Background and Aim
Generative artificial intelligence (GenAI) tools are increasingly used for healthy lifestyle support. Safe use may depend not only on access, but also on users’ AI literacy and health literacy to evaluate AI-generated information. This study examined whether AI literacy and health literacy were associated with trust, accuracy-checking, privacy concern, and future intention to use GenAI for healthy lifestyle advice among adult users.
Methods and Analysis
An international cross-sectional online survey was conducted among adults aged ≥18 years who used GenAI for healthy lifestyle advice. Measures included AI literacy (PAILQ), health literacy (BHLS), GenAI use behaviours, trust in GenAI information, accuracy-checking behaviour, privacy concern, and future use intention. Descriptive analyses, Spearman correlations, and binary logistic regression were conducted. Trust–checking profiles identified individuals with potentially uncalibrated trust (high trust, low checking).
Outcomes
The analysis included 422 participants (mean age 28.6 years; 66.1% female). Most checked accuracy of GenAI health/lifestyle advice (85.1%) and intended future use (82.0%); 58.7% trusted accuracy of GenAI advice and 69.2% reported privacy concerns. Perceived AI literacy was positively associated with trust (ρ=.327, p<.001), accuracy-checking (ρ=.222, p<.001), privacy concern (ρ=.113, p=.020), and future use intention (ρ=.362, p<.001). Health literacy was positively associated with accuracy-checking and negatively with privacy concerns. Trust-checking profile showed that 49.5% participants reported high trust and high checking, while 9.2% reported high trust but low checking, suggesting potentially uncalibrated trust.
Conclusion and Future actions
Perceived AI literacy was associated with greater trust, accuracy-checking, and intention to use GenAI for healthy lifestyle advice, while health literacy was mainly associated with accuracy-checking and privacy concern. Although most checked GenAI advice, a subgroup showed high trust with low checking, suggesting uncalibrated trust. These findings highlight the need for public education and GenAI design to support critical appraisal, verification, and appropriate use of AI health advice.

Background and Aim:
Generative AI tools such as ChatGPT are increasingly used as a source of health information. Public health research shows that communication style can influence trust, acceptability and resistance to advice. However, less is known about how the wording of AI-generated advice shapes users' responses. We tested whether two modifiable communication features (tone and choice framing) change how people respond to otherwise identical cardiovascular advice. Tone compared warm, empathic wording with clinical, neutral wording. Choice framing compared supportive language that offered options with controlling language that used "must/should" wording.

Methods and Analysis:
In a 2 × 2 online experiment, 283 university students read the same scenario about a relative considering a coronary artery calcium scan. Participants were randomly shown one of four ChatGPT-style responses containing identical medical information but differing in tone (warm vs clinical) and choice framing (supportive vs controlling). We measured evaluations of the AI response, with perceived usefulness as the primary outcome. Data were analysed using two-way ANOVAs.

Outcomes:
Tone was the strongest driver of user evaluations. Warmth increased perceived usefulness (p = .004), trust and acceptability, while making the AI advice feel less threatening and less likely to trigger resistance. Choice-supportive wording had narrower effects: it increased perceived autonomy support and felt less threatening (both p < .001), but did not consistently improve broader evaluations. The least favourable judgements occurred when controlling language was combined with a clinical tone; warmth appeared to buffer some of the adverse effects of controlling framing.

Conclusion and Future actions:
These findings suggest AI health information should be evaluated not only for accuracy, but also for how it communicates. Tone and choice-supportive wording may be candidate quality indicators for assessing public-facing AI health tools. Future research should test these effects in more diverse populations and across actual health behaviours.

Background and Aim
Australians are already using general-purpose AI assistants to check symptoms, interpret results and make everyday health decisions - research commissioned in 2025 found a majority had used such tools to check symptoms before seeing a doctor. These tools are instant and accessible, but they do not reliably reflect Australian clinical guidance or local service pathways, lack the governance expected of a health service, and can present inaccurate information with confidence. Healthdirect Australia set out to capture this opportunity - meeting people where they are - while building in clinical safeguards.

