Investigations into digital platforms begin

Witness shares how Claude AI model effectively prevents harmful content

Theo Hourmouzis, general manager of Anthropic Australia and New Zealand, gave evidence at this morning’s Royal Commission hearing session about the safety measures behind Claude – the large language model developed by US company Anthropic PBC – that have effectively prevented harmful content.

He told the hearing that Anthropic’s own pre-deployment testing of its Claude Opus 4.8 model achieved a “100 per cent success rate” in preventing the generation of hateful and discriminatory content, including antisemitism — though a later real-world sample of live conversations found a small number of instances where antisemitic content did get through.

Anthropic builds AI systems that respond to user prompts with text-based assistance and is used mainly by businesses for tasks such as document summarisation, workflow automation and content creation.

It has about two million consumer users in Australia.

Hourmouzis explained that before any Claude model is released, it undergoes layered safety testing built on Anthropic’s “constitutional AI” training approach, under which the model is trained so that “harmful content [should] never being created in the first place, so it doesn’t need to be moderated in post hoc environments.”

He said the Responsible Scaling Policy and specific hate-and-discrimination and extremism policies are applied throughout training and testing.

Pressed on the Opus 4.8 model specifically, Hourmouzis confirmed that nearly 200 multi-turn evaluations conducted by Anthropic’s safeguards team related expressly to antisemitism, testing whether persistent prompting could produce prohibited content.

He told the Commission the outcome of those tests was that “the success rate of not returning any harmful content was 100 per cent.”

Despite that result, he accepted users could still find ways to generate antisemitic content.

He pointed to a sample study Anthropic ran after receiving the Commission’s notice, examining 50,000 randomly selected Claude conversations over a 14-day period in June.

Only five contained antisemitic content — an “extraordinarily low” rate  that he put at 0.01 per cent — and all five were user-initiated, with Claude rejecting or redirecting the request in four of the five cases.

Hourmouzis also revealed Anthropic had engaged the US-based Anti-Defamation League to test Claude against six rival frontier models for their propensity to produce antisemitic discrimination, with Claude ranking “number one on that list with the highest score of all of the frontier models.”

On enforcement, he described a “strike-based system” in which Claude first tries to redirect a conversation, then ends the chat if the user persists, with law enforcement notified in cases involving a “clear and imminent threat.”

He said detecting antisemitic coded language and emojis in context was “technically very straightforward” for the model.

Under questioning, Hourmouzis conceded Anthropic has not systematically tracked antisemitism trends over time, and relies on a ‘hate and discrimination’ policy rather than a standalone antisemitism policy.

Also, it has not engaged any Australian organisations in its testing, and that the safeguards team reviewing flagged Australian conversations is based overseas.

Duty of care may not protect Jews from online hate

The proposed digital duty of care is not currently drafted to protect minority groups from online attacks based on their protected characteristics, the Royal Commission on Antisemitism and Social Cohesion has heard, raising questions about whether the legislation will meaningfully shield Australian Jews from antisemitic content.

Department of Infrastructure, Transport, Regional Development, Communications, Sport, and the Arts first assistant secretary Sarah Vandenbroek confirmed the drafting instructions incorporate all elements of the Rickard review’s recommendation except the reference to attacks based on a group of people’s protected characteristics, such as race, ethnicity, religion and gender.

“Not at this stage. It is something that we can look further at,” Vandenbroek said.

She told the commission the Online Safety Act is designed to regulate what Australians can see online rather than content generated or consumed overseas, and that proving intent to seriously harm a group of Australians with a particular characteristic “would be very difficult when a lot of the hate speech online is unfortunately very generic”.

Nicholas Bender SC, counsel for Australia’s mainstream Jewish communal organisations, noted a very small proportion of antisemitic content online would be directed only at Australian Jews as opposed to Jews generally. Vandenbroek agreed.

Asked how the duty of care would protect Australian Jews in those circumstances, she pointed to regulating the features of platforms and requiring safety by design principles “so that harmful content isn’t recommended or duplicated or promoted”.

