Examination of Witness

Online Safety Bill – in a Public Bill Committee at 3:50 pm on 26th May 2022.

Alert me about debates like this

Frances Haugen gave evidence.

Photo of Roger Gale Roger Gale Conservative, North Thanet 4:36 pm, 26th May 2022

We now have Frances Haugen, a former Facebook employee. Thank you for joining us.

Photo of Alex Davies-Jones Alex Davies-Jones Shadow Minister (Digital, Culture, Media and Sport)

Q Good afternoon, Frances. Thank you for joining us.

Frances Haugen:

Thank you so much for inviting me.

Photo of Alex Davies-Jones Alex Davies-Jones Shadow Minister (Digital, Culture, Media and Sport)

No problem. Could you give us a brief overview of how, in your opinion, platforms such as Meta will be able to respond to the Bill if it is enacted in its current form?

Frances Haugen:

There are going to be some pretty strong challenges in implementing the Bill as it is currently written. I want to be really honest with you about the limitations of artificial intelligence. We call it artificial intelligence, but people who actually build these systems call it machine learning, because it is not actually intelligent. One of the major limitations in the Bill is that there are carve-outs, such as “content of democratic importance”, that computers will not be able to distinguish. That might have very serious implications. If the computers cannot differentiate between whether something is or is not hate speech, imagine a concept even more ambiguous that requires even more context, such as defining what is of democratic importance. If we have carve-outs like that, it may actually prevent the platforms from doing any content moderation, because they will never know whether a piece of content is safe or not safe.

Photo of Alex Davies-Jones Alex Davies-Jones Shadow Minister (Digital, Culture, Media and Sport)

Q You have just answered my question on AI and algorithmic intention. When I questioned Meta in Tuesday’s oral evidence session, they were unable to tell me how many human moderators they had directly working for them and how many had abided by a UK standard and code of conduct. Do you see the lack of human moderators being a problem as the Bill is enacted by platforms such as Meta?

Frances Haugen:

I think it is unacceptable that large corporations such as this do not answer very basic questions. I guarantee you that they know exactly how many moderators they have hired—they have dashboards to track these numbers. The fact that they do not disclose those numbers shows why we need to pass laws to have mandatory accountability. The role of moderators is vital, especially for things like people questioning judgment decisions. Remember, no AI system is going to be perfect, and one of the major ways people can have accountability is to be able to complain and say, “This was inaccurately judged by a computer.” We need to ensure that there is always enough staffing and that moderators can play an active role in this process.

Photo of Alex Davies-Jones Alex Davies-Jones Shadow Minister (Digital, Culture, Media and Sport)

Q One final question from me, because I know others will want to come in. How do you think platforms such as Meta—I know we have used Meta as an example, but there are others—can be incentivised, beyond the statutory duty that we are currently imposing, to publish their data to allow academics and researchers into their platforms to examine exactly what is going on? Or is this the only way?

Frances Haugen:

All industries that live in democratic societies must live within democratic processes, so I do believe that it is absolutely essential that we the public, through our democratic representatives like yourself, have mandatory transparency. The only two other paths I currently see towards getting any transparency out of Meta, because Meta has demonstrated that it does not want to give even the slightest slivers of data—for example, how many moderators there are—are via ESG, so we can threaten then with divestment by saying, “Prosocial companies are transparent with their data,” and via litigation. In the United States, sometimes we can get data out of these companies through the discovery process. If we want consistent and guaranteed access to data, we must put it in the Bill, because those two routes are probabilistic—we cannot ensure that we will get a steady, consistent flow of data, which is what we need to have these systems live within a democratic process.

Photo of Maria Miller Maria Miller Conservative, Basingstoke

Q Turning to the issue of child safety and online abuse with images involving children, what should be added to or removed from the Bill to improve how it protects children online? Have you got any thoughts on that? Some groups have described the Bill’s content as overly broad. Would you make any comments on how effective it will be in terms of online safety for children?

Frances Haugen:

I am not well versed on the exact provisions in the Bill regarding child safety. What I can say is that one of the most important things that we need to have in there is transparency around how the platforms in general keep children under the age of 13 off their systems—transparency on those processes—because we know that Facebook is doing an inadequate job. That is the single biggest lever in terms of child safety.

