I would like to start by welcoming Paul Waugh MP, who is a member of the Culture, Media and Sport Committee and is taking part in this evidence session as a guest. This is the second session in the Committee’s inquiry into social media, misinformation and harmful algorithms. We have previously heard from those impacted by the spread of misinformation, particularly over the summer. We are now hearing from the platform companies. Meta, X and TikTok will join us in the afternoon, but this morning we are pleased to give a warm welcome to Amanda Storey, who is the managing director, trust and safety at Google EMEA. On behalf of the Committee, I thank you again for joining us. Google is one of the largest and most important companies in this space—some may say the most important company in this space. The Committee felt it was very important to hear from you because there is a huge level of public interest in the question of the spread of misinformation. Indeed, I have been surprised by the level of public interest in my constituency, across the country and in Parliament. That brings me to my first question, but first: is it Ms Storey or Amanda?

Either is fine.

Okay, Amanda. First, do you agree that tech companies have a responsibility to be transparent and accountable before Parliament?

Thank you for having me here today to speak to the Committee. The issues that you are looking at are core to what my team does every day—keeping people safe online, looking at AI responsibility and making sure we are compliant with regulations such as the OSA. As you said, billions of people put their trust in Google every day and search for information on our platforms. We see it as our responsibility to make sure that the information that they find is relevant, high-quality and authoritative information and that we are not unintentionally providing misleading, hateful or harmful information to them. We also see it as our responsibility to make sure we provide context on that information. On things such as, for example, “About this result” or “About this image”, we provide people with information so that they can understand the quality and provenance of the information they are consuming across our platforms. Importantly, we are not a social network, so we are optimising for quality and not for engagement. We can potentially talk today a little bit about how some of those obligations and processes are evolving in a world of generative AI, but I will also speak briefly about the Online Safety Act, which I know you are also focused on. We support the goals of that. We think that online safety is absolutely critical, and something we have invested heavily in over the years. We think that the Online Safety Act can raise the bar across the industry in a really important way. It is one of the most complex and comprehensive pieces of safety regulation that my team has seen anywhere in the world. We are working hard to implement the requirements. We are compliant in some areas and we have enhancements to make in others. We are collaborating very closely with Ofcom to do that.

Thank you very much, Amanda. You have touched on a number of points that I know Committee members are interested in, and we will come on to them, but I want to press you on the question regarding whether big tech companies have a responsibility to be transparent and accountable before Parliament. My constituents want to know whether a company like Google feels a responsibility to be accountable before their representatives in Parliament.

We absolutely believe in transparency and accountability in this space. To give you a couple of examples of how we think about transparency, you can think about content safety operations as being like a funnel. You have bad activity or bad actors trying to get in at the top of that funnel, and then you have multiple lines of defence that you put in place to protect users and protect the quality of the information we are serving. Transparency at each of those levels is really important. At the very top level, you can think about the safety by design of our products themselves. Things like SafeSearch, which we turn on for children to blur potentially bad or pornographic content, and making sure that we verify actors who are trying to serve content across our platforms. That prevention layer, including our policy set, is really important, and we are completely transparent about that policy set. We have help centres that explain those policies and explain how to adhere to those policies, as well as our ranking approach, which is part of safety by design, and which, again, we are completely transparent about. We have a 170-page search quality rater manual that you can find online and take a look through. That is the prevention layer, and there are some elements of transparency within that. We then have the detection layer, which is how we actually detect content that violates the policies I just mentioned. That is a combination of what we call classifiers—pieces of machine learning that run across our systems and detect content that violates those policies. We also have flaggers; users flag things to us. That is part of our accountability. We provide transparency about take-downs that we do on the basis of those classifiers and on the basis of flags that we receive. Finally, we have the response layer at the bottom of the funnel. If you have not managed to prevent or detect the bad content, you can then have people reporting that content to you—trusted flaggers or Governments, for example. We have channels to provide to trusted authorities around the world to flag things to us. We have transparency and accountability at all layers of that funnel and absolutely believe that is an important part of doing this work well.

Thank you for that answer, which I think touches on many points. On the relationship that you see towards Governments and the representatives of the public, you talked about trusted channels for Governments to identify issues and so on. Do you see that you have a level of accountability towards Parliament, towards the representatives of the people?

Yes so—

That is fantastic. We will move on because we have a lot of interest in this.

It is a great question. My team focuses on this every day. When you look at Search, there are three levels here. First is the ranking itself, which we have refined over 25 years to make sure we are elevating high-quality content and are not unintentionally elevating potentially misleading content. Direct navigation queries are challenging. We are organising the world’s information and making it universally accessible and useful, but it is the internet and there is a lot of content out there. Secondly, we have safety interventions that we put in place. On Search there are certain queries that we see that are particularly sensitive user journeys, where someone is obviously in crisis and we want to make sure we are providing high-quality information and giving them a recovery journey. Those safety interventions are critical. We have what we call OneBoxes that appear at the top of the search results for particularly sensitive queries, and we make sure that we are directing people to the highest quality information, often in partnership with local entities. The final piece of it is the user context that I mentioned earlier. This is making sure we are providing clear information to users about the content they are consuming—where has that image come from, how old is that result or site that they are seeing? That is an important part of users being able to assess the quality of the information they are consuming, as are all the media literacy investments we do. Think about Be Internet Legends for children, for instance, which is trying to teach children about the quality of information that they might be consuming online.