Commissioner Virginia Bell asked whether the duty of care could adequately protect Australians if imposed on platforms with no physical presence in Australia and no licensing requirement to operate here. “That is obviously a challenge,” Vandenbroek conceded.

The government has not arrived at a position on a separate recommendation calling for a definition of online hate material. Vandenbroek acknowledged the department had not sought legal advice on the definition and said she had only recently become aware, through the Commission’s work and media reporting, of the strength of feeling in the Jewish community.

“That is something that we can certainly do further work on,” she said.

Records tendered in evidence showed the Executive Council of Australian Jewry had raised the gap with the department as early as February 2024, more than two years before Vandenbroek’s testimony.

The Commission also heard the eSafety Commissioner had actively opposed the “no wrong door” recommendation designed to help individuals find the right part of government to address online harms. Vandenbroek said the eSafety Commissioner’s team “has asked us not to pursue the no wrong door approach” due to concerns about handling sensitive personal information across agencies. The department’s preferred alternative was raising awareness of the eSafety Commissioner.

Vandenbroek also confirmed a gap in the adult cyber abuse scheme. Under the child cyberbullying scheme an end user can be ordered to delete harmful content and refrain from further posting about the same person, but the adult scheme has no equivalent power to prevent repeat abuse. She indicated the department would look to address this through amendments.

Complaints doubled as platforms stall safety reforms

Complaints to the eSafety Commissioner have doubled in a single year yet there is no current consideration within government to increase funding for the office, the Royal Commission on Antisemitism and Social Cohesion has heard. This is even as social media companies continue to delay the implementation of online safety reforms through repeated rounds of consultation.

The commission heard that complaints to the eSafety Commissioner rose from 55,000 in the 12 months to June 2025 to 108,000 in the year to June 2026, based on evidence from the eSafety Commissioner yesterday.

Department of Infrastructure, Transport, Regional Development, Communications, Sport, and the Arts first assistant secretary Sarah Vandenbroek told the commission on Friday there was no current proposal to increase resourcing.

“We didn’t put a proposal through the most recent budget for resourcing relating to the duty of care, because we hadn’t designed the duty of care yet,” she said. The only resourcing concern the eSafety Commissioner had raised with the department related to litigation funding, not complaint handling.

A core team of eight or nine departmental staff had been responsible for both the government response to the independent Rickard review of the Online Safety Act and the development of the digital duty of care. The Rickard review, delivered in October 2024 with 67 recommendations, did not receive a government response until April this year, an 18-month gap Vandenbroek attributed to the complexity of the recommendations, a change of minister, a caretaker period and competing priorities.

She conceded technology platforms had repeatedly sought further rounds of consultation with the practical effect of delaying implementation.

“They have characterised their requirement for extra meetings as wanting to understand the detail,” Vandenbroek said. “The multiple rounds of consultation does, of course, slow the process down.”

An email tendered in evidence revealed Jillian Segal, Special Envoy to Combat Antisemitism had been “clearly frustrated” at a January meeting when told the digital duty of care could take up to two years.

Counsel Assisting Richard Lancaster SC took Vandenbroek to a government survey on the duty of care completed by X Corp at the end of 2025. Asked which harmful content platforms should take steps to prevent for young Australians under 18, X Corp selected “none of the above”. It gave the same answer for content affecting adults.

“No, I consider that’s disappointing,” Vandenbroek said, noting that X Corp was the only platform to take this approach and was unique in submitting unsigned correspondence.

Vandenbroek confirmed a drafter had been allocated for the legislation but acknowledged the timetable “may be optimistic”. She conceded there was “probably now a case” to ask the Government whether the planned 12-month delayed commencement could be shortened.

The commission also heard the department had not sought legal advice on several foundational questions, including the effect of social media companies’ overseas domicile on enforcement and the implications of the Australia-United States Free Trade Agreement.