FRONDORF, Mr Evan, Head, External Policy and Partnerships, Safeguards, Anthropic [by video link] Committee met at 0 7 : 59 CHAIR ( Senator Whish-Wilson ): I now declare open this hearing of the Senate Select Committee on Information Integrity on Climate Change and Energy. I begin by acknowledging the traditional custodians of the land on which we meet, the Ngunnawal and Ngambri peoples of Ngunnawal country, and pay my respects to their elders past and present. I extend that respect to all Aboriginal and Torres Strait Islander people here today. The committee will be conducting today's hearing in person. These are public proceedings being videostreamed live via the parliament's website, and a Hansard transcript is being made. I remind overseas witnesses that parliamentary privilege does not extend to giving evidence outside Australia. I also remind witnesses that it is a contempt to give false or misleading evidence. Witnesses also have the right to be heard in camera, which is in private session. If a witness objects to answering a question, they should state the grounds upon which the objection is made, and the committee will determine whether it will insist on an answer, having regard to the grounds on which it's claimed. If the committee determines to insist on an answer, a witness may request that the answer also be given in camera. I remind senators of their obligation under the Behaviour Code for Australian Parliamentarians to treat witnesses with dignity, courtesy, fairness and respect. I remind everyone engaging in the public hearing, including witnesses and members of the public, that the behavioural standards for Commonwealth parliamentary workplaces require people to act respectfully, professionally and with integrity; encourage and value diverse perspectives; and recognise the importance of a free exchange of ideas. Bullying, harassment and discrimination will not be tolerated or condoned. I now welcome a representative of Anthropic, Mr Evan Frondorf, via videoconference. I understand that information on parliamentary privilege and the protection of witnesses giving evidence to Senate committees has been provided to you. Thanks so much for appearing here today—all the way from New York. We very much appreciate it. Would you like to make a statement to the committee? Then we can go to questions. Mr Frondorf : Yes, absolutely. Chair and senators, thank you for the opportunity to appear today. I've been at the company for almost three years and have a previous background in trust and safety and fraud prevention. Anthropic is a leading frontier AI developer, working to build reliable, interpretable and steerable AI systems. Our AI assistant, Claude, is used by millions of people worldwide. That ranges from large enterprises and government agencies to small businesses, researchers and individuals. They're using Claude for tasks like software development, data analysis and scientific research. We believe that AI can deliver enormous benefits for society and we're committed to realising those benefits while developing the technology safely and responsibly. We're also incorporated as a public benefit corporation, with a stated public benefit purpose of the responsible development and maintenance of AI for the long-term benefit of humanity. That structure allows us to weigh the long- and short-term consequences of our work on society. That commitment to safety is reflected across our work and our products. Anthropic had a trust-and-safety function, which we call 'safeguards' before we even shipped our first product, and we're also deepening our commitment to Australia. We recently announced that we'll be opening an office in Sydney later this year, and we look forward to working closely with Australian institutions, researchers and the government. When discussing misuse, it's important to provide some context relevant to this committee's work. Unlike social media platforms, Claude doesn't have a recommendation algorithm or a content feed. It's a 'pull' system, rather than a 'push' system. Content does not surface to a user unless they actively request it, and that means that some of the spread of synthetic media and deepfake concerns that often dominate this conversation are less central to our risk profile. What we see as the more salient risk for a system like Claude is its potential to automate or scale influence operations activity. That's squarely where my team sits. Safeguards is responsible for keeping Claude safe once it's out in the real world. We bring together people across policy, enforcement, threat intelligence, data science and engineering teams. These are people who think like builders but also like our adversaries, anticipating misuse and building the right defences. To misinformation, particularly, we take a layered approach, with defences built in at the training, guardrail and monitoring levels. We train Claude to be factually accurate and honest about the limits of its knowledge but also to engage with a wide variety of perspectives and provide balanced commentary on political topics. Our usage policy explicitly prohibits Claude from creating or spreading misleading and deceptive content. Once our models are deployed, we run automated detection systems and staff a dedicated threat intelligence team that investigates coordinated misuse. It includes influence operations that we've detected, disrupted and reported on publicly. I will do my best to answer all of your questions this morning. Thank you, once again, for the opportunity to appear. CHAIR: Good on you, Evan. I'm going to go to the deputy chair to ask questions, but I do have a cheeky question for you to kick off: did you use Claude to help you with your opening statement today? Mr Frondorf : I would say that it helped with some fact-checking and some drafting, but it mainly came from us. CHAIR: Great, because it might be a bit spooky if Claude were writing his own summary before a Senate committee! Deputy Chair. Senator ANANDA-RAJAH: Thank you, Evan, for appearing today. I know we issued this invitation at short notice, so we're most grateful to Anthropic for accommodating us. Welcome to your first Senate hearing; I'm sure this will be the first of many to come, particularly if you're planning on establishing yourself here in