A great question. We have our 16,000 search quality raters around the world who every day are looking at the quality of the search results and giving us feedback on those. They are working against the 170-page manual that I mentioned. This set of constant evaluations is critical to us looking at how well we are doing against our own policies and our own standards of high-quality information. That evaluation is critical. It is also important to respond when we see that we have got something wrong. We therefore have our strategic command team which is dealing with escalations and incidents in the real world. We are constantly learning from that and feeding that back into the way in which we do ranking, the way we run those evaluations, experiments that we run to improve Search, as well as into the safety interventions and user context tools that I mentioned.

We are dealing with enormous scale across our platforms. If you look at the number of ads we took down, in the most recent report it was 5.5 billion in 2023. We demonetised 2.1 billion publisher pages. You cannot do that all with humans. You have to do that with a blend of what we call classifiers, these pieces of machine learning, and humans. The way that works is the classifiers run across all of our products and assess whether anything is violating our policies. Where they can make a decision they do. Across some of our products you see 95% of the take-downs that we make first detected by a classifier. Where the classifier cannot make the decision, it passes the content to our human review queues. They then assess those pieces of content against our policies and make a decision. Those decisions feed back into improving the classifier, so you have that important constant feedback loop. I mentioned that 95% were first found by a classifier. What about the other 5%? That is the user flags that we get. It is quite a noisy signal; sometimes users will flag things they do not like, rather than things that actually violate policies. But it is complemented by trusted flaggers—academic and civil society experts—as well as Government flags from around the world. That blend of humans and machines to do this work at the scale that we operate at is absolutely critical.

Can I ask for some clarification? It is very helpful understanding how the process works. The speed of the process is also very important, because misinformation and the virality of misinformation is driven by speed. Do you measure the speed of the process? Do you feel that it is successful in reducing the speed of the spread of misinformation?

The classifiers are a really key part of being quick. They can run across our products much faster than humans can, so that is part of how we deal with the speed. The virality you are mentioning is more pertinent to social media companies than it is to the libraries of content or the access to the internet that most of our products provide. But in a crisis situation, if there is a real world incident, we have crisis protocols. We have teams that are dedicated to making sure the intelligence is passing between our product teams and to our trust and safety teams really, really quickly, because in those moments of crisis, that speed really matters.

How quickly would you expect to be able to react?

As I mentioned, 95% of the take-downs we do happen via detection from a classifier, so that is often incredibly quick. Some content is up on our products for many years and then our policies evolve, so sometimes you will see the speed of take-down look slower, but it is that balance between assessing content that has been up for a while and content that is fresh, and making sure that we are making accurate decisions against our policies consistently across all of those.

The way Search works is pretty transparent. The Search rater guidelines are out there; they can be looked at. We are running hundreds of experiments all the time to continue to improve Search. Giving people a high-quality experience when they come to search is a business imperative for us. That is what users expect of us, and that is what we will continue to invest in. Making sure that the ecosystem that people arrive in is safe is also critical to our business. Our publishers expect that high-quality ads will serve on their sites. Advertisers expect that their ads will show up in a brand-safe environment. Creating that safe, high-quality ecosystem is absolutely critical, and we will continue to invest in it.

A lot of this goes back to holding companies accountable on the mission they have set out. Ours is around authoritative information, and we will continue to optimise for that. I think you are also asking about whether we or any company might end up being biased towards one person’s point of view or one country’s point of view. What is really important in this space is that we are building for the world. We are taking into account perspectives from around the world of users and of experts. My team consults with civil society and academia on a regular basis through our Google safety engineering centres, and last year we consulted with over 8,000 people in 150 events. All of that global feedback is part of what ensures we are creating an experience that is appropriate for all, not biased to one person’s point of view or one country’s point of view.

I do not think it is my position to advise on that, but I think that the business imperative for us is the piece that keeps us optimising for high-quality information and building services that are useful to the world.

Before I pass on to Emily, I think Adam has been making a very important point. I suppose the theory would say that competition would help keep a company honest, if you like, and give consumers choice. Where there is a lack of competition in that space, I think you are saying that it is then up to Google’s internal processes and your overall mission.

I am not a competition expert at all, so I do not want to comment unduly, although I am sure the team would be happy to follow up in writing. When you look at the online ecosystem, there is a range of different services where you can find information. It is not just Search; there are dedicated vertical search environments, if you think about travel or shopping. We would be very happy to follow that up and offer more.

Hi, Amanda. I want to go into Google’s new AI Overviews function, which sees Google moving from a search engine to actually generating content using AI. I think you have described yourself as an AI-first company. How do you ensure the accuracy of that content, and that that content meets not only your own content standards but local laws around harmful information and incitement to violence?

Thank you for raising the point about local laws. I have talked a bit less about that, but it is worth taking a moment on that point, and then I can absolutely speak to AI Overviews. We have teams of lawyers who make sure that we are adhering to the local laws in the countries in which we operate, and quickly taking down illegal content that we have detected or that has been flagged to us. That is complemented by the policy set that I talked about before, which is ensuring that we are going further than local laws, and taking down things that could be potentially misleading, hateful or harmful, to create that safe ecosystem that I talked about. AI Overviews is a really interesting new feature within Search; it is like past features that we have launched—Images search, for example. The reason it is particularly interesting is that we think it is going to really help users, instead of getting 10 blue links, to get something that actually takes information from across multiple sites, multiple modalities—text, image and video—and potentially multiple languages, and condenses that into an answer to their question, with citations so that they can dive in and explore further. We think that that is very helpful as an evolution for users. Those citations are really important to make sure that people are able to jump in and dig into the content even further. We also have, as we do across all our products, safety interventions to make sure that the AI Overviews results we are creating meet the safety standards that that we set. That is informed by 25 years of ranking experience in Search and 20-odd years of trust and safety experience in trying to make sure that all the content we have on our platforms is safe.

