AI can help us predict, analyze, and respond faster. But it cannot take responsibility for the decisions it informs.
That distinction matters in public health.
As health systems face growing demands and limited resources, AI and digital tools are increasingly being explored to support everything from appointment reminders and commodity forecasting to data analysis and service delivery.
The opportunities are real. So are the risks.
When these tools are introduced into programmes serving people who may already face stigma, discrimination, or barriers to care, a poorly designed system can reinforce the very inequities we are trying to address. Biased data can produce biased outputs. Weak data protection can compromise confidentiality. And a system that is difficult to understand or challenge can erode trust.
These were some of the questions at the heart of the “Responsible AI and Digital Tools for Improved HIV Outcomes” workshop led by Rhoda Robinson at AIDS 2026 in Rio de Janeiro.
The conversation was not about whether AI belongs in public health. It was about how we make sure it belongs there responsibly.
Start With the Risk, Not the Technology
Not every use of AI carries the same level of risk.
Using a tool to summarize an internal report is fundamentally different from using one to support clinical decision-making. Yet organizations can easily fall into the trap of creating one broad AI policy for everything.
A more useful approach is to assess the risk of each use case before deciding what safeguards are needed.
Low-risk uses: Internal administration
This could include summarizing technical documents, organizing knowledge resources, or supporting operational communications.
These applications still require human verification, particularly where accuracy, sources, or sensitive information are involved. But the consequences of an error are generally more limited.
Moderate-risk uses: Programme-facing tools
Appointment reminders, public health chatbots, and tools that analyze community feedback interact more directly with people.
Here, safeguards need to be stronger. Organizations should consider informed consent, data protection, accessibility, clear communication about automated systems, and regular human review.
High-risk uses: Clinical and consequential decisions
Automated triage, risk scoring, or diagnostic recommendations can directly affect someone’s health and access to care.
This is where the principle of human accountability becomes critical. AI may support a decision, but responsibility cannot be handed over to a machine.
The higher the potential harm, the stronger the human oversight must be.
From Ethical Principles to Practical Safeguards
Global guidance on responsible AI is growing. The challenge for health organisations is turning those principles into decisions that programme teams can actually make.
At HACEY, we have been thinking about this through the SAFEGUARD Framework, which brings together practical questions that should be asked before and after an AI tool is introduced.
Scope & Problem: What problem are we actually trying to solve? Is AI the right solution?
Appropriate Data: Is the data relevant, representative, ethically collected, and appropriate for the intended use?
Fairness & Inclusion: Could the tool disadvantage people because of gender, age, language, disability, literacy, connectivity, or other factors?
Errors & Safety: What could go wrong? What happens when the system produces an incorrect or misleading output?
Guardianship of Data: How will sensitive information be protected? Non-public client information should never be entered into open AI systems without appropriate safeguards.
Understandable: Can users understand when and how an automated system is being used?
Accountable Humans: Who is responsible for monitoring the system and making decisions when something goes wrong?
Review Continuously: How will accuracy, equity, user experience, trust, and community feedback be monitored after deployment?
Discontinue if Harmful: What are the conditions for pausing or stopping a tool if it creates unintended harm?
The important point is that responsible AI is not a checklist completed once at the beginning of a project. It is an ongoing process of testing, learning, monitoring, and being willing to change course.
What This Means for HIV Programmes
The potential applications are significant.
Digital tools can support earlier detection, help people stay connected to appointments and services, strengthen health information systems, improve data quality, and support more personalized care.
But improving efficiency cannot be the only measure of success.
We also need to ask:
- Does this tool make care easier to access?
- Does it work for the people most likely to be left out?
- Does it protect confidentiality?
- Can a person question or challenge the outcome?
- Who is accountable when the technology gets it wrong?
These questions are particularly important in HIV programming, where trust, confidentiality, stigma, and continuity of care can directly influence whether people seek and remain in services.
Technology Should Strengthen Trust, Not Replace It
There is a temptation to think of AI as a way to do more with less. But public health is not simply a problem of efficiency.
It is also a problem of trust.
People need to know that their information will be handled responsibly. Healthcare workers need to understand the systems they are being asked to use. Programme managers need to know when a tool is reliable and when it should not be trusted. And communities need a voice in decisions about technologies that affect them.
The question, then, is not whether public health should embrace AI.
We should.
The more important question is what kind of AI-enabled health systems are we willing to build?
If we get the governance right, AI can help health systems extend their reach, use resources more effectively, and improve the experience of care.
If we get it wrong, we risk digitising existing inequalities and giving them greater scale.
Responsible innovation means being ambitious about what technology can do while being equally serious about what it must never compromise: privacy, equity, dignity, and human accountability.
That is the ethical frontier we need to focus on now.
Resources for Public Health Practitioners
The conversation cannot stop at principles. Practitioners need practical resources they can use when assessing, designing, and implementing digital and AI-enabled health interventions.
We are sharing the following resources from the workshop:
- WHO Guidance on Ethics and Governance of Artificial Intelligence for Health: a foundational resource covering ethical and governance considerations for AI in health.
- WHO Guidance on Large Multi-Modal Models: practical guidance for considering the governance and responsible use of generative AI and large multi-modal models in health.
- UNAIDS/PEPFAR Privacy, Confidentiality and Security Assessment Tool: a resource for assessing how personal health information is collected, stored, and used and whether appropriate privacy and security measures are in place.
- NIST AI Risk Management Framework: a practical framework for organizations developing, deploying, or using AI to identify and manage risks and promote trustworthy AI.
- UNAIDS Global AIDS Strategy 2026–2031: a framework for the future of the HIV response, with a strong focus on country leadership, reducing inequalities, rights, and community leadership.
- KoboToolbox and KoboCollect: practical tools for secure, offline-capable field data collection, particularly relevant in low-connectivity settings.
These resources are intended to support programme managers, healthcare workers, policymakers, educators, researchers, and organisations thinking about how to use AI and digital tools without losing sight of the people behind the data.
Technology should make public health stronger. It should never make it less human.
For the complete AIDS 2026 participant pack, workshop canvases, and interactive community resources, visit the HACEY resource hub: https://tinyurl.com/hivworkshop2026
