VERGANI, Professor Matteo, Co-director, Tackling Hate Lab [14:08] CHAIR: Welcome. Although the committee does not require you to give evidence under oath, I should advise you that this hearing is a legal proceeding of the parliament and therefore has the same standing as proceedings of the respective houses. The giving of false or misleading evidence is a serious matter and may be regarded as a contempt of parliament. I now invite you to make an opening statement. Prof. Vergani : Thank you for inviting me and the lab to this hearing. We conducted a large-scale study of online language. We are a lab specialised in looking at big data, so, instead of looking at individual-level posts, we look at millions of data points. To do this, we use artificial intelligence in different ways. We use existing tools to capture what is toxic language. For example, we used a state-of-the-art classifier that looks at toxic language, which is language that is likely to result in people leaving the conversation online. This is the definition. And we built a custom artificial intelligence classifier trained using decisions by First Nations community members, which we identified through researchers in our field, specifically the Call It Out register. We reached out and asked if they had researchers who were able to classify what is hate, in their view. Then, we trained the AI classifier and analysed three million data points for the report we used to make the submission, and we looked at how hate spreads related to specific events. We looked at the campaign of the referendum in 2023, so we collected data between January and the referendum. And in 2025 we collected data around two events which were clearly marked by episodes of First Nations hate: Anzac Day, which in 2025 there were lots of events we can talk about but especially during the celebrations in Melbourne, when they were booing the welcome to country, and March for Australia—of course, there was the attack on Camp Sovereignty on 31 August. Around these events, we collected data and we study what hate is and how it spreads. We have, we think, interesting findings related to the different types of hate that can exist in the Australian discussions online. Also, we found that First Nations public figures are disproportionately targeted by hate compared to public figures with similar followings. We did a proper comparison. We found that certain types of misinformation are more likely to be associated with spikes in hate targeting First Nations. These are the main findings, but we have also some recommendations which we submitted, including having a permanent observatory of hate online, which helps find the best interventions, with evidence based approaches, to tackle hate. Also there should be attention on types of legal language—for example, misinformation in many cases is not illegal, but it is empirically associated with hateful language. We want to be able to tackle that type of misinformation as well. There should be more protection for First Nations public figures and people disproportionately targeted online, as we found in our data, and also platform accountability, because of their clear amplification of hate and misinformation online. CHAIR: Thank you very much. We will proceed to questions, and I'll begin with Senator Thorpe. Senator THORPE: Did my name pop up? Prof. Vergani : Yes, you were disproportionately targeted. Senator THORPE: I feel like I need a badge of honour. Your analysis of social media activity during the campaign of the failed Voice showed that racism escalated after some politicians made speeches or published articles, including John Howard, John Anderson and Jacinta Nampijinpa Price. A lot of that racism was obviously directed at me, but what responsibility do these figures have for inciting racism? Prof. Vergani : That's a great question. Clearly there is an empirical association; it's in the data. We cannot disregard the data. What happens online is clearly connected with what happens offline. It's not in a separate world. What we have seen in this study and in many other similar studies when we looked at Islamophobia or hate targeting women and other communities, is that hate is often a response. What we have seen in this study—and in many other similar studies when we looked at Islamophobia or hate targeting women and other communities—we found clearly, and also here, that hate is often a response. It comes after certain events online because it gives a platform, really, to people who hold prejudice to express this prejudice. It's really hard to say if there is a personal intent. Well, we can all speculate and make our judgement, I guess, but the data clearly says that these instances were associated with a spike in hate. Senator THORPE: My workplace is the federal parliament, otherwise known as colonisers headquarters. It's okay to be racist in my workplace. It's an everyday occurrence in the Senate chamber. There are no protections, particularly for women of colour. Mehreen Faruqi, Fatima Payman and I have raised this with the Senate president, with the Special Minister of State—I always forget his name—Don Farrell. We've raised it with the Parliamentary Workplace Support Service. We've raised it with the Human Rights Commission. We have no right to be safe in our workplace from racists. You can be racist in the federal parliament. What message does that send to the rest of the country? Prof. Vergani : Clearly that it's okay to be racist, I guess. But it's what we see, of course, in social media. I mean, the main evidence we have from a range of studies, and very well structured and reliable studies, is