I do not think that that quite answers the question I was asking. How, within AI Overviews, are you ensuring that the information is not harmful, not hateful, not inciting violence and not promoting harm?

Just like our policies elsewhere across our platforms, we have policies for AI Overviews. We do rigorous testing of all those features before they roll out—think about adversarial red teaming and performance testing, where we set benchmarks. We are testing in all sorts of languages. We are testing all sorts of sensitive queries that someone might put in. That is in order to make sure that the AI Overview that is returned meets those quality standards. That is the pre-launch part of it. Then, there is the post-launch, responsive approach that we take. If we do see someone flagging an AI Overview as being harmful or inappropriate, we will absolutely dig into that feedback and look at what went wrong. Why did it violate our policies? Why did that AI Overview trigger on an area where, potentially, there was not high-quality information to be surfacing? That feedback loop is really important for us to continue to refine those evals and get safety right.

Is misinformation one of those things that you evaluated against? For example, we saw in the news that the AI Overview asserted that the average IQ in Pakistan was 80, Kenya was 75.2, and Sierra Leone was 45.07. That is clearly racism and has absolutely been debunked by scientific research. That is clear misinformation. How do you ensure that clear misinformation is stripped from your AI Overview? Or is misinformation considered less important than something that is potentially illegal?

Absolutely not. We are very concerned about real-world harm and information quality. The example you give is from a Wired article that people may have seen. It was not the sort of answer that we would have wanted to give. It absolutely violated our policies, and we have taken action on that particular space.

I followed that, and the action was not within hours. It was not even within days. It took some time for that to come down.

That is totally fair. This was the very early stages of AI Overviews, and this is a new feature. It is something that we labelled as experimental and that we are continuing to refine over time. But you are right: it was not as quick a reaction as we would have expected, but we have taken action. It did violate our policies, and what we learned from that experience is that there are certain areas where, if a user puts in a query, there just is not enough high-quality content on the internet to create an AI Overview. So we have AI Overviews triggering on certain queries and not triggering on others. If we see what we call a data void, which is where there is not enough high-quality information, we will not trigger an AI Overview. So we have implemented a more systemic fix to exactly that problem space.

Obviously, the Committee is very interested in the spread of both misinformation and incitement to violence and hatred, in the context of what we have seen increasingly in the UK, particularly during the summer. What changes have you made as a result of these examples to make sure that they do not happen again? That is what I am not clear about. I get that you took action on this one, but what will stop it happening in future? And what gives the public confidence to look at these AI Overviews or to allow their children to use them to write a paper for school?

It is critical for us to be raising authoritative information, not serving information to people that is misleading or harmful. We have strict policies to make sure that we meet that benchmark. We are constantly learning and taking feedback from users and other authorities and making sure that we are responsive and learning from incidents like this. That constant learning process, and judging us on the outcomes that you see on our products, is critical here. Again, we are not social media, and some of the problems you are articulating—viral spread of misinformation, for example—are more pertinent in those contexts.

I get what you are saying. I get that you are trying to make that distinction between a search company and social media, but the reality is that your AI scraping to create your AI Overview is scraping social media. You cannot really disaggregate that world any more; you do not live in two different spaces.

It is a really fair question. Again, I would go back to how 25 years of search ranking has taught us a lot about how you distinguish high-quality information from lower-quality information. The way of assessing relevance quality is really nuanced, based on all those search experiments that we have done and all the years of experience in search ranking. We are very much using that to make sure that the sources that we use to create an AI Overview are high-quality sources. In some cases, where something is a really fast-moving event, there is a lot of discussion on social media. We also have treatments like breaking news to make sure that we put high-quality news information at the top of the search results, not necessarily triggering an AI Overview, and using authoritative sources in any AI Overview that is created.

Thank you, Emily. Allison is going to move on to a very relevant area, but perhaps you could write to us, Amanda—I take the point about virality and social media—to give us an idea of how quickly you would expect to take down something like racist IQ results. You talked about us holding you to account according to your mission and processes, so can you write to us with how quickly you would expect to be able to take down things like that?

Sure.

Allison, I know you have something very relevant.

It is very relevant. We have been talking a lot about assessing and moderating output, and Emily rightly commented on the data that you scrape to help train your AI tools. I am interested in how you prevent that output from happening in the first place, by assessing your input data. How do you assess the credibility of the sources you use to train your AI tools, and how much human oversight is there in evaluating that process?

It is a really important question. I will admit that I am not an expert in the training of models, but I will tell you how we think about the training of models as related to application safety and, ultimately, content safety. You have to look at it in terms of those three layers. The first is the foundational model itself. We have things like our frontier safety framework and research papers on things like responsible capability scaling, which talk about how you assess a frontier model and make sure you are staying ahead of capabilities that might be emerging from that model. As you say, a lot of that is also about the data that we use to train those models, and ensuring that is filtered, safe and authoritative information. Again, the ranking experience we have is really important there. But it is not just the foundational models. You have to think about the application layer and the content layer as well. At the application layer, once you have the frontier model, you then tune it and train it for a particular use case—so, a particular product application. At that stage, it is important to run evaluations, to run adversarial red teaming, to set benchmarks and to understand the performance of that model for its particular application in different languages and different cultural contexts around the world. That is something that the trust and safety team is responsible for doing. The content that the model actually creates is critical here. Once the model is out there and being used, how do you make sure people understand that that content was created by an AI? We have invested a lot in our content provenance standard and tools—things like SynthID and the C2PA provenance standard, which is a coalition approach—to make sure people can understand whether the content they are using is AI generated. Those three layers are quite helpful. I take the point that the authoritativeness and quality of the training data is key here, and I am sure that the team will be happy to follow up and share more. Again, I am not an expert on that particular piece of it.