Australia—which we welcome. I have a question regarding Claude. Claude is trained to be a helpful, honest and thoughtful assistant. Can you explain how these AI models and the training weight scientific consensus and authoritative sources in response to questions around climate or energy? Mr Frondorf : Thank you for that question. At a high level, training an AI model like Claude involves two main stages. Firstly, the model is learning from a really large body of diverse information sources—that's including publicly available information from the internet, licensed data from third parties and data that we generate internally. That allows Claude to build a general foundation of language and knowledge. The second stage is where we shape the model's behaviour, which is how it responds to questions, what tone it takes and how it responds to sensitive or controversial topics. That's where one of our really important tools comes in, which is called our 'constitution'. The constitution is a set of principles that we've developed that guide Claude's behaviour. Some of those principles are directly related to how it should think about information like this: it should be factually accurate; it should represent consensus where it exists; and it should be honest about the limits of its knowledge, but it should still engage with a range of perspectives, where appropriate, on contested topics. Senator ANANDA-RAJAH: Could you provide the committee with a list of the principles in the constitution that you use, if that's not proprietary? Mr Frondorf : Sure. Our constitution is publicly available and is published on our website. I will say it is quite the lengthy document, but one of the principles that I think is at play here when we think about these topics is that Claude should be honest and truthful. It should sincerely only assert things it believes to be true. Other principles are that: it should also have calibrated uncertainty, based on evidence and sound reasoning; it should acknowledge its own uncertainty and lack of knowledge, where relevant; and it should approach questions empirically and rely on legitimate epistemic actions like sharing evidence or providing examples. This is the language that's in the constitution that speaks to these areas. Senator ANANDA-RAJAH: Is it conceivable that Claude could inadvertently amplify fringe or other views, particularly around climate or energy? Mr Frondorf : This is where our approach to misinformation and disinformation comes in, and it's multifaceted. I would say it relies on multiple layers of defence. There's that training level that we spoke about, where we prioritise these behaviours and go through significant training and learning steps to reinforce these policies. We also then undergo extensive testing before we deploy any of our models to test how they perform against those principles and against our policies. Our system cards include results on both even-handedness and how neutral the model is. There are also honesty evaluations on whether Claude is accurately answering questions or appropriately stating that it doesn't know the answer. But, as you know, malicious use or misuse is possible. When our models are deployed, we rely on a number of techniques to detect that sort of usage, or that behaviour, that might violate our usage policy. We use a number of smaller models or detection tools that are called 'classifiers'. These are trained to screen models, and inputs and outputs, for potential violations of our policy. They can then take action, such as blocking the response—in a severe case—or surfacing accounts for review by our enforcement team. That's that additional layer. Senator ANANDA-RAJAH: No, that's good, it sounds like there's an internal safeguard there. If you had a determined bad actor, would they be able to skew the model just by bombarding it with misleading or false information and content? Mr Frondorf : This is where we aim to have those many layers. We think those multiple layers are important to stop behaviour like that. Along with the training, and along with the classifiers that detect that policy-violating activity, we also maintain a threat intelligence team; they follow up on external leads or other indications of abuse. We work with a number of external monitoring solutions and we monitor outside networks to see where content might be spreading or what kind of influence operations content is out there. We check to see if any of this can be traced back to Claude generation and, if so, we will put a stop to that—obviously—and take down those accounts. But that is kind of an additional layer of defence beyond what we already do. No AI model is perfect or deterministic on every response it makes, but that is why we have these multiple layers. Senator ANANDA-RAJAH: Okay. As you know, Australia is on the front line of climate change; even as we speak, there are parts of Queensland that are underwater right now. If we're not underwater, we're burning somewhere. This is the reality—this is the lived reality—for thousands, if not millions, of Australians every single year. We have heard testimony that there is a surge of false or misleading content around major events, particularly on social media. Given our vulnerabilities in this space, what kinds of safeguards would Anthropic put in place, particularly around natural disasters in Australia, to prevent the surge of false and misleading content that could mislead both Australians in the heart of those communities that are affected as well as first responders? We do have evidence that this is happening in Australia. Mr Frondorf : Thank you for the question. I think many of the guardrails that we put in place around training, around those classifiers, are designed to be evergreen and to work across all harms that we might encounter and all misinformation harms. But we do pay close attention to current events and to news, which would include some of the natural disasters or the other events that you mentioned. We do investigate, including through our threat intelligence team, the types of misinformation campaigns or what types of storylines are out there. We are able to trace those back to potential usage on Claude, to see if Claude has been implicated in generating that content. So we do focus