that hate in online environments can facilitate more hate because it basically normalises a certain type of language, which over time is maybe assumed as being normal and okay by people. So the more hate you have in a certain social environment, in social media, the more hate you will likely have in the future. Senator THORPE: My last question: according to your heat map of racist social media activity, Victoria is one of the most racist states—is that correct? Why is so-called Victoria so racist? Prof. Vergani : There are lots of methodological caveats in there because we used geolocation metadata, which is not available for all users, so it could be that the data is biased in a way that we don't know. Nonetheless, we decided to present the data because we think it's also interesting, and we cannot find an obvious bias. Honestly, I don't know why; we don't know. We have tried to understand it, but it's certainly a future direction for research. Users in the states of Victoria, Western Australia and Queensland look more likely to express hate. I don't know why. Senator THORPE: I will ask one more question really quickly. It's about the orange elephant in the room that I spoke about yesterday: Senator Pauline Hanson. Has there been an increase in hate towards Aboriginal people since the rise of Senator Hanson? Does it correlate with the rise of Senator Hanson? Prof. Vergani : Do you mean the rise in opinion polls? Senator THORPE: The rise in hate. Racists find a platform, they get on the TV, they can get a free jet and a free Ranger, and it becomes normalised. Everyone loves that, but has the rise in the racist party also seen a rise in hate and racism against First Peoples? Prof. Vergani : Unfortunately, we don't have data in our dataset to respond, but it would be great to have a look at it really. It'd be interesting to see if visibility of Pauline Hanson's party and associated political figures is associated with spikes in hate. I don't know. I don't have the answer in our data—not yet. Senator THORPE: If you do do the work—I'll put it on notice and, if you get it done by then, you do. I'd be interested in what that looks like too. Thank you very much for your work. CHAIR: I'll go online to Mr Gee. Mr GEE: Thank you very much for your evidence. I'm interested to know how pattern based monitoring works in practice. What are you looking for, and, in practice, how does that lead to the prevention of online racism using AI? How easy is it to implement? Prof. Vergani : I understand you want to know more about the methodology we use to classify hate and how our studies can suggest intervention points. Is this the question? Mr GEE: You've basically said that effective responses require a shift from reactive approaches to approaches that address patterns, amplification and underlying narratives. I'm wondering, practically speaking, what that looks like and how those approaches in practice can lead to the prevention of racism online. Prof. Vergani : Let me give you a few examples. No. 1, we saw that certain types of misinformation—and we specify the type of misinformation, which is, for example, the one suggesting that a victory of the 'yes' at the referendum would have meant that people could have been evicted from their properties. That type of misinformation was clearly associated with increases in hate. If we are able to identify early—and we can with artificial intelligence—the spread of these narratives, we can, for example, ask platforms to avoid algorithmically recommending this type of content because, although it is not illegal, it is very likely to create a spike in hate, so avoiding amplifying it would be a way to reduce hate in platforms. This is one example. Another example is that we know that the majority of hate originates from certain clusters of users and targets certain, specific people—we know already online where. At the moment, for example, platforms and anyone looking at online hate, even government agencies, are reactive in the sense that they try to remove individual posts online after they are published, and sometimes they stay in the platforms for a long time. If you already know where to look for them, for example—this is a second example—you can immediately go and remove these posts because you know that they are more likely to originate from certain people and target other peoples. They are just two practical examples—I hope they make sense—to tell you how our work can help identify evidence based responses that can be effective in mitigating hate. Mr GEE: I suppose though that it's either the platforms having to sign up to doing this or having to be made to do it. Is that the bottom line here? Prof. Vergani : In relation to avoiding algorithmically spread hate, yes, we rely on the platforms. So I agree with you—either we convince them to do it or we make them do it. I don't know how—what is most effective—but totally we need them. There are other types of interventions which are proven to be effective—for example, inoculation and other text based intervention—which can be done without the intervention of the platforms. They are clearly less effective than intervening when you can tweak the algorithms within a platform, but there are other things that can be done without necessarily having the platforms on board. Mr GEE: What would they be? Prof. Vergani : One example is inoculation. We have done lots of studies on that, and it has proven to be effective in randomised controlled trials and good scientific studies. That consists of basically educating online users in recognising the ways in which people try to manipulate them. Instead of