Methods and Analysis
Healthdirect first delivered an AI Virtual Assistant proof of concept on its secure infrastructure, using clinically governed content, service-navigation and triage assets and synthetic health-record data. The agentic assistant was developed through rapid prototyping and tested with a consumer panel spanning diverse ages and backgrounds. Building on this, Healthdirect is now working with OpenAI to deliver an assistant within ChatGPT, drawing on a clinical triage engine (Infermedica) and skin-assessment technology (Skin Analytics), with seamless escalation to a Healthdirect nurse.

Outcomes
The proof of concept demonstrated agentic capability across preventive health, symptom assessment, service finding and health-record interactions, with strongly positive consumer reception. Notably, the assistant could deliver targeted preventive health prompts - such as screening, immunisation and chronic-disease reminders - strengthening the link between everyday AI health conversations and population health priorities. This presentation will report early findings from the alpha release of the ChatGPT assistant - how consumers actually engage with trusted, guard-railed health guidance in the environment they already use.

Conclusion and Future actions
Surfacing trusted national health infrastructure inside consumer AI offers a scalable path to safer health guidance. Findings will inform guardrails, governance and regulatory pathways ahead of wider release.

Problem:
Stigma, misinformation and low trust continue to limit access to timely, evidence-based alcohol and other drug (AOD) information. While many people seek support online, the quality and safety of available information vary widely. Artificial intelligence (AI) presents an opportunity to improve access and personalisation; however, its use raises important public health concerns, including misinformation, inappropriate guidance, bias, and lack of clinical oversight, particularly for vulnerable populations.

What we did:
The Alcohol and Drug Foundation (ADF) developed “dib”, a large language model (LLM)-powered chatbot designed to provide accessible, evidence-based AOD information and harm reduction support. The system was developed within an ethical and governance framework, guided by a clinical advisory group.
dib integrates WHO ASSIST screening and curated ADF content to deliver personalised information, with a strong focus on safety, trust and accessibility. Governance mechanisms include clinical oversight of content and outputs, monitoring of conversational data to identify risks and misinformation, safeguards to prevent inappropriate or unsafe advice, and ongoing human review to support quality assurance and iterative improvement.
Evaluation included user testing (200+ participants) and real-world analytics to assess engagement, accuracy and accessibility.

Results:
Over eighteen months, dib has facilitated more than 47,600 conversations, with higher engagement than traditional web-based information (2:05 vs 1:07 minutes). Use is highest among young people (18–24 years), a group with known barriers to help-seeking.

Lessons:
AI-enabled public health tools require careful balancing of safety, accuracy, relevance and user experience. Improving responsiveness and engagement increases value for users but can introduce risks around misleading or inappropriate information if not actively managed. Conversely, overly restrictive controls can limit usefulness and reduce impact. This work highlights the need to pair strong governance and monitoring with rapid prototyping and iteration, using real-world data to continuously refine both content quality and the safeguards required for safe, effective deployment.

Background and Aim

Domestic and family violence (DFV) describes physical, sexual, psychological abuse, and coercive control. Nearly half (45%) of women aged > 18 years are affected. However, when a survivor discloses DFV to a chatbot, who is responsible for what happens next? In globally deployed systems, the answer is rarely specified. Australian survivors are already accessing conversational artificial intelligence (AI) through international platforms and purpose-built chatbots outside any DFV sector framework. No sector-specific guidance exists for the use of AI in DFV disclosure contexts. Available frameworks address government use of AI or privacy compliance, not survivor-facing tools. Nearly all frontline practitioners report clients experiencing technology-facilitated abuse, and the same tools promoted for survivor support are weaponised by perpetrators.

Methods and Analysis

Two studies address this gap. A feasibility study involved consultations with specialists in digital health ethics, natural language processing, and data privacy, and focus groups with frontline DFV practitioners; the findings shaped the scoping review's analytical framework, surfacing gaps in the published evidence. This scoping review maps peer-reviewed and grey literature on conversational AI in DFV. This study is the first to examine these systems through a disclosure facilitation rather than a detection lens, applying intersectional feminist and epistemic injustice frameworks.

Outcomes

Findings suggest conversational AI may offer new pathways to information, referral and early support in an overstretched sector. Existing literature focuses on detection and risk stratification, leaving post-disclosure accountability and reproductive coercion (abuse over reproductive health decision-making) unaddressed.

