UQ Reveals Australia's AI Safety Gap is 4000 Times Bigger Than Thought

Unveiling Australia's AI Safety Gap: UQ's Groundbreaking Survey Findings

The University of Queensland (UQ) has dropped a bombshell in the world of artificial intelligence (AI) research with its latest Survey Assessing Risks from Artificial Intelligence (SARA). This nationally representative poll of 933 Australians paints a picture of a massive disconnect between what the public expects from AI systems and the reality assessed by experts. At its core, the Australia's AI safety gap refers to the chasm between perceived safety levels and actual risk probabilities, particularly when benchmarked against ultra-reliable sectors like commercial aviation.

AI safety, in this context, encompasses measures to ensure advanced AI systems—those capable of outperforming humans in economically valuable tasks—do not cause unintended harm, from everyday biases to catastrophic existential threats. UQ researchers, led by Associate Professor Michael Noetel from the School of Psychology, reveal that 94% of respondents believe AI should match or surpass aviation safety standards. Yet, expert evaluations suggest current AI development lags by at least 4,000 times, with some estimates reaching 30,000 times higher risk.

This revelation isn't just alarming; it's a call to action for Australian universities, where much of the foundational AI safety research is happening. As higher education institutions like UQ pioneer these studies, they position themselves at the forefront of a field poised to shape national policy and global standards.

Public Perceptions: Demanding Airline-Grade Reliability from AI

Commercial aviation boasts an enviable safety record: roughly a 1-in-30-million chance of death per flight, translating to about 150 global fatalities annually despite billions of passengers. Australians, familiar with the Civil Aviation Safety Authority (CASA)'s stringent oversight, project similar rigor onto AI. The SARA survey underscores this, with nearly all participants (94%) insisting on equivalent or better safeguards for AI deployments in healthcare, autonomous vehicles, and decision-making tools.

This expectation stems from cultural trust in regulated industries. Step-by-step, aviation safety involves rigorous certification, mandatory incident reporting, independent audits, and continuous improvements based on data. Public surveys show Australians want the same for AI: pre-deployment testing, transparency in failures, and regulatory bodies akin to CASA. However, without such frameworks, trust erodes—only 36% currently trust AI, per related polls.

For universities, this highlights a teaching opportunity. Programs in AI ethics and safety at institutions like UQ are equipping students to bridge this perceptual gap through evidence-based education.

Expert Assessments: The Stark Reality of Current AI Risks

AI experts, drawing from global benchmarks like the AI Index and expert elicitations, peg current frontier models' safety at orders of magnitude riskier. Median estimates place existential catastrophe risks from misaligned superintelligent AI at 5% by 2100, but per-use failure rates for critical systems hover around 10^-3 to 10^-4—far from aviation's 10^-8. The 4,000-fold gap arises because AI lacks aviation-style verification: no standardized testing suites for emergent capabilities like deception or goal misgeneralization exist at scale.

Take real-world cases: AI systems in hiring have amplified biases, costing Australian firms millions in lawsuits, while autonomous vehicle trials reveal edge-case fatalities. UQ's SARA quantifies this mismatch quantitatively, urging a shift from optimism to rigorous evaluation.

Higher ed implication: Demand surges for PhD researchers in AI alignment at UQ and peers. Explore research assistant jobs in AI safety to contribute.

Behind the Study: UQ's Michael Noetel and Alexander Saeri Lead the Charge

Associate Professor Michael Noetel and Dr. Alexander K. Saeri, both from UQ's School of Psychology, helm the SARA series. Noetel specializes in mapping AI hazards via the MIT AI Risk Repository, while Saeri focuses on sociotechnical transitions and policy. Their methodology: nationally representative online surveys tracking longitudinal trends since 2020.

Previous iterations (2024: 1,141 respondents) showed 80% viewing AI catastrophe risks as pandemic-level priorities. The 2025/2026 update refines with safety benchmarks, revealing declining support for unchecked development (13% drop). Their work exemplifies how Australian universities drive empirical AI governance research.

Rate professors like Noetel on Rate My Professor to see student impacts.

AI Safety Research Across Australian Higher Education

UQ isn't alone. The University of Technology Sydney (UTS) Human Technology Institute praises government initiatives, while the University of Sydney launches the 2026 AI Safety Fellowship—a 10-week program for aspiring safety researchers. Adelaide's Australian Institute for Machine Learning (AIML) tackles robustness, and Wollongong secures AEA grants for AI commercialization with safety nets.

• Sydney AI Safety Fellowship: Hands-on governance training.
• UTS HTI: Interdisciplinary human-AI risk studies.
• AIML: Robust ML for real-world deployment.

These initiatives create ecosystems for PhDs and postdocs. Check Australian university jobs for openings.

Government Steps Up: Australian AI Safety Institute (AISI) in 2026

In response to calls like SARA's, the Albanese Government funds AISI with $29.9M, operational early 2026 under Industry, Science and Resources. It will test frontier models, assess risks, and recommend reforms—mirroring UK/US institutes. Universities will partner, providing expertise and talent pipelines.

AISI addresses the voluntary-to-mandatory shift, focusing on transparency absent in current frameworks.

Implications for Higher Education: Careers and Curriculum Shifts

The AI safety gap boosts demand for university talent. Roles in alignment research, risk auditing, and policy explode. UQ's programs train the next wave, with actionable advice: Master Python/ML basics, then specialize in verification techniques.

Read how to excel as a research assistant or pursue postdoc positions. Institutions adapt curricula, embedding safety modules.

Stakeholder Views: From Industry to Policymakers

Industry leaders welcome AISI for competitiveness; 90% of SARA respondents favor safeguards boosting trust. Policymakers note skills gaps alongside tech. Universities Australia hails the National AI Plan for productivity-safety balance.

Balanced perspectives: Optimists see innovation; skeptics urge pauses on frontier models.

Recommendations: Closing the Gap Through Actionable Steps

SARA proposes:

• Mandatory testing for high-risk AI.
• Independent audits like aviation.
• Incident reporting databases.
• Whistleblower protections.

Photo by Datingjungle on Unsplash

Future Outlook: Universities Shaping Australia's AI Destiny

By 2030, AI could add $15T globally; Australia risks falling behind without safety. UQ et al. position higher ed as leaders. Actionable insights: Pursue AI safety courses, engage in fellowships, advocate policy.

Discover opportunities at higher ed jobs, university jobs, career advice, and rate my professor. Australia’s universities are key to safe prosperity.