1 of 11

Responsible AI in Practice�Embedding Ethics and Transparency in AI

2 of 11

Jose Meza

AI Ethics Researcher & Academic

Lecturer at The University of Sydney, Researcher, and Affiliate at All Tech Is Human; Contributor to IEEE Working Groups, and ForHumanity AI, Algorithmic, and Autonomous Systems Auditor.

Women in AI Australia

Angelica Chowdhury

Data Scientist & AI Safety Researcher

Current Supervisory program for Alignment Research Fellow

Incoming Mentorship for Alignment Research Student Fellow and Future Impact Group Fellow

3 of 11

4

A lot has happened in the last few months locally

The Responsible AI Index

25 August 2025

Source: https://www.industry.gov.au/news/australias-national-benchmark-responsible-ai-adoption-now-available

4 of 11

4

A lot has happened in the last few months locally

Guidance for AI Adoption

21 October 2025

Guidance for AI Adoption

21 October 2025

Source: https://www.industry.gov.au/publications/guidance-for-ai-adoption

5 of 11

4

A lot has happened in the last few months locally

Announcing Australia AI Safety Institute (AISI)

25 November 2025

Source: https://www.industry.gov.au/news/australia-establishes-new-institute-strengthen-ai-safety

The AISI will provide trusted, expert capability to monitor, test and share information on emerging AI technologies, risks and harms. By identifying future risks, it will help the government ensure fit-for-purpose protections for Australians.

The AISI will be an important capability in government, working directly with regulators to make sure we’re ready to safely capture the benefits of AI with confidence.

Its work will include:

helping government keep pace with rapid developments in AI technologies, dynamically addressing emerging risks and harms
enhancing our understanding of technical developments in advanced AI and potential impacts
serving as a central hub to share insights and support coordinated government action
giving guidance on AI opportunity, risk and safety for businesses, government and the public through established channels including the National AI Centre (NAIC)
supporting Australia’s commitments under international AI safety agreements.
Its work will complement existing legal and regulatory frameworks that already protect Australians’ rights and safety.
The AISI will become operational in early 2026.

By working with the National AI Centre and international partners, the AISI will help position Australia as a leader in global efforts to address AI safety. It will join the International Network of AI Safety Institutes, leveraging world-class safety testing expertise from leading AI nations.

6 of 11

4

And where we contributed globally

International�AI Safety Report

29 January 2025

Source: https://internationalaisafetyreport.org/sites/default/files/2025-10/international_ai_safety_report_2025_english.pdf

This is the first comprehensive international review of scientific research on general-purpose AI capabilities and risks, written by over 100 independent experts led by Turing Award winner Yoshua Bengio, and backed by 30 countries and international organizations

Who Contributed?:

Bronwyn Fox, the University of New South Wales
Australia’s Temporary AI Expert Group
Gradient Institute
Harmony Intelligence

What this report is about?

What can general-purpose AI do? (Current and future capabilities)
What are the risks? (From malicious use, malfunctions, and systemic impacts)
How can we manage these risks? (Technical approaches and their limitations)

It is not about:

It does NOT recommend specific policies - it provides scientific information to support informed policymaking
It focuses on risks, not benefits - while acknowledging AI has vast potential benefits, the report specifically examines safety concerns so those benefits can be realized safely
It covers current harms AND potential future risks - from problems already happening (like deepfakes and bias) to emerging concerns (like AI-enabled cyberattacks and labor market disruption)

7 of 11

4

And where we contributed globally

Research agenda on risks from synthetic content

14 July 2025

Source: https://www.industry.gov.au/publications/research-agenda-risks-synthetic-content#download-a-copy-of-the-research-agenda

Synthetic AI-generated content is exploding across text, audio, image and video, posing risks through harmful content like CSAM (Child Sexual Abuse Material) and NCII (Non-Consensual Intimate Imagery), enabling fraud and impersonation, and undermining trust in digital information by making it difficult to distinguish real from AI-generated content

Currently no silver bullet. Even the strongest watermarks can be attacked, detection tools are criticized for being reactive and inaccurate, and there is no consensus in the research community on how well specific techniques work for their intended purpose

This research agenda on risks from AI-generated content is co-led by Australia and Canada as part of our participation in the International Network of AI Safety Institutes.