on events and that does drive a lot of our investigation priorities, based on what we're seeing in terms of current events. Senator ANANDA-RAJAH: I'll go to the role of civic society, particularly researchers. I noticed from your Transparency Hub document that you do a lot of work with third parties on extremist policing and extremist content. You work with the US entity, the NAIRR—the National Artificial Intelligence Research Resource—which I want to draw you out on. It sounds like it's some sort of new public-private partnership. We have a deep interest in Australia in empowering our civic society—particularly our researchers, whether they be in academic institutions or not-for-profit. They have yielded a treasure trove of information to this committee during this hearing. Can you just talk us through what your work is like in the US, or globally, with third-party researchers, and what you're considering here in Australia? Mr Frondorf : Yes— Senator ANANDA-RAJAH: Just to be clear, it's not that I don't believe what you're telling me regarding your safeguards—and you've given very comprehensive responses—but we need external parties to mark your homework! Mr Frondorf : I certainly understand where you're coming from there. Working with third parties is key to our approach on safeguards and also widely at Anthropic, so I'll speak to both of those elements. From a safeguards perspective, we partner with various groups who stress test our models for us and provide feedback. That has ranged from groups like the Institute for Strategic Dialogue, which has actually helped us research the topic of election related misinformation and influenced our model behaviour there. We've also worked with groups like ThroughLine in New Zealand and the International Association for Suicide Prevention on topics around mental health and caring for our users. We've actually partnered with many of those organisations to help provide up-to-date external resources in critical situations when we detect conversations like these. That's partly the safeguards approach. On the overall Anthropic approach, including when it comes to testing model capabilities and risks, we also work with a number of third-party organisations on that testing. You mentioned US bodies. We worked with the US's AI safety institute as well as the UK's AI safety institute on assessing our models before release, and we certainly welcome the opportunity to have those conversations with civil society and with the government in Australia. Senator ANANDA-RAJAH: In Australia, I think it is incredibly important that our civic groups and our researchers are oxygenated in this environment. We don't want to see them pushed to the fringes by big tech companies coming in. I think they have a really, really important role in civic oversight of tech platforms, and I'd really ask the company to consider engaging in funding some of our civic groups in Australia for periods of three years or more. I think that it would be seen as an act of good faith, and it would help maintain your social licence in this country. That's just something I'd like your company to consider. Mr Frondorf : Yes, I appreciate that. It's absolutely something that we will take under advisement and consider. We understand the importance of working with civil society. That's a core part of my role. We know we can't take on these problems alone, and we want our work to be informed by the context of experts and local expertise as well. So, we certainly appreciate that. Senator ANANDA-RAJAH: Thank you. CHAIR: I'll ask a few quick questions, Evan, and then I'll rotate the call. As you may be aware, we've had many of the key tech platforms provide evidence. One thing I've asked every one of them consistently is: does your company have a position statement or a policy on climate change, reflecting either the science and the consensus or your own ESG sustainability guidelines? The reason I ask that question is I think it goes to your credibility in dealing with climate misinformation and taking it seriously. Mr Frondorf : Yes. Thanks for that, Senator. I would say that I am not aware of—or I do not work specifically on—what our overall policy statements are as a company. But, when it comes to climate responses in the model, Claude should acknowledge where the scientific literature points and where scientific consensus exists. When there are questions around policy and politics, it should engage with the perspectives there. I can't speak to any kind of official statement from Anthropic on the issue. CHAIR: If you could take that on notice, whether you do have a position statement or a policy, I'd appreciate it. Mr Frondorf : Yes. CHAIR: I understand it's a tough environment for you in the US at the moment, with President Trump being an open climate denier on the world stage, but if you do have a position I'd be interested in it. In line with this, I read that Anthropic was lobbying the US administration for a national fossil fuel target to help power its centres. Are you able to tell us if that is correct or if you have, for example, a renewable energy target that you've been lobbying for—but mostly for your data centres. Mr Frondorf : Unfortunately, I can't speak to that topic. I'm not aware of what we've lobbied for or what commitments might exist in that area, as it's somewhat outside the scope of my role. CHAIR: No worries. Once again, if you could take that on notice, it would be appreciated. You mentioned scientific consensus a bit earlier. One thing that's different about information integrity around climate change—this was recently raised at the COP of truth in Brazil. You may note that disinformation and climate impacts have been labelled, for many years in a row, by the World Economic Forum as the biggest risk facing humanity. I'm interested in how your model, especially with search engines, would incorporate scientific consensus—decades and decades of tens of thousands of scientists with hundreds of thousands of research papers. How do you incorporate that when, for example, people are doing searches? Mr Frondorf : The first note I'd make is that Claude's ability to search the web is only one component of its overall functionality. For the most part, people are asking Claude for information and then