eSafety Commissioner issues recommendations to limit Gen AI harms

The eSafety Commissioner, Australia’s independent regulator for online safety, has issued industry recommendations intended to prevent the weaponisation of artificial intelligence (AI).

The Commissioner, who warned that the agency already has a number of AI-generated child sexual abuse materials and deepfakes under investigation, said the new safety measures would include interventions that industry could adopt to immediately improve safety.

The measures and interventions, which are based on the expertise of domestic and international AI experts and are aimed at helping the industry stay protected from more widespread and sophisticated technology, include:

  • appropriately resourced trust and safety teams
  • age-appropriate design supported by robust age-assurance measures
  • red-teaming and violet-teaming before deployment
  • routine stress tests with diverse teams to identify potential harms
  • informed consent measures for data collection and use
  • escalation pathways to engage with law enforcement, support services or illegal content hotlines, like eSafety
  • real-time support and reporting
  • regular evaluation and third-party audits.

eSafety Commissioner, Julie Inman Grant, said that this month the agency received its first reports of sexually explicit content generated by students using this technology to bully other students. It followed “a small but growing number of distressing and increasingly realistic deepfake porn” being identified.

Grant described the danger of generative AI as “causing incalculable harm to some of our most vulnerable”.

“While our regulatory powers around online safety are the longest standing in the world, regulation can only take us so far,” she said.

“Online safety requires a coordinated, collaborative global effort by law enforcement agencies, regulators, NGOs, educators, community groups and the tech industry itself. Harnessing all the positives of generative AI, while engineering out the harms, requires a whole-society response.”

The Commissioner stressed that it has long been a concern that AI was being trained on huge datasets whose balance, quality and provenance had not been established, reinforcing stereotypes and perpetuating discrimination.

“Industry cannot ignore what’s happening now. Our collective safety and well-being are at stake if we don’t get this right,” she argued.

Grant also warned that advanced AI promised “more accurate illegal and harmful content detection”, helping to disrupt serious online abuse at pace and scale”.

Therefore a shift from “moving fast and breaking things” to a culture where “safety is not sacrificed” was crucial and relying solely on post-facto regulation could result in “a huge game of whack-a-troll”.

The tech position statement sets out eSafety’s current tools and approaches to generative AI, which include education for young people, parents and educators, reporting schemes for serious online abuse, transparency tools, and the status of mandatory codes and standards.

Commissioner Grant also stressed that recently registered codes would soon require some social media services to take proactive steps to detect and remove child sexual exploitation material and eSafety is also currently considering a revised Search Engine Code, which directly considers generative AI.

Mandatory standards are also being developed for Relevant Electronic Services and Designated Internet Services.

According to eSafety’s “Generative AI – position statement”, the term generative AI is used to describe the process of using machine learning to create digital content such as new text, images, audio, video and multimodal simulation of experiences.

The difference between generative AI and other forms of AI, such as machine learning, is that its models can create new outputs, instead of just making predictions and classifications like other machine learning systems.