telling them what they believe is wrong, we tell them, 'Look, have you noticed that people try to manipulate your emotions by using these strategies?' It is an effective intervention because it is associated with recognition of attempts to manipulate someone, and then this person exhibits a resistance against the message. Mr GEE: Thank you very much. That's very interesting and very helpful. No further questions, Chair. CHAIR: Senator Liddle. Senator LIDDLE: Professor Vergani, I just want to ask: how did you define racism? Prof. Vergani : Like— Senator LIDDLE: How did you define it? Prof. Vergani : In the study? Senator LIDDLE: Yes. Prof. Vergani : I have written a large systematic review on definitions of hate and racism. There's a huge number of definitions and disagreements. If you ask me, I believe racism is found in very different places, not just interpersonal interactions; it's found in structures, in algorithms, in institutions and in policies that people experience—in many places. For this specific study, we used two approaches. One was to classify language that, like toxicity, has features that are clearly harmful for users and targets. In that case, we used Google's classifiers, developed with billions of data points— Senator LIDDLE: I just want to be clear. I'm trying to understand the threshold you used to define how you counted, quite clearly, what you defined as racism. Prof. Vergani : Okay. We don't exactly talk about racism, because we know what we are measuring in the report. In the report, we talk about toxicity and identity attacks when we refer to classification using existing tools. The other approach we used was to train an artificial intelligence algorithm based on annotations provided by community members, which we identified through the Call it Out register. We used common definitions in the literature to identify hate as language that is dehumanising, is harmful and attacks the identity of the target. Senator LIDDLE: Do you say how you defined hate or racism in the report? Prof. Vergani : I think we have a couple of pages on which we explain the different types of hate. We captured specifically Voice based hate and racial identity hate. They're on one page. It's a page on which we define and provide examples of these types of language. Senator LIDDLE: You gave us an example of what you said you saw as misinformation. What else? How did you define what misinformation was? Prof. Vergani : Misinformation? We relied on the Associated Press Fact Check initiative, which provides a list of verified false claims that were spread during the referendum campaign. We only focused on the ones associated with First Nations peoples and the Voice. It's also there in the report. I think in a footnote we also provided the link to the database that we used to identify misinformation items, and we looked exactly for them in the data. Senator LIDDLE: My colleague Senator Thorpe asked whether she featured in here. You've got some people featured in this report. I can't see in here. Did you identify hate related to those who signalled a 'no' vote for that campaign or did you just do it on people like Senator Thorpe—how did you define who you would do that on? CHAIR: Senator Thorpe was a 'no' vote. Senator LIDDLE: Yes, that's what I'm saying. CHAIR: Sorry. Senator LIDDLE: I'm saying: how did you identify who you would look at? Prof. Vergani : That's a really good question. I think we explained in the methodology section that we collected data through a query, and we have provided a list of hashtags and keywords that we used to retrieve the data. These hashtags and keywords represented a range of political opinions—so both the ones, let's say, associated with the 'yes' vote and the ones associated with the 'no' vote—to make sure we had a comprehensive dataset, not a biased one. Then we ran our AI classifier through the whole dataset. To understand whether specific First Nations public figures were more targeted than others, we compared them with non-First Nations public figures with comparable following. Say there was someone followed by 15,000 people on social media platform. We compared that person with another public figure followed by more or less 15,000 people. Of course, clearly, the more visible a person is, the more hate and also interaction they drive, but, if you compare people with the same following, you are pretty much comparing apples with apples, right? Senator LIDDLE: Have you done the same work on other things? Earlier this year, parliament passed legislation for combatting antisemitism, hate and extremism federally. Did you look at what happened then? Prof. Vergani : No, the dataset is from 2023. As we said, we have limited—I would love to look at everything, but— Senator LIDDLE: Who sponsored this work? Who paid for this work? The university? Prof. Vergani : It was internal funds we have at Deakin, so it's a bit limited. Senator LIDDLE: It would be interesting to understand what happens when you pass significant legislation like that. Prof. Vergani : I really agree. This was specifically looking at anti-First Nations hate online, so this is why we focused on events that were likely to drive that type of language. If we focus on Islamophobia or antisemitism, we focus on different events. Senator LIDDLE: If I look at one of these charts in your report, daily counts of Voice based hate—the red line is racial identity hate and the blue line is during the referendum campaign—they're kind of tracking a little bit up and down, but there is a considerable change. It starts in July 2023 and significantly changes in September 2023. I'm talking about significant