receiving a response. In doing that, Claude is trained to draw on everything that it has been trained on. When you speak to those decades of scientific research and all of that information that exists, it should be drawing on that information in its response and acknowledging that history when somebody is asking questions on this topic. Of course, if somebody is bringing new evidence or asking questions, what it should do is think about that empirically. It should consider the balance of evidence that exists on one topic versus the new information that has been presented, and it shouldn't be swayed by just one new piece of information, given how much it's been grounded in that literature from its training. CHAIR: I have a couple of quick ones. We had interesting evidence from a scientist here, who's one of Australia's key science educators across the board, that he's developing his own AI chatbot to provide information about climate misinformation and disinformation. What guardrails are in place to ensure that chatbots powered by Claude are providing factual and accurate information, and what's needed to ensure that into the future? It could be for good, but it could undermine climate integrity as well. Mr Frondorf : Absolutely. I think this is where our multilayer approach comes into play. We are training Claude, from the outset, to avoid misinformation, address questions factually, being honest about its limitations or lack of knowledge while being able to engage in political discussion where warranted. I think it's that training layer that first tries to stop this content from even being generated in the first place. That's where we come in, on the safeguards team, of then having this additional layer of defence. We put in our additional screening models, those classifiers. We put those in place to also block or otherwise flag activity that may violate our usage policy. That could be disinformation or misinformation that's being created to deceive or mislead others. For us, it really goes beyond just the model training. Somebody needs to then be monitoring what that model activity is or what is happening. That's where our safeguards functionality comes in. CHAIR: Just on that model, I think the deputy chair has a couple of follow up questions. Senator ANANDA-RAJAH: Has Anthropic detected any coordinated attempts to use Claude to generate large volumes of misleading information, and, if you have, where have these hotspots been identified? Feel free to venture beyond climate. We are focused on climate and energy, but obviously this relates to a whole range of other harms. Mr Frondorf : Thank you for that question. We think it's important to be transparent about the risks that exist. There are two main areas where we believe misinformation poses a particular risk for AI. The first is around models perpetuating that misinformation or falsehoods, which we aim to mitigate with our constitution and our training process. The second, as we've been discussing, is around individuals or groups using AI to spread misinformation or automate tasks and influence operations, which we detect through those guardrails and that ongoing monitoring. Given Claude's capabilities, the greater concern that we have is less around, say, the generation of the content itself and more around the power of agentic tools and coding tools to automate that. For example, in a threat intelligence report that we published in March 2025, we did detect the use of Claude by actors to run what we called an influence-as-a-service operation. In that campaign, Claude was used in violation of our policies to basically orchestrate the operation of multiple social media accounts. This was less about the generation of the content itself. It was a really novel attack where Claude was misused to dictate posting schedules and to dictate who these accounts should interact with. That sort of activity is a concern and is a focus for our team. When we detected that, we shut that campaign down, and it influenced our monitoring capabilities going forward to detect that activity earlier. That is the form of activity that we've seen, beyond just misinformation, in the world of fraud and scams, where Claude was used as an orchestrator to help automate some of these activities. Senator ANANDA-RAJAH: That's really interesting. Were there any regional hotspots in the world where you identified this sort of coordinated inauthentic behaviour? Mr Frondorf : I can't speak to this in exact detail, but I do know that, for this campaign in particular, it was pretty wide ranging. It actually appeared to be a service that maybe multiple politically motivated actors from across the world were using. We saw very disparate activity of influence in a European political campaign and discussions around the Middle East. So, at least for this campaign, I don't think I can say there was any particular hotspot that was identified. But we did see that activity be wide ranging. Senator ANANDA-RAJAH: Thank you. CHAIR: This is the last question from me. I want to give you an example. It's not theoretical; it was provided to us yesterday. The ABC here in Australia did an article on hundreds of deepfake political posts around a politician here called Pauline Hanson from the One Nation party. I understand that they were generated in Vietnam—this is what was reported—and they've been flooding social media here. It's not just deepfake but also a lot of fake content. Could you talk us through how that could be allowed to happen. Your guardrails are at the point of creation, I understand. Is this because another form of AI was being used or another company? Where did it fail? How did this kind of thing happen? Mr Frondorf : Without speaking to Anthropic in particular, when I think in general about where AI defences fail, it can be from not having the guardrails to specifically detect that activity—perhaps because a model developer or another actor wasn't aware of the concern or was not thoughtful about it. But I think another area that we see in AI models is that there can be a susceptibility to jailbreaks, where, if you come to a model and pose questions in a particular way or otherwise feed the model certain types of text, depending on the model, you can get responses that are out of the intended