It is supported by and will be used as a reference in consideration of work planning and prioritization by, Australia, Canada, France, the European Commission, Japan, Kenya, Korea, Singapore, and the United Kingdom.

Digital Content Transparency (DCT) Techniques: DCT techniques help distinguish synthetic from non-synthetic content

Four Priority Research Areas

1. Evaluating and Improving Built-in Safeguards

Preventing harmful content generation (especially NCII)
Testing how fine-tuning affects safety measures
Improving safeguards for multimodal AI systems

2. Evaluating Current DCT Techniques

Mapping content authentication standards and their vulnerabilities
Testing watermark resilience against removal and tampering
Developing better evaluation benchmarks

3. Understanding Systemic Impacts

How synthetic content affects public trust in information
How people actually use transparency tools
Impact on global information ecosystems, especially outside wealthy nations
Challenges in validating real content versus AI-generated

4. Advancing Technical Approaches

Developing text watermarks that survive translation between languages
Creating better benchmarks for detection methods across languages
Improving multimodal detection using broader contextual signals
Understanding how humans work with AI detection tools
Developing novel techniques to overcome current weaknesses

8 of 11

But Is this Enough?

Because existing risks are looking like these�

And emerging risks could look like these�

Broad AI adoption, but limited labour market effects

More AI agents, less human oversight

Stellar benchmark performance, limited practical utility

Biological and Cyber threats

AI knows it is being tested (Sandbagging)

Gradual disempowerment

Source: https://internationalaisafetyreport.org/publication/first-key-update-capabilities-and-risk-implications

one study found that reasoning models cannot solve problems above certain complexity levels, even when given adequate computational resources at inference time. This suggests that these models’ success may rely on sophisticated pattern-matching rather than ‘true’ reasoning

Though many workers have begun to use AI, the labour market impacts of AI systems remain limited. Evidence points to some workplace adoption and minimal aggregate employment disruption to date, though some targeted impacts on specific demographics have been documented.

noise, missing entries, or demographic differences.

It is difficult to understand how accurate and useful the evaluations used to assess AI models are. For example, data contamination – the inclusion of evaluation questions in training data – can inflate AI models’ evaluation scores.Most evaluations are conducted only in English, which limits conclusions about AI models’ global performance and may overestimate their capabilities in languages other than English.

One study showed that current language models can troubleshoot virology lab protocols better than 94% of tested subject experts, drawing on knowledge considered rare by virologists.¹⁰³ Such advice could assist both experts and novices, and many current safeguards can be bypassed, such as if the user claims that they need the information for legitimate research

Europol reports the rise of malicious LLMs on both surface and dark Web, lowering entry barriers for criminal offenders.

Some research shows that AI systems may be able to detect when they are in an evaluation setting and alter their behaviour accordingly. Studies have documented models producing outputs that can mislead evaluators and showing an ability to distinguish between evaluation and deployment contexts. A small number of demonstrations have shown that, under certain conditions, AI models can produce outputs that could systematically mislead evaluators, such as underperforming in assessment contexts.This could make it more difficult to assess their true capabilities

Previous technological disruptions (like the Industrial Revolution) displaced specific jobs but created new ones, and humans remained essential to the economy. With sufficiently advanced AI, we might reach a point where AI is better than humans at everything economically valuable. At that point, the systems that currently serve human interests (because they need humans) would have no structural reason to continue doing so - even if individual AIs are "aligned" and doing what they're told. It's essentially the risk that we voluntarily make ourselves obsolete, one incremental improvement at a time.

9 of 11

4

What can be done to

best prepare and keep up?�

Invest in independent testing capabilities beyond current voluntary standards.
More research into realistic evaluation methodologies
Keep each other informed
Keep the responsible bodies accountable through advocacy
Use AI, but use your own judgement

Source: https://internationalaisafetyreport.org/publication/first-key-update-capabilities-and-risk-implications

10 of 11

What can be done to

best prepare and keep up?�

https://www.innovationaus.com/new-deepfake-bill-to-protect-individuals-faces-voices/

11 of 11

Also Keep an Eye on WAI’s Workshops in 2026 to be Informed

where we will cover these in detail and more