changes. And then, in November 2023, it literally drops back to below where it was even when all this conversation started. How do you explain that? Prof. Vergani : It's a really good question. We are not able to look at hate everywhere online. We have to collect a sample of data. So there are three million data points, which is a lot, but they are nothing. They are a small fraction of all the data that Australians produce every day. So we have a query and—think about having a telescope—we are zooming in on one area of the landscape. So in that area, which is what we see in the report, we saw a massive increase because we were looking at conversations using campaign hashtags and all that. Now, those conversations ended. The conversation using the campaign hashtag, which we were monitoring, ended. It doesn't mean that the same people were not discussing it using other hashtags or in other ways or that hate ended. But it means that, within those conversations associated with the campaign and campaign hashtags, actually these conversations dropped. People stopped using these hashtags after the referendum. Does it make sense? Senator LIDDLE: Hashtags or hate commentary? Prof. Vergani : Hashtags. Hashtags are the way that we used to retrieve the data. Within that data, which is the zoom—the part of the landscape we see with our telescope—we monitor the hate and how hate increased or decreased. But it's always within a specific sample of conversations. Senator LIDDLE: I'm trying to understand what value this actually has in telling us how much hate and where it was directed during that campaign, because— Prof. Vergani : During the campaign, a lot of value. Senator LIDDLE: it just completely drops off, according to this. Prof. Vergani : No, but the conversations containing—people talking about the Voice campaign and the referendum, between January and the referendum date, used mostly hashtags associated with the referendum, which is #TheVoiceReferendum or #VoteYes or #VoteNo or everything else—#IfYouDontKnowVoteNo. We comprised the discussions including all these hashtags and keywords, which means we captured, we think, a big portion of the conversation online about the referendum. Within these discussions, hate massively increased. These discussions stopped happening with these specific keywords and hashtags in the week after the referendum, so people stopped talking or using #VoteYes. Does that make sense—how the data was collected? Senator LIDDLE: I still think it would be useful for you to provide the committee how you defined what hate was when you actually went in there or you defined what racism was, rather than people who were actually talking about it. What was the threshold where you went from free speech, people objecting, people aggressively objecting, to hate— Prof. Vergani : Yes, totally. I can do that. Senator LIDDLE: and racism? If you can provide the committee with that—not now but on notice—that would be helpful. Prof. Vergani : But it's in the report. It's in the page providing examples with Voice based hate and race based hate. It's a full page with two different paragraphs and examples. I can— Senator LIDDLE: You can send that to us, because I think I got this yesterday. So it'd be very useful for you to direct us to where that is, rather than us having to search through a whole heap of data points. Prof. Vergani : It's in the methodology— Senator LIDDLE: Thank you. CHAIR: Yesterday morning, we heard evidence in particular about the targeting of First Nations content from bots online. Do you have any evidence or data about bots and how they target First Nations content? Prof. Vergani : I wish we had. Old bots were easy to detect because they were behaving like bots—spreading and publishing 15 posts in 30 seconds. Current bots, contemporary ones, are really good, and it's impossible—I don't know if it's impossible, but it's a really hard technical task to distinguish them from real people. So I wish I had this information, but no. CHAIR: We had Meta at the Brisbane hearing. They have a combating online hate advisory group. Are you a part of that at all? I'm wondering what advice the group gives Meta that tells them that it's okay not to have any of their moderators on Australian shores when they are making decisions about what is racist or not in Australia. Prof. Vergani : I'm also wondering. They established the advisory group maybe two years ago, if I'm not mistaken. They invited a range of academics and community activists. I actually participated because I thought, 'Well, I'm going to learn something about how they do, for example, moderation, or I'm going to finally see their data,' because they never release the data. If they say, 'We remove 70 per cent of the hate,' how do we know it's 70 per cent and not three per cent and not one? We have to trust what they say because there is no accountability and no transparency. So my biggest request was, 'Can we see the data?' I never got into the data. We met probably twice—and not in the past six or seven months. They never asked for any advice. I would have loved to give some advice. CHAIR: So what happens at those meetings then? Prof. Vergani : There are presentations about how they do moderation or the introduction of new policies, and then people say, 'We think it doesn't work,' and they take notes. CHAIR: So you don't know if the things you raise in there are actioned. The experience of users online would probably suggest they are not. Prof. Vergani : Probably not. CHAIR: There was an ABC article that came out yesterday talking about Facebook paying controversial creators to produce rage-bait