scope. That's why our multiple guardrails exist. Agnostic of AI platform, that is where you might see that type of behaviour be possible. CHAIR: Can AI be used to assist hackers in terms of their operations? Mr Frondorf : Yes. Thank you for bringing that up. That is also a topic that we have discussed in some of our published threat intelligence reports, where hackers or other malicious cyberactors have been aided by AI systems, including Claude, to perpetuate fraud or scams, develop ransomware and seek to extort ransomed parties for that ransom. Increasing the cybercapabilities of models is one of our key concerns in terms of their ability to assist in that type of activity. We think there's also great strength in AI models to help detect these bad actors, detect where systems may be vulnerable and use these cybercapabilities in this way. But it's tricky to find that right balance between assisting good actors in these cyberactivities and making sure that we prevent the bad activity. CHAIR: Thank you. Senator Pocock. Senator DAVID POCOCK: Thank you, Mr Frondorf. We really appreciate your time. President Trump recently called Anthropic 'left-wing nut jobs' and an 'out-of-control, radical left AI company', and a White House statement referred to your 'woke' AI company's terms of service. We've obviously seen a lot of reporting around this, but I was hoping that you could maybe describe what actually led to this happening. Mr Frondorf : I appreciate you asking this question. That situation with the department of war is ongoing and quickly evolving. Unfortunately, I am not equipped to comment on active litigation at this time, and it's outside of the scope of my role in particular. Senator DAVID POCOCK: Sure. I don't expect you to comment on litigation, but I'm just wondering if you could inform us what actually led to that. My understanding is that restrictions were written into contracts with the department of war around lethal autonomous warfare and the surveillance of Americans en masse. I'm just wondering if that is true. Mr Frondorf : As mentioned, this is out of the scope of what I work on and my expertise. As Anthropic have stated, we do support the lawful use of Claude for defence and intelligence purposes, but with two exceptions: the mass domestic surveillance of Americans, because we believe that it is incompatible with democratic values, and the use of fully autonomous weapons, because we believe AI is not reliable enough at the moment to take humans out of the loop. I'm not able to comment beyond that, but that is our core principle when it comes to defence and intelligence. Senator DAVID POCOCK: I understand there is litigation underway, so I understand you can't comment. After being labelled a supply chain risk by the Trump administration, we've seen reports of legal action. I don't want any details there, but, as I understand it, at least two of your competitors, Google and OpenAI, have supported the court action. I'm just wondering what you think it says about the AI industry when competitors are supporting each other in court. Mr Frondorf : On this question, I think this is something that I would need to take on notice, given it's outside of my scope of building safeguards to protect the models. Senator DAVID POCOCK: I appreciate that, and I'm happy for you to take it on notice. When it comes to AI mis- and disinformation, do you accept that AI tools present a significant challenge in mis- and disinformation being created and spread? Mr Frondorf : I think yes. We agree that misinformation is one of the risks that is associated with AI. As we discussed before, it really arises in two ways. One is the potential for AI models to create or perpetuate misinformation, and the other is a model or AI's ability to automate certain operations or otherwise coordinate those operations. So I think there are both risks that are associated with generation, where we have multiple defences at the training level and the screening level, and also those risks of Claude being used as a tool for automation and misuse in that way, which is another set of defences that we have in place. I certainly agree that it is a risk. Senator DAVID POCOCK: Anthropic has been a lot more transparent around threats, from what I've seen, than some of the other AI companies. I'm interested in what evidence you've found that Claude has been used to create misleading content or to use bots. Mr Frondorf : We have published reports from our threat intelligence team of cases where we have seen Claude both used in situations that are related to fraud or scams but also in an influence operations campaign. This wasn't related to climate disinformation or misinformation, but we did find that that influence-as-a-service operation, where Claude was used in violation of our policies and to orchestrate and direct hundreds of social media accounts and do so with different politically motivated personas. So it is certainly activity that we have found and published about in our threat intelligence reporting. Senator DAVID POCOCK: What happens when you detect that? Mr Frondorf : When we detect a campaign like that, we seek to shut down that operation. I think an important part of what we do on the threat intelligence side is—since this is activity that we have investigated and shut down, then that needs to feed back into our training and detection systems. Then: what do we need to build or what do we need to enhance around our real time classifiers or around our training to block that content or block those actions from being generated in the first place? We see that as a really important feedback loop. Of course, you detect that type of activity, you shut it down, and you work to notify victims or relevant actors as needed, but then that needs to feed back into our ongoing building of defences to make it less likely for that to happen the next time around. Senator DAVID POCOCK: Thank you. Senator ANANDA-RAJAH: Can you ask if the social media companies responded. Did they take down as well? Does it count? Senator DAVID POCOCK: Sure. How many people are dedicated to looking at this issue at Anthropic? Mr Frondorf : Our safeguards team includes a number of groups. It's policy enforcement, threat intelligence, trust and safety operations teams. We also