless implementation could violate the law.’
Flagging content for fact check
I undertook to provide more information on the ways in which content is flagged to fact checkers including what share of content flagged to fact checkers was flagged manually by technology or by users.
Content is surfaced to our independent third-party fact-checking partners through a combination of technology and human review. On the technology side, we have misinformation classifiers that use machine learning to synthesize thousands of data points to make predictions about whether content would likely be rated false by a fact-checker. Such data points include things like user reports, disbelief comments, and whether the person posting has a history of sharing false rated content.
Content that is prioritized for fact-checkers to review is populated in an internal tool we have developed for our fact-checking partners to review. On the human review side, our fact-checking partners can also proactively surface public content on Facebook or Instagram that they can then choose to review and rate and submit into the Fact Checking tool.
Once one of our fact-checking partner submits a rating through the fact-checking tool, we also use technology to surface content that is similar to something that they previously fact-checked, thereby helping them rate more content quickly.
As should be clear from the above the interplay between technology and human reporting is not hard and fast. For example if a claim was first manually reported and found false, but numerous similar claims were then also flagged as false because technology found that they were similar claims it is unclear how many should count as manually flagged and how many manually.
In order to provide a sense of the interplay between the two, so as to try to provide as satisfactory an answer as possible to the question, I can provide the committee with the latest figures on actions we are taking on misinformation relating to the COVID-19 epidemic.
In the UK we work with Full Fact, FactCheckNI, and as of last month, Reuters UK. We also know that for this content where we have applied a warning label, 95% of people who saw it did not click through to view it. In addition to misinformation that we’ve actioned on, since March 1st, we’ve removed more than 2.5 million pieces of content for the sale of masks, hand sanitizers, surface disinfecting wipes and Covid-19 test kits.
The committee also asked for more clarity on the definition of political candidate or elected official in relation to fact checking.
For the purposes of fact-checking, politicians on Facebook are profiles and Pages from politicians, government officials, political candidates, and political parties or campaigns--many of which have blue-badge verification.
The political candidates list would include full time elected office holders or candidates for full time elected office - given their deeper knowledge of the context we ask our independent fact checkers to make the judgement as to whether part time elected office holders or candidates for those office holders should be covered by this specific exemption.
As I told the committee the difference between our treatment of politicians applies in one specific area: We continue to remove and action their content against our community standards (the content rules) on the same basis as all other users of our platform.
However, when it comes to fact-checking we don’t send content or ads from politicians and political parties to our third party fact-checking partners. We believe that by limiting speech or removing content, we’d leave people less informed about what their elected officials are saying and leave politicians less accountable for their words.
As I said earlier, this doesn’t mean that politicians can say whatever they want on Facebook. They can’t spread misinformation about where, when or how to vote, incite violence or post anything else that would violate our Community Standards.
Additionally, when a politician shares content that has previously been debunked as part of our third-party fact-checking program that content will be demoted and have the fact checkers debunking attached to it. But in general we believe political speech should be heard and we don’t feel it is right for private companies like us to fact-check or judge the veracity of what politicians and political parties say - when they are being transparent about who they are.
Unicef Myanmar Content
The Chair raised with us content appearing in Myanmar that Unicef had told him they had flagged. At the time I told the committee that we had removed significant amounts of the content and that we were continuing to investigate. We recently published our regular monthly report into take-downs we have undertaken as part of our global efforts to stop what we describe as ‘coordinated inauthentic behaviour’. This included networks of accounts that our investigation linked to members of the Myanmar police force. A report on this investigation was published and can be found here: https://about.fb.com/wp-content/uploads/2020/05/April-2020-CIB-Report.pdf
UK Public Policy Manager