Google had an AI ethics team with Joy Buolamwini and Meredith Whittaker, and you had the AI audit tool. Is that used?

I am afraid I do not know the details of that specific one, but I am happy to follow up.

Okay—it is an old one. I might come back to the point, which you touched on, about what information users can use to assess the validity of the output, but I am going to go back a little, to search tools. It was only in October 2024 that you optimised your search for images and links. Images and videos are quite difficult for automated tools to assess to make sure that they are safe and harm-free. I have sort of answered my own question, but what challenges do images and videos pose to AI moderation, compared with text, and what changes have you made to account for this?

It is a great question. You are right that classifiers tend to be tuned for a text, audio, image or video environment, so we use a blend of classifiers to deal with any potential real-world harm issue or policy area. The image classifiers are tuned on things like Google Images search, which we have had for many years, and our video classifiers are tuned on things like YouTube videos, which, again, we have been running for many years. So we do have very good classifiers in those spaces, and you see, in the quality of the outputs, that those are as good as the classifiers running on text or audio.

I was wondering whether you were going to mention using your AI tools to generate images and videos. You have been talking about having embedded watermarks and metadata so that your classifiers can assess these. That might help with the speed, which you have been referring to, in crisis situations, although I think you are probably not there. I wonder whether you thought about extending that watermark labelling so that users can assess whether what they are seeing is AI generated—a sort of anti-trust mark, if you like.

Yes, absolutely. We have what is called SynthID, which is exactly that. It is a watermark that a user can see to assess that something is AI generated. And then C2PA is an industry coalition that is trying to come up with a provenance standard that is applied consistently across the web, so that people can consume that information. I hope that we get to a place where those provenance standards are really widely adopted, whether it is our own or the C2PA one.

Hi, Amanda. Through DSIT, the Government has its own national security online information team. It has been rebranded, but essentially it is attempting to counter misinformation online. Can you tell us how Google engages with that team? Are you proactive? Are you reactive? How frequently are you in contact with them?

We do engage with that team—more on some products than others—but we are not dependent on that team. We have our own policies and our own technical teams who look for signals of state-backed activity. Our threat analysis group is tracking 270 state actors across 50 countries on a real-time basis. That then feeds signals to all our products to take action against state-backed co-ordinated influence operations. When we find examples of that, we are transparent about it, and in crisis situations, we share information directly with the unit that you are describing.

Moving on to digital advertising, Google is obviously a major player in the online advertising space. Adalytics, an American firm, only this month published research showing that Google’s ad networks had placed adverts on a sickening child abuse image website called IBB—you will be aware of this. The companies whose adverts appeared there were obviously appalled. They said that they sometimes had no idea where their products and brands were being advertised when they sent their content to Google. Do you have any real answers on how you can cope with that as a phenomenon? The worry is that this is not just a one-off—something that you reacted quickly to and countered—but that there is something systemic about the lack of clarity and transparency in the way your digital networks work.

I am familiar with this case. Obviously, it is abhorrent to see any child sexual abuse material online, and it is certainly not acceptable to see that content monetising in any way. It is a challenge with user-generated content sites that you have to constantly assess the content that is being generated on those sites. In that particular case, we had already taken action against some of the sites that were identified, and the monetisation was in the order of $2 to $3—so, not acceptable, but very tiny amounts. In terms of child sexual abuse imagery and material generally, we have a really strong record in this space. We have invested heavily to make sure that we are able to identify this content really quickly and accurately and to take action. We have what is called our child safety API, which we offer to the industry and is used by many players, large and small, to detect CSAM and then prioritise it for review and action. We provide all of those signals to NCMEC, the National Centre for Missing and Exploited Children, and our tip reports to it are known as being some of the highest quality in the industry. We provided 3 million of those tip reports in the first half of last year. Our record on CSAM is excellent. This was one case where we got it wrong, and we have taken action on that. On your broader question about trust and transparency in the online ads ecosystem, I talked earlier about this triumvirate of users, publishers and advertisers who we are optimising for. It is obviously critical that we get that right; that is a business imperative for us. Publishers have to be able to trust that the ads that are serving on their sites are high-quality ads that operate within our policies. Advertisers have to be able to trust that their ads are not going to show up next to that kind of content, and we give them controls to make sure they can tailor by category or even by URL if they want to really restrict where their ads are showing up. That is all part of our transparency and accountability to the advertising ecosystem, to make sure people can have confidence in the safety of that.

A lot of advertisers still think it is too opaque and that there is not enough transparency. Are you taking any fresh steps to reassure them?

We would be happy to share more details on recent changes that we have made. The key pieces for us are, first, the controls that I mentioned for advertisers, which give them that granular ability to shift the dial on the environments in which their ads will show up; and, second, the Ads Safety report that we publish every year, which talks about the actions that we have taken at the content layer and the actor layer. For example, in the last report that we put out, we had taken action on 5.5 billion ads and 2.1 billion publisher pages that violated our policies, so we are very active in this space, to ensure that the ecosystem is safe and trustworthy.