work quite closely with our research and engineering teams, who build the underlying models and technical safeguards. I'd hesitate to give a single headcount number because that work spans multiple teams and it's really embedded into the product and building of the product at Anthropic. Senator DAVID POCOCK: You don't have a team specifically looking at mis- and disinformation and the use of bots? Mr Frondorf : It is one of the multiple harm areas that fall under the scope of our safeguards team. Our safeguards team manages and develops the usage policy for how Claude can and can't be used. That includes a variety of topic areas, including mis- and disinformation. Senator DAVID POCOCK: How big is your safeguards team? Mr Frondorf : That's where, given the work with the multiple teams, I hesitate to give a single headcount number. I'm certainly happy to take on notice a more detailed breakdown if that would be useful. Senator DAVID POCOCK: Thank you. It would be very useful if you could take that on notice. On this area—and it relates to some of the questions that other members of the committee have asked—you've obviously said that you look for the use of Claude to generate mis- and disinformation, potentially direct bots to spread it, and you look to shut that down. But I imagine, once that's happened, a lot of those posts or information are out there on social media. What are you doing with social media companies to then flag that and get them to also remove that content? What has the response been from the likes of Meta and X? Mr Frondorf : Thank you for asking that. We do work with a number of external monitoring vendors to track misuse across platforms. We're also a founding member of the Frontier Model Forum to facilitate information sharing on these topics and others. That includes some of the platforms where this content might be spread. I am not party to or familiar with conversations that we've had specifically with social media platforms and how they've actioned this in particular, but we do seek to work with them and share information. Senator DAVID POCOCK: Maybe you could take on notice what the workflow is and how that actually happens and if—or if not—social media companies have actually taken down those posts. This is stretching the terms of reference, but, given we've got the opportunity with you, I really wanted to ask it. The people I represent here in the Australian Capital Territory—and, I know, Australians around the country—are concerned about the climate and environmental impacts of data centres. We're told, here in Australia, that data centre demand in Sydney alone is forecast to be roughly the equivalent of our national capital's—Canberra's—total drinking water. It was reported today in the Australian Financial Review that Anthropic is looking to expand the construction of data centres. What is Anthropic doing to ensure that there isn't a spike in emissions, water consumption and, potentially, electricity prices? Mr Frondorf : Thank you for bringing up this important topic. This topic is outside of the scope of what I do day to day, but I can address it briefly. In the United States, where our data centre build out is currently underway, we've made public commitments to cover the full cost of grid upgrades needed to connect our facilities and bring net new power generation online to match our demand. We've invested in curtailment systems that reduce the draw of power at peak times, and we also seek to deploy water-efficient cooling. We expect to take a similar approach in any market where we build infrastructure. Senator DAVID POCOCK: Just to confirm, you'll voluntarily take that approach or do you need guardrails put in place, guidelines, by the government? Mr Frondorf : I'm not the right person to speak to specific regulation, but we'd expect to take that same approach that we've taken in the US to any market. CHAIR: Before I go to Senator Roberts, I have a follow-up question to Senator Pocock's questions. A platform like Meta, Evan, requires you to indicate if you're using AI when creating a post, but I think it might be voluntary. Are the tools sophisticated enough for platforms or even third-party fact checkers, for example, to be able to detect the use of AI, and do you assist any of these groups with the detection of deepfakes and other misuses of AI? Mr Frondorf : I think one important note to call out is that Claude is a text based model today. It does not generate images or video. That certainly mitigates some of the risk associated with deepfake or synthetic content that's available. When it does come to text based content, as part of our usage policy, if someone is operating a chatbot that is interfacing with consumers, they must disclose that AI is being used. Similarly, if content is being created for a media or journalistic approach, we do require a human in the loop to be in that process, to evaluate that content before it is posted, so our policies wouldn't allow automated posting to social media platforms. CHAIR: Excellent, thank you. Senator Roberts. Senator ROBERTS: You're from an AI company, so I'm going to step back a bit and ask some broad questions, because there is a lot of ignorance about AI. What is AI? Could you include in your definition the breadth of scope that many AI providers provide. Mr Frondorf : Sure. An AI model like Claude is a large language model. What Claude does—it is trained on a very large body of text from across the internet and published literature. It learns from that very large body of information sources—internet data, licensed data from third parties, data that we generate internally—and is able to develop a broad understanding of language and knowledge. We then apply additional steps in a second stage to shape the model's behaviour around how it responds to questions, what tone it takes, how it handles sensitive topics. But, at the end of the day, when you provide an input to a large language model like Claude, you are getting a predictive function that predicts the next response or the next word or token—as we call it—based on all of that training data that it has seen before. It is taking the input and, in a sophisticated way, predicting what the next word or response would be, based on that input. Senator ROBERTS: So it's based