I have talked to researchers at places like Oxford and they talk about how, with social media, one of the critical windows is when children transition through puberty, because they are more sensitive on issues, they do not have great judgment yet and their lives are changing in really profound ways. Having mandatory transparency on what platforms are doing to keep kids off their platforms, and the ability to push for stronger interventions, is vital, because keeping kids off them until they are at least 13, if not 16, is probably the biggest single thing we can do to move the ball down the field for child safety.

Photo of Maria Miller Maria Miller Conservative, Basingstoke

Q You say that transparency is so important. Can you give us any specifics about particular areas that should be subject to transparency?

Frances Haugen:

Specifically for children or across the whole platform?

Photo of Maria Miller Maria Miller Conservative, Basingstoke

Specifically for children.

Frances Haugen:

I will give you an example. Facebook has estimated ages for every single person on the platform, because the reality is that lots of adults also lie about their ages when they join, and advertisers want to target very specific demographics—for example, if you are selling a kit for a 40th birthday, you do not want to mis-target that by 10 years. Facebook has estimated ages for everyone on the platform. It could be required to publish every year, so that we could say, “Hey, there are four kids on the platform who you currently believe, using your estimated ages, are 14 years old—based not on how old they say they are, but on your estimate that this person is 14 years old. When did they join the platform? What fraction of your 14-year-olds have been on the platform since they were 10?” That is a vital statistic.

If the platforms were required to publish that every single quarter, we could say, “Wow! You were doing really badly four years ago, and you need to get a lot better.” Those kinds of lagging metrics are a way of allowing the public to grade Facebook’s homework, instead of just trusting Facebook to do a good job.

Facebook already does analyses like this today. They already know that on Facebook Blue, for example, for some age cohorts, 20% of 11-year-olds were on the platform—and back then, not that many kids were online. Today, I would guess a much larger fraction of 11-year-olds are on Instagram. We need to have transparency into how badly they are doing their jobs.

Photo of Kim Leadbeater Kim Leadbeater Labour, Batley and Spen

Q Frances, do you think that the Bill needs to set statutory minimum standards for things such as risk assessments and codes of practice? What will a company such as Facebook do without a minimum standard to go by?

Frances Haugen:

It is vital to get into the statute minimum standards for things such as risk assessments and codes of conduct. Facebook has demonstrated time and again—the reality is that other social media platforms have too—that it does the bare minimum to avoid really egregious reputational damage. It does not ensure the level of quality needed for public safety. If you do not put that into the Bill, I worry that it will be watered down by the mountains of lobbyists that Facebook will throw at this problem.

Photo of Kim Leadbeater Kim Leadbeater Labour, Batley and Spen

Q Thank you. You alluded earlier to the fact that the Bill contains duties to protect content of democratic importance and journalistic content. What is your view on those measures and their likely effectiveness?

Frances Haugen:

I want to reiterate that AI struggles to do even really basic tasks. For example, Facebook’s own document said that it only took down 0.8% of violence-inciting content. Let us look at a much broader category, such as content of democratic importance—if you include that in the Bill, I guarantee you that the platforms will come back to you and say that they have no idea how to implement the Bill. There is no chance that AI will do a good job of identifying content of democratic importance at any point in the next 30 years.

The second question is about carve-outs for media. At a minimum, we need to greatly tighten the standards for what counts as a publication. Right now, I could get together with a friend and start a blog and, as citizen journalists, get the exact same protections as an established, thoughtful, well-staffed publication with an editorial board and other forms of accountability. Time and again, we have seen countries such as Russia use small media outlets as part of their misinformation and disinformation strategies. At a minimum, we need to really tighten that standard.

We have even seen situations where they will use very established publications, such as CNN. They will take an article that says, “Ukrainians destroyed a bunch of Russian tanks,” and intentionally have their bot networks spread that out. They will just paste the link and say, “Russia destroyed a bunch of tanks.” People briefly glance at the snippet, they see the picture of the tank, they see “CNN”, and they think, “Ah, Russia is winning.” We need to remember that even real media outlets can be abused by our enemies to manipulate the public.

Photo of Caroline Ansell Caroline Ansell Conservative, Eastbourne

Q Good afternoon, Frances. I want to ask you about anonymity and striking a balance. We have heard variously that anonymity affords some users safe engagement and actually reduces harm, while for others anonymity has been seen to fuel abuse. How do you see the balance, and how do you see the Bill striving to achieve that?