Conclusion and Future Actions

Safety and justice by design must be preconditions, not afterthoughts. Future action should focus on sector-specific governance for survivor-facing AI tools, including standards for data privacy, post-disclosure responsibility, equity for survivors experiencing marginalisation, and protection against perpetrator misuse. Policy reform must extend beyond preventing harm to create conditions for safe access, innovation, and enhanced accessibility.

BACKGROUND AND AIM
Rapid uptake of generative AI in health communication is underway. Clinical organisations use AI for notetaking and patient-facing materials. Public policy encourages AI literacy and adoption across government. In this context, a risk-based, human-in-the-loop approach is emerging as a norm for responsible AI implementation. However, there is limited guidance on evaluating AI use in health promotion, often deemed low-risk and exempt from ongoing evaluation. This creates methodological gaps around AI applied to health communication.

METHODS AND ANALYSIS
We explored a mixed-methods interdisciplinary approach to understanding AI risk using the novel 4-Factor Framework to Assess the Suitability of AI Applications in Health Communication. The framework integrates media studies, computational social science, and health literacy perspectives to evaluate AI models across performance, fairness, flexibility, and explainability. We applied the framework to two AI approaches for assessing health communication clarity. Custom bag-of-words and fine-tuned BERT models were trained to predict expert ratings of US public health agencies’ social media posts based on CDC clear communication guidelines. Performance was assessed against expert-labelled data; explainability through training documentation and inter-rater agreement; flexibility through model specifications; and fairness through comparative content analysis of posts receiving correct and incorrect AI labels.

OUTCOMES
Results revealed trade-offs between explainability and performance. Less transparent BERT models consistently outperformed bag-of-words models, although both achieved satisfactory performance for some communication guidelines. Findings suggest value in combining modelling approaches to fine-tune these trade-offs. Both models also showed potential bias when evaluating tailored communication, raising concerns about their risks at scale.

CONCLUSION AND FUTURE ACTIONS
We recommend combining AI modelling approaches and exploring simpler dictionary-based methods for assessing communication clarity. Our findings demonstrate the value of a mixed-methods interdisciplinary framework for evaluating applied AI in health communication and highlight limitations of relying on proprietary, closed-source AI tools in health agency communication.

Problem:
Harms from alcohol and other drug (AOD) use remain a major public health issue, with limited access to timely, evidence-based early intervention. While Screening, Brief Intervention and Referral to Treatment (SBIRT) is effective, delivery is constrained by workforce availability and service access. Artificial intelligence (AI) offers an opportunity to scale early intervention; however, concerns remain regarding safety, clinical appropriateness, bias, and governance in AI-enabled health tools.

What we did:
The Alcohol and Drug Foundation developed a prototype enhancement to its AI chatbot ‘dib’, embedding a structured motivational interviewing (MI) intervention within a validated SBIRT pathway. dib integrates WHO ASSIST screening with the FRAMES model to support behaviour change and connection to care.

A core focus of development was AI safety and governance. The model was designed within a clinical governance framework and iteratively refined through transcript-based testing, independent review by senior clinicians, and continuous assessment of safety risks.

Development prioritised:
• detection and response to high-risk scenarios (e.g. withdrawal, crisis)
• maintenance of appropriate clinical boundaries
• alignment with MI principles (empathy, autonomy, pacing)
• mitigation of risks associated with inappropriate or premature advice

Results:
The prototype demonstrates feasibility as an AI-enabled early intervention. Across development, all critical safety issues were identified and resolved, including improvements in crisis detection, help-seeking pathways and conversational safeguards. Clinical review indicates improved alignment with MI fidelity and reduced risk of unsafe or directive interactions.

Lessons:
Delivering AI-enabled public health interventions requires balancing safety, model fidelity, efficacy and user experience. Enhancing conversational quality and engagement must not compromise clinical boundaries or safety detection, while overly restrictive controls risk reducing usefulness and impact. This work highlights the need to balance strong governance and oversight with rapid prototyping and continuous improvement, using real-world data to iteratively refine both model performance and safeguards over time.

Background and Aim: Critical health literacy is increasingly positioned as a key resource for navigating and appraising online health information. This qualitative study examined how multi-lingual adults from diverse migrant backgrounds in Australia access and evaluate such information.