You have mentioned that it is very damaging to the advertisers, but ultimately it is very damaging to Google, isn’t it?

Absolutely.

Even if it is just two or three dollars on a sickening child abuse website, that is appalling.

Absolutely.

Is it not the case that Google, as it stands, cannot guarantee that it does not profit from such websites, from misinformation or from hate speech? You simply cannot say that, can you?

We have really rigorous policies, and we have really rigorous processes to detect this kind of content and take action, or to detect bad actors and take action at the actor level. We take very seriously doing that work well. That is what my team does every day. We are trying to stick to those policies at speed, scale and high quality, and we continually learn from where we have made errors to make sure the system is improving.

But at the moment, you cannot guarantee that Google does not profit from all that hate?

We have been audited by multiple third parties, who have not found significant evidence in this space. We will continue to be transparent and open to those sorts of assessments. Ultimately, it comes down to publishers, advertisers and users trusting us and using our services. That is the real test for us.

I appreciate the answers you are giving. This is a complex area but it is important, and we will have a couple of follow-up questions to Paul’s points. Paul emphasised Google’s role in internet advertising. Google’s stock market capitalisation is $2.2 trillion—a huge amount—and it has profits of $74 billion. 87% of your revenue is from advertising, and you have 40% of worldwide digital advertising, which is clearly significant. It is also important that, as I understand it, you have 90% of the market share of the sell side of advertising because you own key infrastructure such as Google Ad Manager, Google Ad Exchange and Google AdSense. Advertising has an absolutely critical role in driving revenues for the platform companies—the infrastructure holders such as yourself—and a critical role as a driver of misinformation, in which the Committee is particularly interested. So, given your role in orchestrating and owning critical parts of the advertising infrastructure, it is inevitable that you will be profiting—I am not saying intentionally—from misinformation. Researchers at Stanford and Carnegie Mellon found that 75% of disinformation websites are monetised by advertising. Is it not the case that you will inevitably have a significant share of that, in some way?

I want to draw an important distinction between co-ordinated influence operations—disinformation—which is state-backed activity, and misinformation and information quality more generally. I have talked a little about some of our technical teams that are tracking state-backed activity and taking action. On the information quality side, it really comes back to our policies. We are trying to avoid anything that could create real-world harm, or that could be misleading or hateful. We are constantly evolving policies based on feedback from experts around the world, and we are running our classifiers and our human review queues to make sure that we are adhering to those policies with high quality, consistency and speed. We are doing everything we can to make sure we are creating a safe, trustworthy and high-quality information ecosystem, which is critical to our business.

I appreciate that. You have gone through a number of different ways in which Google does that, and you have set out the different processes and checks. This Committee is focused on misinformation. You draw an important distinction, but you have 40% of global digital advertising, and the evidence shows that some misinformation, or the drive behind misinformation, is caused by advertising revenue. Advertising revenue is the incentive for communicating misinformation, and you must accept that Google plays a significant part in that.

The numbers you state are absolutely correct. I would go back to the processes that my team runs, the policies we set and the way we enforce. That is how we take our responsibility in this space, to manifest that we want to ensure a safe ecosystem.

Before we move on, I want to focus on this particular example. The Committee has seen an unpublished report by a non-profit organisation called Check My Ads, which presents what it says is evidence that Google helped to monetise Channel3Now—a site that published misinformation about the Southport attacks, particularly and specifically a wrong name of the attacker and information that the attacker was a Muslim migrant, which was also wrong. Check My Ads has evidence that the site claimed to be monetised by two major companies, Ezoic and Google, at the time that it published that misinformation. I do not know whether you have been contacted by Check My Ads. You accept on the balance of probabilities that there will be cases like that, and you speak to your own responsibilities and processes to address them, but is that really good enough, given what we have seen and do not want to see again?

I completely agree that monetising any form of low-quality information, particularly information associated with an atrocious real-world attack, is absolutely not acceptable. I have not seen that report yet. I am very happy to take a look once it comes out and respond in writing, if that is helpful. But I very much agree that that would violate our policies. It is something that we would look into to understand what had gone wrong. These fast-moving, real-world situations are very challenging. There is viral spread of information on social media, and we have to deal with the echo across the sites we operate with.

I appreciate that, and we look forward to your response. Just to clarify, Channel3Now is not social media. It is a website. It was part of the drive of misinformation through social media, but it is a website that has been monetised through advertising.

It is quite interesting that you are saying there is a difference between the illegal content that my colleague Paul raised and misinformation sites. This is an interesting use of language, but it clearly comes down to the quality of the systems within Google. When it comes to misinformation that leads to violence, which is exactly the case the Chair raises, I still do not understand how, through your ad processes, you ensure that you are not monetising and encouraging content creators to knowingly create harmful misinformation that incites people to commit violence against others or themselves. Your ad model seems to promote that kind of activity. I am trying to understand how you are disincentivising the creation of that material on websites.

Our publisher policies are how we deal with this, and then the enforcement of those policies. The policies prohibit content that is hateful and that incites violence, exactly as you describe. We can then take action at the content level, at the page level and at the actor level. We employ a strike system if a publisher repeatedly violates our policies. We can take action at the whole site level, and we do take those actions. Our most recent report says that we took down 2.1 billion publisher pages from being able to monetise, so cutting off their ability to make money from this sort of content that violates our policies, and 395,000 publishers were de-monetised at the publisher level. We are absolutely taking action.