upon hoovering up a lot of data, collecting a lot of data and aggregating it, and it can be shaped by algorithms or directions or instructions within the AI developer. Correct? Mr Frondorf : Yes. AI models are trained on large amounts of data from diverse information sources, and that gives it its general foundation of language and knowledge and how that works. Then, there is a second stage, a training stage, to shape that behaviour and provide principles to the model. The main way we do that is through our public constitution that lays out those principles that we've developed to guide Claude's behaviour. Senator ROBERTS: There are many different AIs. I don't expect you to name your competitors. Can you give us an understanding of the industry, the different types of models and the different types of restrictions and opportunities for each of them without naming them? I'm sorry, just the range, not each of them—the scope of AI. Mr Frondorf : Sure. I think what I can best do is speak to the scope of large language models. I think artificial intelligence has been used in many capacities and includes areas like machine learning or other places where artificial intelligence is being used in a predictive way. For the large language models that we're familiar with, there is a set of general purpose models that are available. I won't speak to all of the names, but these are like Anthropic's Claude model in that they can be used for a variety of purposes and really can provide answers on any topic. Many product makers or companies may seek to integrate those AI models or develop with those AI models in a way that is more specific to their use case. They may take a model, like Claude, to build a chatbot for their users to interact with for customer service, or they may use the model in a way to drive scientific research or to help with the analysis of documents. You start at those foundation models that exist, and then there's a variety of work streams or ways that developers, product makers and companies use those AI models to create their own AI products. Senator ROBERTS: Thank you, that's comprehensive. What we've got with AI is something that learns from the internet or from selected parts of the internet. So, if there's a lot of crap in the internet, it learns crap. Then you've got instructions within each AI manufacturer, and then you've got tailoring for specific purposes. Is that correct and a good summary? Mr Frondorf : I would say that all AI models are trained on a large amount of data and then go through a learning process to determine how to respond to those questions in a way that is useful to the user. Senator ROBERTS: Let's go to something the chair raised a minute ago: Pauline Hanson's party. It's actually now called One Nation not Pauline Hanson's One Nation. We have had phenomenal growth—staggering growth—and we get so many compliments for our work. We're inundated with membership, with people wanting to support us and praising us, and I mean that sincerely. Now we've heard that the ABC, which is notorious for propaganda, is saying that it's due to Vietnamese bots. Is AI often blamed for things that are quite natural and truthful? Do you get the blame for someone doing something? Mr Frondorf : I'm not familiar with this particular situation so can't speak to that. I would say that we take seriously our responsibility to make the benefits of AI and maximise the benefits of AI for the world. That includes providing a useful product and useful models while also having various guardrails in place to prevent harm. Senator ROBERTS: I don't want to put you in a position of talking about a competitor, but ChatGPT version 4 did what most AI does. It goes out and hoovers the internet. ChatGPT version 4o can be used in super-rational mode, meaning it goes out and only uses verified data. The responses under ChatGPT version 4o super-rational mode are quite astoundingly different from the normal ChatGPT. Do you have anything like super-rational mode within Claude? Mr Frondorf : I can't speak to this particular ChatGPT feature. I would say the way that we train Claude applies to all Claude models. That foundation comes from our constitution, and the key principles in that constitution are to approach all questions empirically. It should be factually accurate. It should be honest about where it's limited on its knowledge, but also, where there are a range of perspectives on contested topics, it should be able to put forth the best and strongest viewpoints on those topics to educate our users and promote discussion. Senator ROBERTS: What do you mean by factually accurate? Do you mean using legitimate peer-reviewed scientific papers as the basis? Is it verified data? How do you verify the data to make sure it's factually accurate? Mr Frondorf : For our purposes we are not verifying specific points of fact. We're not looking at specific topics and judging what information is in or out. Overall, the process for training a model is: first, a model critiques and revises its own responses against the principles that we have in the constitution; and then, second, the model is shown different possible responses to the same question, and it uses that to train itself to prefer the one that better aligns with those principles. So it's seeing many thousands and thousands of examples, and it learns to internalise those principles in how it generates responses. In an area like this, it is designed to pull from the data it has seen, but it is using judgement on presenting factual information in areas of consensus based on what it's seen in the training data. Senator ROBERTS: My last question is probably the most important. Some humans seek to control others through censorship, limiting opportunities or whatever. This has been shown throughout history—correct, Mr Frondorf? What I want to know is: how do you accommodate such things in your AI, in particular the Intergovernmental Panel on Climate Change, the UN's so-called climate body? I've read every report. I've read all six. Every report of the IPCC, starting from its first in 1990, has been based on a lie. How do you make sure that people don't refer back to such lies? Mr Frondorf : I'm not familiar with the specific reports, but our goal is that Claude is designed to