content. Prof. Vergani : I saw that. CHAIR: The article suggests that, in effect, Meta is in a direct commercial relationship with content creators that have links to Neo-Nazi far-right-wing extremist groups. It seems like this would be an important thing to bring together a group like the one that you're on in Meta to talk about. Meta is paying for far-right extremist content on its platforms. How does that contribute to the hate and racism that's directed at First Nations people? Prof. Vergani : Massively, of course. CHAIR: We asked them in Brisbane about hate and racism being profitable. They of course said that they tried not to have it happen. But this article says that this is an invitation-only program—so they are explicitly inviting groups who are creating hate content. It seems very at odds with the evidence they gave to the committee that they are not profiting from hate and racist content on their platform. What advice are you going to give them the next time you meet? Prof. Vergani : The question is: what can the government do to make them stop and behave in a responsible way? CHAIR: We're happy to hear what recommendations you think government— Prof. Vergani : No, I don't know the answer. Their headquarters are in the United States. Clearly, they are protected by the current administration. So what can a country like Australia do? There are already a few issues with the Trump administration tariffs. What can Australia do to enforce a different approach to regulation of content online when these companies are protected by an administration in the US which defends them? CHAIR: We have seen moves from the Australian government, like the banning of under-16s from social media. The other thing Meta pointed to in their evidence is that they comply with the laws of the countries they operate within. There are some levers there, in thinking about what their presence in Australia looks like, or should look like, or how we protect Australian social media users from harmful content and things that I would suggest don't align with Australian values. Prof. Vergani : I agree, but the problem we face is that they are not accountable. They tell us they comply, but how do we know they do? My research shows that there is still a lot of hate in platforms that is not removed. Some of the hate is removed—we find that out in our data analysis—but we don't know the proportion that has been removed. Many community organisations report specific individual data points; sometimes they remove them, sometimes not. But how do we check that they comply if they own the data and only they can see the data? This is the question. CHAIR: What does accountability look like—transparency in data? Prof. Vergani : Transparency in data, absolutely—and in the algorithms. CHAIR: And in the algorithms—the recommender systems. I get pushed far-right content, and I definitely don't want to see it. Senator LIDDLE: Can I ask for a clarification on the transparency of the data; you can take this on notice. On the page that talks about the heat map of hate activity per 100,000 people, it says that Western Australia and Queensland remain among the higher activity states, while the ACT and the Northern Territory remain lower. Is that 100,000 based on a proportion of the population or is that just based on 100,000 people? This doesn't seem consistent with how people voted on a geographic basis. This doesn't look right to me based on the actual vote per state and territory. This is inconsistent, it appears. If you could explain that to me, that would be good. Prof. Vergani : Let's say we have—and I'm making up the figure—100 hateful users in Victoria and 50 in the Northern Territory. Does that mean there is more hate in the population of Victoria or of the Northern Territory? It depends on the way in which you look at the data. In absolute numbers, which is not what that map shows, you can say that, yes, there is more in Victoria. But if you divide 100 by the population of Victoria and 100 by the population of Northern Territory, you will probably find that there is more hate per capita per person. In the case of that figure, per 100,000 people, there are more hateful users per each 100,000 people. Senator LIDDLE: But there are also fewer people. Prof. Vergani : That's the point, exactly; there are fewer people. There might be less hate in absolute numbers, but in relative numbers, relative to the population of the state, there is more hate. That is what the map shows—so, relative to the population of the state, how many hateful users you have per 100,000 people. It's a hate rate, not an absolute number. Senator LIDDLE: But you can't, based on that, equate hate with the actual vote. If you look at this, you've got, with the actual vote for each state and territory—you can't. It says South Australia is sitting there quite different from the other states, yet it was the second-highest 'no' vote in the country. Prof. Vergani : That tells you—basically, in the heat map, the colour changes with the change of how many hateful users you have per 100,000 people on average. CHAIR: So how active people were potentially on their social media platforms by state rather than the vote outcome. Thank you very much for your evidence to the committee this afternoon. I imagine you'll get some questions on notice—certainly from me—after today. Answers to questions on notice are due by 28 August 2026; the secretariat will make sure you know about that. You'll be sent a copy of the transcript of your evidence from today, and if, when you have a read through it, there are any errors you want to see fixed, you can request that that happens..