Could you give an example of where you have taken action against something we would know about, in the UK, to de-monetise a site?

I could follow up with a specific example of a UK publisher, but I do not have an example top of mind. We are taking these actions at scale, every single day, all around the world.

I would appreciate that follow-up, because I fail to see the action you are talking about. Google controls between 40% and 80% of the market share of the buy side. That is putting together Google ads and YouTube ads, which are similar. You own YouTube, and they work together. Are you aware of the recent report from the Center for Countering Digital Hate? It looked at the monetisation of anorexia and eating disorder material, with companies like Ralph Lauren and L’Oréal advertising to young people on these sites that are clearly about self-harm. How are you taking action to stop that?

That space is absolutely concerning. My heart goes out to anyone who is experiencing that sort of situation. We try to make sure that in those very sensitive user journeys, where someone is looking for that kind of information, we elevate what are called OneBoxes. Those are particular treatments within search where we elevate information from, for example, the Samaritans or Shout text—local, UK-based entities to which someone can reach out to get support and help towards a recovery journey. That is particularly critical in the sorts of areas you describe.

Right. How much money has Google made off these ads promoted by eating disorder content? One in four items of material suggested by YouTube breaches your own policies, in terms of harmful eating content, and these adverts are from quite high-profile, big brand names. How much money do you reckon you have made from that?

I am afraid that I cannot answer that question. I do not know, and I am not a YouTube expert. We have a dedicated YouTube trust and safety team. We would have to follow up and provide some more information. I go back to the fact that, in these instances of things like eating disorder journeys, we are really focused on making sure that people have access to support and high-quality information. We cannot do anything about a navigational query where someone types in the URL of a site they want to go to, but we can make sure those sites adhere to our publisher policies, and we can make sure that the search page is flooded with information to help people to recover.

To come back to the Chair’s point, how fast is fast? Only 22% of that material was removed within two weeks of being reported. Of that material, 74% had not been removed after two weeks. Is that the kind of time period we should expect? When you say “swiftly,” do you mean a month, a day or an hour?

The time between a user reporting something and the takedown is a relevant fact, but it also depends on our policies. Some of those user reports will be about things that do not violate our policies, so we will not take action. When someone reports something that is policy-violative, we would seek to take action as quickly as possible. We are using those signals—user reports, expert reports, Government reports—across all our products to take action quickly, complementing what we do with our classifiers and human reviewers.

I think we need to move on now, but we would appreciate your writing to us with more detail on the timing. I have the impression that you are saying that, where it does not violate your policies, you may try to show counter-misinformation or signpost help, such as to the Samaritans, rather than taking down the misinformation. I think we would like to understand that. Amanda Storey indicated assent.

Thank you for joining us today; it has been very interesting. I am also interested in the algorithms that have potentially led youngsters and people to a space where they are on that journey. Harm occurs, and I think that has been widely recognised. It is social media addiction, online bullying and mis- and disinformation, such as what we saw in the Southport riots. Harm occurs. What we are really trying to find out and pin down is who then has responsibility for that harm and how we rectify it. I would welcome what you have to say on that.

Harm in the real world is what my team focuses on every day. That is exactly how we are designing our policies and testing our classifiers to make sure we are mitigating those real-world instances of harm. You asked about algorithms, and I think that is an important piece of this. You are not just dealing with harmful content through takedowns, which is obviously one tool; the other mechanism is the ranking. To spend a minute on how we do search-ranking, it starts by taking into account the meaning of the user’s question. Sometimes things will be mistyped, so we have algorithms and natural language processing running across that to make sure we really understand the meaning of the user’s query. We then think about relevance and what is going to be the most relevant result for that user, based on what people have clicked on in the past and an assessment of the quality of the underlying information. When we then look at quality, it is about expertise, authoritativeness and trustworthiness, so there are multiple factors that go into assessing the quality of the site. You can think about that quality in different contexts as well. Freshness is going to matter much more in a news context, for example, than if you are looking for a dictionary definition or for factual information about a particular area that might be older. You then look at the usability of the site. Is it going to be a good user experience when someone jumps on to it? Is it a high-quality site that loads properly? It is also about the context of the user and where they are based. Those are the factors that determine how we rank, and that is critical to making sure we are elevating authoritative, high-quality information not lower quality information, or potentially misinformation.

But what do you then think about the ads that people can buy, and the ads that have been able to stay up for a number of weeks, which do not violate your policies but are potentially harmful to those particularly vulnerable individuals? That is a danger, and it causes harm. How do we rectify that harm?

I want to pull apart the policies we have for publishers, which include the content of the page and whether the page, based on that content, can monetise or not. That is one lever. There is also the content of the ad itself, and we do not allow certain types of ads to serve, and again there is a policy set that dictates that. Those are the two levers that we are using here, and both are really important for your question.

Do you think that you are putting out a potentially unsafe product? With the algorithms and the checking that the ads will not be promoting mis- and disinformation, do you think they need to be proven safety-wise before they are unleashed? You speak a lot about the experiments that are going on, and you say, “We are trialling it all the time.” But there are living consequences, and people’s lives have been ruined by mis- and disinformation through websites and ads that have linked them there. How can we help to stop this?

We absolutely feel an accountability to our users, publishers and advertisers to create and maintain that safe ecosystem, and to learn from experiences where we get it wrong. Judging us for the outcomes and for that continuous learning process is what I think is important here.

Do you think the digital advertising market is sufficiently regulated and transparent to help us hold you to account?