reflect scientific evidence and reflect the majority of what it is seeing in its data. But Claude is also designed to engage with a wide variety of perspectives on climate and energy policy, and that includes sceptical ones. So, if someone is asking Claude policy questions around the energy transition or weighing economic costs, Claude should engage with that full range of perspectives. But we overall draw a distinction between factual questions or empirical questions, where we want Claude to draw from the literature, and those policy questions where we want Claude to be even-handed. Senator ROBERTS: Thank you, Mr Frondorf. I appreciate the manner in which you've answered these questions. CHAIR: I have a couple of quick questions, and then we'll wrap up. We're very, very appreciative of your time. In February 2026, Time magazine reported that Anthropic had dropped some key pillars of its responsible scaling policy, or RSP, including scrapping a promise to not release AI models if Anthropic couldn't guarantee proper risk mitigations in advance. Can you talk us through why that was reported and why you did that? Mr Frondorf : The responsible scaling policy is not my area of expertise, which is working on safeguards. The RSP is focused on a narrow set of risks, while my focus is on the usage policy, which addresses areas like the misuse and misinformation that we've been speaking about, which doesn't fall under our RSP. CHAIR: This committee has received hundreds of submissions from a whole range of stakeholders on this issue, and we have had some submissions that have been looked at in detail and that have used AI. For example, some academics have said, 'Well, this is not an appropriate reference of my work,' or 'My work's been misquoted or used out of context.' Is that a function of these models not being developed enough, or is it impossible to police this issue—not police this issue but have guardrails around, for example, individuals writing substantial submissions to a government body and then saying, 'AI is responsible for the mistakes in those submissions'? Mr Frondorf : On this question, I would say that we take this risk of AI being used in a way to perpetuate misinformation and disinformation seriously, to avoid these instances where an answer isn't factual or maybe is being used out of context. First, you have all of the training processes that we've discussed today. This is partially where our usage policy comes in that prohibits that creation of this misleading content or otherwise deceptive content. When we're detecting that, what we're doing is often not looking on a specific topic-by-topic basis but, instead, finding this by the types of requests that are being made, like someone requesting content about a person but asking for certain information to be added or embellished. Is someone asking for Claude to automate posting or create social media accounts? We do put in place detection based on those techniques or actions that the user is taking to try and catch misuse regardless of the subject matter. This is one place where it is important to have these multiple layers of defence, both to ensure accurate and factual responses at the model level and to catch cases that are against our usage policy. CHAIR: Evan, you mentioned you're moving to Australia. You've set up shop here. I understand you've appointed some key personnel now in Sydney. What guarantees can you give Australians that the use of your product is safe? Mr Frondorf : On that question I would say that, from the safeguards perspective, we have built these multifaceted layers of defence. That includes at the-model training, guardrail enforcement and monitoring levels. We take care for that information to be transparent and available to our users. Our constitution is a publicly available document that details what our intentions are for Claude's values and behaviour. When we release our models, they undergo extensive testing, and we publish those results in our system cards that are available. We also do seek to be transparent about where we have seen our systems be misused through our threat intelligence reporting. We do think transparency about AI risk and about AI capabilities is crucial to providing that comfort to users, including those in Australia. Senator ANANDA-RAJAH: Evan, Anthropic is a company that's domiciled in the US, so obviously you fall under the purview of US law. My question relates to the CLOUD Act. Is there any Australian user data that would fall under the remit, theoretically, of the CLOUD Act? Mr Frondorf : I'm not aware of the CLOUD Act or the implications here. This is a question that I would need to take on notice. Senator ANANDA-RAJAH: If you could take it on notice, that would be useful, and if you could include in your answer if there are any means for Australian user data to be protected from the CLOUD Act that would also be very helpful, if you don't mind. On the topic of oversight, I want to go back to researcher access and civic oversight. What public access does Anthropic provide so that Australian researchers have oversight over the functioning of your models? At the moment, it's Claude but there may be others in the future—for example, public content libraries, search libraries, whatever libraries. Mr Frondorf : We're supportive of external research on AI risks and AI usage. We engage constructively with researchers that study AI safety and misuse and work with a number of third parties to stress test and evaluate our models. We've also published extensively ourselves on how Claude is being used and other privacy-preserving analyses on how people use Claude and its impact on society. We're always happy to have discussions on further researcher access that balances user privacy concerns with those legitimate aims. CHAIR: Thank you, Evan, and, once again, thanks for appearing at short notice. It's a brave new world, but sometimes it does feel like a grave new world, so we appreciate the evidence you've given today. That concludes today's hearing. Thank you to all witnesses who appeared and to Hansard and Broadcasting for their assistance. The committee has agreed that responses to questions on notice should be provided by 17 March 2026. The committee now stands adjourned. Committee adjourned at 09:00