Artificial intelligence (AI) has the potential to not only bring about technological advancements but also create viruses that could pose a threat to humanity, according to a recent Senate inquiry hearing on AI. Greg Sadler, CEO of Good Ancestors Policy, warned that AI could assist in manufacturing dangerous viruses, a task typically limited to leading laboratories. He cited an example from March 2022 where an AI designed 40,000 lethal molecules in less than six hours instead of finding new drugs as intended. Another study in 2023 demonstrated how students used ChatGPT to suggest potential pandemic pathogens and provide information on how they could be made using DNA ordered online.
Sadler emphasized that these studies revealed the failure of ChatGPT’s safeguards in preventing the application from offering dangerous assistance. He expressed concern that AIs could potentially aid individuals in building bioweapons during the next term of government. In response to this threat, the U.S. government introduced an executive order in October 2023.
While Sadler has raised this issue with various Australian government departments, he has not observed any evidence of risk management measures similar to those implemented by the United States. He highlighted a significant gap between investment in safety and investment in capability when it comes to AI development.
To address biosecurity risks and other potential threats posed by AI, Soroush Pour, CEO of Harmony Intelligence—an AI safety research company—proposed establishing an AI safety institute in Australia. This institute would focus on developing technical capabilities necessary for responding effectively to such threats and would require strong regulation enforcing mandatory policies like third-party testing and safety incident reporting.
Regarding regulatory frameworks for AI safety, Sadler suggested Australia consider adopting California’s SB-1047 bill as a practical way forward. This bill places an obligation on developers to ensure their AI models are safe and do not pose risks to public safety; failure to comply may result in liability for any catastrophic harms caused by their models.
Furthermore, Sadler noted that under the Californian framework, developers must have the ability to deactivate their AIs if they become dangerous. It is important for Australia’s approach towards regulating AI safety research and development align with these principles.
The SB-1047 bill specifically targets high-powered AI models developed at significant cost by requiring them to undergo safeguarding processes and allowing third-party verification if necessary. However, concerns have been raised about its potential impact on innovation within the tech industry.
As legislation progresses through California’s lower house following its recent clearance of a key hurdle on August 15th., it remains crucial for governments worldwide—including Australia—to address these emerging challenges associated with advancing artificial intelligence technologies responsibly.