I know that Ofcom is looking at the digital ads market as one of its potential codes under the OSA, and I think that is a really good forum to discuss exactly that point. We have engaged very closely with Ofcom on the legal and child safety codes that are already out, and we will continue to do so.

On that point, you have talked about the checks you conduct before serving ads on a website or an app. When it comes to banking—a totally different sector—there is a huge emphasis on avoiding money laundering and on knowing your customer. We received evidence calling for basic “know your customer” checks to be required. Do you agree with that?

In certain circumstances, that, as one of the preventive techniques, is really valuable. In the case of financial services ads, for example, we do pre-certify our advertisers against, for example, the FCA to make sure they are valid, trustworthy financial advertisers.

Is that a legal requirement?

That is something that we implemented ourselves; we didn’t wait for regulation. In fact, I gave evidence in 2021 to the Treasury Select Committee on exactly that point. We do implement verification for our advertisers in certain circumstances, and we have found that to be a very important part of the overall funnel, which I described earlier, of prevention, detection and response.

There is an “Intermediary and Platform Principles” preliminary agreement. Do you think that implementing that is enough to create a digital advertising system in which bad actors cannot profit from the spread of misinformation? That is what we have been talking about. Are you familiar with it?

I am afraid I am not, so I don’t want to comment on that. I want to make sure you get a good response so I am very happy to follow up on it, but I am not familiar with it.

Well, it is a global “Intermediary and Platform Principles” pilot, which the UK’s Advertising Standards Authority launched—so it is a pilot rather than a programme as such. It aims to explore the UK’s online advertising system and bring greater transparency. Some companies have agreed to it in principle—actually, some have participated, including Google; I am not sure that is the same thing as agreeing. Perhaps you could let us know your views on it. The online harms Act gives a legal requirement for you to meet the standards that you hold yourself to. Do you think that will change anything significantly? If the online harms Act had been in force over the summer, do you think Google would have acted differently with regard to some of the issues that we have been talking about—whether that is about advertising or about the misinformation being returned from search results?

Maybe I could back up for a second and mention that although I am not familiar with that particular example, the Advertising Standards Authority is what is called a “priority flagger” for us. We integrate with its alert system and it sends us flags. We do have collaborations with it in multiple spaces, but we can follow up on that point. On your question about the online harms Act, we have not waited for regulation to do this work around the world, including in the UK, and we are making changes to be ready for the Online Safety Act. For the illegal harms codes, we have our systemic risk assessment. We have made some adjustments to our appeals flows, for example, and our illegal content takedown flows, to adhere to that regulation. We are also working closely with Ofcom on the child safety codes that come into force later this year. In terms of whether it would have made any difference in the example over the summer, I think the illegal harm codes would have probably made a difference to overall ecosystem safety. I think that our policies and our approaches meant that we were not particularly implicated in what happened in the Southport situation, but obviously, it was an atrocious attack and our thoughts and sympathies are with the families who were impacted.

Obviously, we echo your words about the Southport attack, which was horrendous for all those involved and those touched by it. We heard evidence from witnesses about the impact in communities across the country of the riots that followed and the misinformation and disinformation that Ofcom found had driven and helped to incite those riots. My understanding from what you are saying is that you are doing everything you can to limit the spread of misinformation but everything you do is really based on your own processes, mission and standards. Would external regulation be helpful in making you or some of your competitors act more quickly? Would it enable us to have more detailed answers about how quickly things should be taken down and the transparency of the algorithms you use?

Online safety is a “squeeze the balloon” problem: when you have policies and processes that can mitigate harm in a particular part of the online ecosystem, that harm will migrate to other places. What will be really interesting with the Online Safety Act are things like the categorisation of services—making sure that that includes services where the harm is actually playing out. As the Online Safety Act continues to evolve—and we are collaborating closely with Ofcom on it—it is also really important that we should not tie our hands when it comes to the use of some of the new technologies such as AI to support our detection work. We have seen really interesting experiments using AI to support our classifiers: creating synthetic datasets to train classifiers faster and using AI to make sure that a classifier can adapt to an emerging hate narrative more quickly, for example. We want to make sure that our hands are not tied when it comes to that really interesting space. The final interesting area is that there is a lot of value in signal sharing across the ecosystem. We see that in the scam space, with things like the global signals exchange, which we have rolled out to share signals between online platforms and banks, telcos and law enforcement. That signal matching, so you can see a full picture of what is playing out and react more quickly and appropriately, is a really important part that will require Ofcom and the ICO to collaborate and make sure that the signal sharing is facilitated.

I appreciate that response. You have talked about the proportion of Google’s revenues and profits that come from advertising. Do you think that those signals and that sharing can be as strong as the incentive to have more adverts to serve and have more content that people are accessing regardless of the standard of that content?

It does not serve our business to monetise unsafe content or run unsafe ads. That would violate people’s trust in the ecosystem so it is absolutely in our business interests to make sure that that ecosystem is safe and trustworthy. My team spends a lot of time ensuring that every day.

What concerns me is the statement that you made earlier: that you had very little to do with the Southport misinformation. Yet you have been presented with evidence here that you helped to monetise the creation of a key piece of misinformation that was used to incentivise the riots. Have you done any reflection post the Southport riots about how your advertising and monetisation of content may have contributed to the riots that we saw across the UK?

Absolutely. Any time an incident plays out in the real world—a terrorist attack or any real-world harm situation—our strategic command teams do a root cause and corrective action assessment. We post mortem whether anything did go wrong and what we could learn from that. We roll those learnings into our policies and important processes on a real-time basis—so absolutely.