Frances Haugen:

It is important for people to understand what anonymity really is and what it would really mean to have confirmed identities. Platforms already have a huge amount of data on their users. We bleed information about ourselves on to these platforms. It is not about whether the platforms could identify people to the authorities; it is that they choose not to do that.

Secondly, if we did, say, mandate IDs, platforms would have two choices. The first would be to require IDs, so that every single user on their platform would have to have an ID that is verifiable via a computer database—you would have to show your ID and the platform would confirm it off the computer. Platforms would suddenly lose users in many countries around the world that do not have well-integrated computerised databases. The platforms will come back to you and say that they cannot lose a third or half of their users. As long as they are allowed to have users from countries that do not have those levels of sophisticated systems, users in the UK will just use VPNs—a kind of software that allows you to kind of teleport to a different place in the world—and pretend to be users from those other places. Things such as ID identification are not very effective.

Lastly, we need to remember that there is a lot of nuance in things like encryption and anonymity. As a whistleblower, I believe there is a vital need for having access to private communications, but I believe we need to view these things in context. There is a huge difference between, say, Signal, which is open source and anyone in the world can read the code for it—the US Department of Defence only endorses Signal for its employees, because it knows exactly what is being used—and something like Messenger. Messenger is very different, because we have no idea how it actually works. Facebook says, “We use this protocol,” but we cannot see the code; we have no idea. It is the same for Telegram; it is a private company with dubious connections.

If people think that they are safe and anonymous, but they are not actually anonymous, they can put themselves at a lot of risk. The secondary thing is that when we have anonymity in context with more sensitive data—for example, Instagram and Facebook act like directories for finding children—that is a very different context for having anonymity and privacy from something like Signal, where you have to know someone’s phone number in order to contact them.

These things are not cut-and-dried, black-or-white issues. I think it is difficult to have mandatory identity. I think it is really important to have privacy. We have to view them in context.

Photo of Caroline Ansell Caroline Ansell Conservative, Eastbourne

Thank you. That is very helpful.

Photo of Chris Philp Chris Philp The Parliamentary Under-Secretary of State for Digital, Culture, Media and Sport

Q Thank you for joining us and giving evidence, Frances; it is nice to see you again. We had evidence from Meta, your former employer, on Tuesday, in which its representative suggested that it engages in open and constructive co-operation with researchers. Do you think that testimony was true?

Frances Haugen:

I think that shows a commendable level of chutzpah. Researchers have been trying to get really basic datasets out of Facebook for years. When I talk about a basic dataset, it is things as simple as, “Just show us the top 10,000 links that are distributed in any given week.” When you ask for information like that in a country like the United States, no one’s privacy is violated: every one of those links will have been viewed by hundreds of thousands, if not millions of people. Facebook will not give out even basic data like that, even though hundreds if not thousands of academics have begged for this data.

The idea that they have worked in close co-operation with researchers is a farce. The only way that they are going to give us even the most basic data that we need to keep ourselves safe is if it is mandated in the Bill. We need to not wait two years after the Bill passes—and remember, it does not even say that it will happen; Ofcom might say, “Oh, maybe not.” We need to take a page from the Digital Services Act and say, “On the day that the Bill passes, we get access to data,” or, at worst, “Within three months, we are going to figure out how to do it.” It needs to be not, “Should we do it?” but “How will we do it?”

Photo of Chris Philp Chris Philp The Parliamentary Under-Secretary of State for Digital, Culture, Media and Sport

Q When I was asking questions on Tuesday, the representative of Meta made a second claim that raised my eyebrow. He claimed that, in designing its algorithms, it did not primarily seek to optimise for engagement. Do you think that was true?

Frances Haugen:

First, I left the company a year ago. Because we have no transparency with these companies, they do not have to publish their algorithms or the consequences of their algorithms, so who knows? Maybe they use astrology now to rank the content. We have no idea. All I know is that Meta definitely still uses signals—did users click on it, did they dwell on it, did they re-share it, or did they put a comment on it? There is no way it is not using those. It is very unlikely that they do not still use engagement in their ranking.

The secondary question is, do they optimise for engagement? Are they trying to maximise it? It is possible that they might interpret that and say, “No, we have multiple things we optimise for,” because that is true. They look at multiple metrics every single time they try to decide whether or not to shift things. But I think it is very likely that they are still trying to optimise for engagement, either as their top metric or as one of their top metrics.