Methods and Analysis: Grounded in a Participatory Action Research (PAR) approach, semi-structured qualitative interviews were conducted by bilingual community co-researchers (n=11) in participants’ preferred languages. Data were analysed using a purposefully strengths-based inductive Framework analysis approach, underpinned by a critical realist epistemology.

Outcomes: Interviews were conducted with 55 participants; a majority identified as female (75%) and most were aged 18–44 years (69%). Participants were born in 17 countries outside of Australia and twenty-three (41.8%) had lived in Australia for <5 years.

We constructed two themes from the data: (1) Navigating “a jungle” and (2) “It’s always a mix of trust and scepticism”. Artificial Intelligence (AI) was identified as a sub-theme across both themes. Participants described using AI tools as adaptive supports that extended their cognitive and linguistic capabilities, enabling them to translate, simplify, summarise and verify health information encountered in increasingly complex online environments. As it pertains to critical appraisal, participants positioned AI both as a mechanism to verify information from other sources, and as an object requiring verification itself. For some participants, AI occupied an uneasy middle ground where it was perceived as useful for simplifying complex information, yet fundamentally uncertain as a trustworthy source.

Conclusion and Future Actions: These findings suggest that AI-mediated health information seeking is shaped by both the quality and clarity of the underlying online information environment and individuals' capacity to engage effectively with AI tools. They position AI literacy as an important extension of health literacy, and highlight the need for further research and co-designed interventions to support critical appraisal in AI-supported health information seeking.

Background and Aim:
Conversational artificial intelligence (AI) health coaching may extend preventive health support through on-demand lifestyle advice. However, its public health value depends on more than usability or technical accuracy. Conversational AI systems collect sensitive personal information, simulate relational interaction, and may blur boundaries between health advice, self-management support, and clinical care. This study explored community perspectives on trust, privacy, data governance, and the role of conversational AI within existing health services.

Methods and Analysis:
Using a qualitative interpretive design, we conducted a secondary analysis of data collected during a co-design study developing a conversational AI health coach. Data were drawn from eight workshops involving 24 adults living with or at risk of chronic conditions, with four workshop designs repeated across two independent participant groups. Participants completed ten structured activities exploring experiences, priorities, and concerns relating to conversational AI health coaching. Reflexive thematic analysis was used to identify themes relating to trust, privacy, governance, and health service integration from workshop transcripts, participant-generated materials, and polling outputs.

Outcomes:
Five themes were identified to inform conversational AI governance and implementation: credibility and trustworthiness, particularly user-interface design cues; privacy, data control, and reversibility, including processes for withdrawing personal data from AI systems; boundaries and interfaces between conversational AI and healthcare, including who was responsible for advice that is provided; conditions for trusted engagement; and equitable access to conversational AI-supported care.

Conclusion and Future actions:
Conversational AI health coaching may expand access to preventive support, but only if credibility, privacy, data control, and accountability are embedded into systems from the outset. Future public health action should establish clear standards for evidence, consent withdrawal, human escalation, health service integration, and equity monitoring to safeguard communities and build public trust.

Background and Aim
Young Australians have taken up vaping in recent years, yet when they want to quit, there is still a gap in national, youth-specific, co-designed digital vaping cessation support. Cancer Council NSW, supported by nib foundation, is developing a multi-channel digital platform combining evidence-based guidance with conversational AI to deliver personalised, confidential vaping cessation support for young people aged 14-25.
Methods and Analysis
Scoping engaged over 150 young people and 28 stakeholder groups through co-design workshops, surveys, consultations, and a rapid evidence review. Young people emphasised the need for empathetic, non-judgmental, personalised support, rather than fear-based messaging. Anonymity and confidentiality were especially important for those aged 14 to 17. A proof of concept validated the AI approach. Phase 1 delivers a website and self-guided cessation modules. Phase 2 introduces an AI-enabled native mobile app using Retrieval-Augmented Generation (RAG), grounded in a curated evidence base of Australian clinical guidelines and behavioural science, to deliver adaptive, context-aware support including goal-setting, stress-management tools, nudges, and referral pathways to Quitline. Ethical AI guardrails, moderation processes, and content boundaries guide responsible deployment.
Outcomes
Phase 1 launches late August 2026, with Phase 2 piloting from late September 2026. Co-design has shaped the platform's tone, content, and user experience, supported by a curated evidence registry that underpins the AI knowledge base. Early data insights on reach, engagement, and user experience, as well as lessons on co-designing and safely deploying AI in youth preventive health will be presented.
Conclusion and Future Actions
This project offers a replicable model for integrating AI into preventive public health, grounded in co-design and safety-by-design. The platform empowers young people to access the right support at the right time, positioned as an early intervention with pathways to clinical care. Future priorities include evaluating cessation outcomes and strengthening equity for priority populations.

