The first step is to understand the context. South Africa faces significant digital inequality and limited digital capacity. The reality is that less than 20 percent of South African Internet users report that they use e-government services. This means when setting up an AI chatbot, we need to consider that online interaction with the government is a relatively a new practice in the country.

For the chatbot to be effective, it must be as effortless as possible for users to access and use. This means avoiding any additional steps, such as requiring citizens to download a new software. Since WhatsApp is already widely used by the target audience (93 percent of active social media users), it offers a convenient and familiar platform for hosting the chatbot. However, this choice also introduces certain risks, which will be discussed later.

Transparency begins with clear disclosure. Users should be informed upfront that they are communicating with an AI chatbot, not a human agent. This disclosure should be unambiguous, visible at the start of the interaction, and repeated when necessary —especially in complex or sensitive contexts, such as legal, financial, or health-related services. The Terms of Use document, while necessary, tends to be complicated and should further be optimized to ensure clear disclosure.

Perhaps most critically, users should be clearly informed about what data is being collected, how it will be used, and who will have access to it. This includes metadata, message content, and any personal identifiers provided during the conversation. For instance, the South-African citizen chatbot only stores de-identified data, which is relevant to its purpose.

In principle,, national data protection regulations ought to be respected. This also means adopting a data minimization approach: only collect what is strictly necessary to deliver the relevant service. Collected data must be stored securely, encrypted, and anonymized wherever possible to prevent unauthorized access or re-identification.

There should also always be a clearly designated authority accountable for the chatbot’s development, deployment, maintenance, and oversight. Citizens should be able to easily find out which authority is in charge and what standards the chatbot is expected to meet – technically, legally and ethically.

Without these measures, public confidence and trust in public institutions and their digital service delivery can quickly erode. It can also be worsened by the the lack of transparency or obfuscation that can fuel fears of surveillance, discrimination, or misuse. By contrast, a transparent approach reinforces the idea that digital tools serve the public interest, not just technological ambition.

Citizens should receive clear and measurable benefits in exchange for sharing their personal data. For example, a South African citizen uses the chatbot to report a broken water pipe in their village. Such a report should trigger a specific administrative action – in a visible and timely manner. Trust is strengthened and user engagement increases when citizens see their feedback is responded to time with clear follow up actions, such as providing an update what steps will be taken to repair the water pipe.

To ensure trust, governments must also take precautionary measures to prevent misuse of the chatbot. This includes implementing strong authentication processes to block malicious bots and ensuring the chatbot limits interactions to its designated purpose. By automatically filtering out attempts to engage on unrelated or harmful topics, governments can protect both users and the integrity of the tool.

AI chatbots come with a unique set of risks that go beyond traditional software systems, and governments must approach these with caution and foresight. One major concern is algorithmic bias — the tendency of AI systems to reflect and amplify existing inequalities in the data they were trained on. This can result in unequal treatment of certain groups, particularly in contexts where language, cultural nuances, or access to services play a role. It may also lead to inconsistencies in how public services and information are communicated and delivered, disproportionately affecting those who speak underrepresented languages or dialects, practice diverse cultures, or have historically faced barriers to accessing public services.

Another well-documented issue is hallucination, where AI models generate confident but false or misleading responses — potentially having negative implications when citizens interact with government chatbots. This risks spreading misinformation, causing confusion or even harm. Citizens may lose trust or be discouraged from using public services. Over time, this undermines credibility, accessibility, and public participation.

Compounding these risks is the black-box nature of many AI models: they produce outputs without offering clear reasoning, making it nearly impossible to trace how or why a particular response was generated. This threatens due process, as individuals can't know how or why decisions were made. It limits appealability, since there's no clear basis for challenging outcomes. Lack of transparency also weakens documentation and accountability in public service.

Lastly, over-reliance on AI can lead to technology dependency, where human expertise is devalued and essential public services become fragile in the face of outages, system errors, or model drift.

Therefore, the question of proportionality must be asked here: considering all the risks AI poses as well as the significant environmental costs associated with it, does the benefit truly outweigh the cost for the problem it's trying to solve?