Remember, Meta is not trying to optimise for engagement to keep you there as long as possible; it is optimising for engagement to get you and your friends to produce as much content as possible, because without content production, there can be no content consumption. So that is another thing. They might say, “No, we are optimising for content production, not engagement,” but that is one step off.

Photo of Chris Philp Chris Philp The Parliamentary Under-Secretary of State for Digital, Culture, Media and Sport

Q The Bill contains provisions that require companies to do risk assessments that cover their algorithms, and then to be transparent about those risk assessments with Ofcom. Do you think those provisions will deliver the change required in the approach that the companies take?

Frances Haugen:

I have a feeling that there is going to be a period of growing pains after the first time these risk assessments happen. I can almost entirely guarantee you that Facebook will try to give you very little. It will likely be a process of back and forth with the regulator, where you are going to have to have very specific standards for the level of transparency, because Facebook is always going to try to give you the least possible.

One of the things that I am actually quite scared about is that, in things like the Digital Services Act, penalties go up to 10% of global profits. Facebook as a company has something like 35% profit margins. One of the things I fear is that these reports may be so damning— that we have such strong opinions after we see the real, hard consequences of what they are doing—that Facebook might say, “This isn’t worth the risk. We’re just going to give you 10% of our profits.” That is one of the things I worry about: that they may just say, “Okay, now we’re 25% profitable instead of 35% profitable. We’re that ashamed.”

Photo of Chris Philp Chris Philp The Parliamentary Under-Secretary of State for Digital, Culture, Media and Sport

Q Let me offer a word of reassurance on that. In this Bill, the penalties are up to 10% of global revenue, not profit. Secondly, in relation to the provision of information to Ofcom, there is personal criminal liability for named executives, with a period of incarceration of up to two years, for the reason you mentioned.

Frances Haugen:

Oh, good. That’s wonderful.

Photo of Chris Philp Chris Philp The Parliamentary Under-Secretary of State for Digital, Culture, Media and Sport

We had a case last year where Facebook—it was actually Facebook—failed to provide some information to the CMA in a takeover case, and it paid a £50 million fine rather than provide the information, hence the provision for personal criminal liability for failing to provide information that is now in this Bill.

My final question is a simple one. From your perspective, at the moment, when online tech companies are making product design decisions, what priority do they give to safety versus profit?

Frances Haugen:

What I saw when I was at Facebook was that there was a culture that encouraged people to always have the most positive interpretation of things. If things are still the same as when I left—like I said, I do not know; I left last May—what I saw was that people routinely had to weigh little changes in growth versus changes in safety metrics, and unless they were major changes in safety metrics, they would continue to pursue growth. The only problem with a strategy like that is that those little deficits add up to very large harms over time, so we must have mandated transparency. The public have to have access to data, because unless Facebook has to add the public cost of the harm of its products, it is not going to prioritise enough those little incremental harms as they add up.

Photo of Roger Gale Roger Gale Conservative, North Thanet

Ms Haugen, thank you very much indeed for joining us today, and thank you also for the candour with which you have answered your questions. We are very grateful to you indeed.

The Committee will meet again on Tuesday 7 June at 9.25 am for the start of its line-by-line consideration of the Bill. That session will be in Committee Room 14.

Ordered, That further consideration be now adjourned. —(Steve Double.)

Adjourned till Tuesday 7 June at twenty-five minutes past Nine o’clock.

Written evidence to be reported to the House

OSB24 The Investment Association

OSB25 Jeremy Peckam

OSB26 Mid-Sized Platform Group

OSB27 Carnegie UK

OSB28 Full Fact

OSB29 Together Association

OSB30 The Christian Institute

OSB31 Clean up the Internet

OSB32 Joint Submission on Children's Amendments on the Online Safety Bill submitted by 5Rights Foundation, NSPCC and Children’s Charities’ Coalition on Internet Safety (CHIS) (and others)

OSB33 Internet Advertising Bureau UK (IAB UK)

OSB33A Annex - IAB UK Digital advertising industry commitment to tackle scam advertising online

OSB34 Victims’ Commissioner

OSB35 The British Psychological Society

OSB36 Paul Wragg

OSB37 Joint submission from Global Encryption Coalition signatories

OSB38 Internet Matters