Background and Aim
Dengue has become a major public health threat, escalating from seasonal outbreaks to persistent endemicity across Oceania. Despite climate change intensifying transmission risk, governance remains reactive. A critical gap exists in harnessing emerging technologies to translate localised climate and health data into timely early warning and targeted action. This study identifies climate drivers of dengue transmission in two ecologically distinct sites in Indonesia, establishing foundations for an AI-integrated, data-driven early warning system (EWS).

Methods and Analysis
A Distributed Lag Non-linear Model (DLNM) was applied to nine years of dengue case data (2015–2023) alongside temperature, humidity, and rainfall data from Palu City and Central Buton to estimate non-linear exposure–lag–response relationships, quantify outbreak probability, and validate predictive performance.

Outcomes
Climate drivers are site-specific. In Palu City, temperatures of 25–26°C with high humidity increase dengue risk by 14–30% over a 0–3 month lag. In Central Buton, extreme heat followed by elevated rainfall drives risk with a 2–3 month delay. Outbreak detection coverage exceeded 80% at both sites, confirming a meaningful early warning window. Findings reinforce that one-size-fits-all strategies are insufficient across ecologically heterogeneous landscapes.

Conclusion and Future Actions
The 2–3 month early warning window offers opportunity for transformative AI integration to reduce population risk to negative health outcomes. Machine learning applied to real-time climate monitoring and surveillance efforts could sharpen EWS accuracy beyond DLNM. Federated learning architectures provide a pathway to scale locally-calibrated models, which is critical for the heterogeneous conditions of Indonesia’s vast archipelago. AI-assisted education campaigns could deliver culturally tailored health messaging to build community health literacy during high-risk scenarios. AI tools can support healthcare workers to personalise risk communication while preserving the trusted human relationship that drives uptake. This research positions climate data-driven EWS as a foundation for broader, AI-augmented, equity-conscious approaches to anticipatory infectious disease governance.

Background:
Artificial intelligence (AI) and digital technologies are rapidly transforming healthcare, reshaping diagnosis, decision-making, and patient engagement. However, digital health literacy, encompassing both individual capability and the accessibility, and complexity of the digital health environment, is quickly becoming a critical determinant of health outcomes. In Australia, over 60% of the population has low health literacy, with poorer digital capability concentrated among already disadvantaged groups. The rapid uptake of generative AI, including consumer use of tools such as ChatGPT for health, risks further widening inequities unless health literacy principles are embedded in AI development and implementation.

Methods:
This NHMRC Synergy program brings together multidisciplinary experts across health, social, and computer sciences to address AI health literacy across the full technology lifecycle. The research adopts a four-stage pipeline: (0) public engagement and horizon scanning, (1) health literacy-informed AI development, (2) capacity building for consumers and the health workforce, and (3) deployment of AI to protect population health. Methods include citizens’ juries, co-design with priority populations, experimental trials, implementation studies, and AI-enabled monitoring systems. Partnerships with government, industry, and community organisations will ensure relevance and translation.

Expected Outcomes:
The program will deliver a suite of co-designed, evidence-based interventions, including AI-assisted communication tools, multilingual digital health resources, workforce training, and insights for regulatory monitoring systems. It will generate new evidence on integrating health literacy into AI design, improve accessibility and safety of digital health tools, and strengthen capabilities among consumers, clinicians, and developers.

Conclusions and Future Actions:
This program positions health literacy as foundational to equitable AI-enabled healthcare. By embedding inclusivity, trust, and evidence into digital innovation, it aims to ensure all Australians can benefit from the digital health revolution and reduce widening health inequities.