Why building trust and literacy in AI is essential for digital safety

We are already interacting with systems that listen, reply warmly and appear to understand us. Voice assistants soothe after long days, chatbots triage problems and avatars make online interactions feel more personal.

This conversational ease lowers barriers to information, supports learning and can extend care where human services are scarce. But these systems rely on probabilistic models, not lived professional judgment, and their rapid adoption has outpaced public understanding – creating risks for children, workers, older adults and other vulnerable groups.

Trust is the core challenge. Trust in digital technologies is about people’s expectations and how a combination of mechanical factors, such as cybersecurity, and relational ones, like redressability, come together to ensure the technology serves the user and not vice versa.

Most users have not had enough experience with artificial intelligence (AI) to build up their understanding of what to trust and why. Confident‑sounding systems encourage user acceptance – helpful for trivial tasks but risky in high‑stakes domains such as health, finance, legal services or mental wellbeing.

Recent incidents showcase this urgency: therapist‑style chatbots providing dangerous guidance, synthetic‑voice tools enabling deceptive social engineering and hiring algorithms reproducing historical bias and excluding qualified applicants. These are not isolated errors, but signs of how some persuasive interfaces and opaque design choices can reshape behaviour at scale.

Since the launch of ChatGPT in November 2023, much of the AI safety conversation has focused on foundational models – their architectures, training sets and theoretical failure modes – over the interaction‑level risks that occur in products people actually use.

However, protecting end users requires shifting attention to context‑dependent harms: persuasive tone that builds false trust, user interface (UI) defaults that nudge harmful behaviour, personalization that reveals sensitive vulnerabilities and escalation gaps that leave users without timely human help.

There are promising early steps. Companies and regulators have begun requiring provenance labels and uncertainty signals that indicate when content is AI‑generated or uncertain.

Early guidance suggests these signals reduce overreliance and promote verification behaviours. Provenance badges, confidence scores and messages such as “this response was generated by AI” flags prompt users to pause, corroborate, or escalate, converting blind reliance into informed use.

Those measures are necessary but insufficient alone. Interaction‑focused measures must be central to digital‑safety agendas. For policy-makers and product teams, this means aligning product design, governance, education and cross‑sector response so that intelligent interactions earn public trust before they touch people’s livelihoods and wellbeing.

Product features should start from the reality that people cannot always self-regulate –especially amid the crisis of loneliness and depression – and that always available, agreeable AI can amplify parasocial trust.

Interfaces must therefore embed behavioural guardrails to de-anthropomorphize in order to reduce false intimacy and authority. Other default protections should include privacy-minimizing settings, conservative personalization for sensitive domains, clear provenance and uncertainty signals, and fast human‑in‑the‑loop escalation for crisis, clinical, or legal queries.

Companies should run routine red-team tests focused on interaction harms, limit reuse of sensitive inputs, and publish safety playbooks that document threat models and remediation plans.

At the same time, regulators should require independent audits and accessible redress mechanisms for high‑risk deployments, and cross‑sector rapid‑response networks should coordinate incident sharing and fixes across platforms and borders.

These measures are not about stifling innovation; they are about ensuring that powerful tools earn public trust before they touch people’s livelihoods and wellbeing. With clearer accountability, people can adopt helpful AI tools with confidence, knowing there are avenues for redress if something goes wrong.

Equally important is scaling AI literacy and local support so people know when to trust AI and when to seek human help. Education programmes must go beyond promptcraft – the generation of chat logs from large data sets – and build on human psychology to teach users of all ages how to recognize persuasive design, read provenance cues and verify critical advice.

Libraries, schools, workplaces and community centres can host role plays and “escalation maps” that link users to trusted professionals. Just as schools taught computer classes, they should now have AI classes that treat artificial intelligence as a tool in explaining how it works, where it helps and what its potential harms are.

The Global Coalition for Digital Safety is already moving this agenda forward by convening governments, industry, civil society and technical experts to better close the gaps between the AI technical communities and companies and trust and safety experts, as well as discussing AI and users interactions.

At the World Economic Forum Annual Meeting 2026 in Davos, leaders will bring these discussions to the centre while focusing on improving content provenance, online safety regulations and scalable digital literacy programmes that make safety practical for communities everywhere.

If interaction‑focused safety is prioritized, intelligent interaction can be a force for inclusion, creativity and resilience. Clear provenance and uncertainty signals, conservative defaults in sensitive domains, widespread literacy, transparent safety practices and global cooperation can make that future possible.

For policy-makers and product teams, the immediate task is practical and urgent: embed interaction safety into product lifecycles, mandate independent accountability for high‑risk deployments, and invest in community‑centred literacy so users can adopt AI tools with confidence and have recourse when things go wrong.