Can you share that with the Committee?

The post mortem?

Yes, the post mortem.

Let me take that back and see what we can make transparent.

I have a final question. One of the things you keep saying is, “Hold us to account.” You are, essentially, the monopoly in this area. You are—let us call it as it is. You are market dominant.

You have significant market power—I do not think Google could argue with that.

Within the Data (Use and Access) Bill that is going through the House, section 123 is about ensuring that researchers, non-profit companies and others can hold companies like yours to account. It is about ensuring that they can access the information necessary for academics and researchers to evaluate and hold you to account, and to share that information with legislators such as ourselves. Do you support that?

I am not an expert in the data Bill specifically, but I can talk to you about our research and access programmes. My team actually runs some of those programmes. We have an API that allows researchers to scrape content that is publicly visible, but to do so in a way that is scalable and makes the research easier for them. If you think about scraping download, view or page visit numbers, for example, those APIs are available to researchers affiliated with academic institutions. We also provide grants to researchers to do research in the safety space, and we recognise the value of others researching as well as our own research. We engage really closely with researchers all over the world. My team runs a programme of consultation with researchers and civil society to ensure that we are getting feedback on our policies and enforcement approaches on a regular basis. We are absolutely engaging closely with researchers.

Do you also give that access to your advertising space and processes?

The research and access programmes do touch on search, play and a bunch of our product spaces. I would have to check specifically what people can get access to in terms of the advertising pieces of those products, but we could follow up and provide more on that.

Thank you for your response, and we welcome the programmes that you run. Indeed, we have received some evidence on them, and not all companies do offer the same kind of access. Is the access open to everyone? For example, on the questions that we have asked about how quickly something came down, could a researcher using API access that effectively? Google has a huge amount of data and influence. I will give an example from the Antisemitism Policy Trust, which stated that there are 170,000 Google searches with antisemitic content made each year, and when Google altered its algorithm to remove “Are Jews evil?” from its autocomplete function, the number of people making that search fell by 10%. I would like to understand whether the access that you offer as part of the programmes you mentioned is available to everyone, so that bodies other than Google can find out how quickly posts with antisemitic content get taken down, and how many of them there are. Is there that kind of transparency and access for research purposes?

There are a few criteria, including being affiliated with an academic institution. We can follow up and share those criteria, but the access is pretty open to researchers around the world for products like YouTube, for example. On your question about the time between something being uploaded and taken down, let me ask about that. I do not know that specific point, but I would imagine that the date stamp of when something was uploaded, if it is affiliated with publicly visible content and you can see when it was taken down, would be in the dataset.

Could you let us have the key parameters that the API gives access to?

Just a last follow-up from me. There is a perception today that Google can profit from false advertisement or advertisement that does not comply with your—essentially profiting from harm. Where is the investment in the repair and recovery, and do you feel that Google has a place for that? For instance, our mental health systems are at crisis point, in large part—certainly for our youngsters—due to the social media side of things. Do you think that Google has a role and a place in reinvesting in preventing harm and investing in harm recovery?

It is a really interesting question. I think you are right that harm recovery and who bears the accountability for that is a critical part. We invest a lot in our media literacy programme, to try to make sure that children, including children in the UK, are able to assess the quality of the information they are consuming and make decisions about what they consume. That includes things like Be Internet Legends and our Family Link tool, which allows children and parents to have a conversation about what the child wants to have access to and what the parent is comfortable with them having access to. We have training materials to help parents through that conversation, so that it can be managed together. I absolutely see responsibilities in terms of making sure that children are being educated on information quality and having some boundaries put in place, as well as our regular approaches to ensuring safety across our platforms.

That is the before bit—the prevention bit—and I welcome that, but I am looking at the harm being done now, while we are working on all the different AI and algorithms. Do you think Google has a role in helping to facilitate their recovery?

It is a really interesting question. We do consult with victims as part of evolving our policies. For example, in the NCII space—non-consensual intimate imagery—we have made our notification systems much easier for victims to use, and we have committed to removing duplicates, to try to prevent that ongoing harm. We work very closely with civil society groups in these spaces like Refuge and the Samaritans. That consultation with victims and experts is a key part of our obligations here, but I think it is a really fair question about how you repair, not just how you prevent.

Does Google have a charitable foundation?

We do: we have Google.org, which invests in charitable endeavours, whether that be making sure that we are providing support to charities, ad grants or training. We can share more on that if it is of interest.

I know that companies like Google often resort to tech jargon, but I point out that you referred earlier in the context of the Southport attacks to your own post mortem. I think that is a singularly inappropriate use of the word, given the actual, real-life post mortems that took place after that attack.

My apologies; I was not meaning to cause offence with that. You are absolutely right.

Thank you, Amanda. You have faced and responded to all our questions with a level of sincerity and knowledge that we have appreciated; where that has not been possible, you have offered or at least agreed to send us further information. As an engineer myself, I am really aware of the sense of a lack of accountability and a lack of agency among many of my constituents and the public generally—other MPs are nodding at me—when it comes to technology, and particularly technology from huge, multitrillion-dollar companies headquartered elsewhere. By being here and being as open with us as you have been, you have helped to increase the understanding of the accountability and the support for accountability from some companies, which I hope will give my constituents a slightly greater sense of agency. We really appreciate that. Thank you for taking the time to join us.

Thank you very much for having me here today.