Artificial intelligence can help organize fragmented information, but tourist health coordination is not a safe place for vague automation. Travelers may be stressed, languages may differ, provider information may be incomplete, and the next action may depend on a clinician, dispatcher, payer, public authority, or authorized representative. An AI system must not blur those responsibilities.

Tourist SOS calls its AI assistant Terra. Terra's appropriate role is to support defined information and administrative tasks within the feature and service being used. Terra is not a clinician, emergency dispatcher, insurer, government official, or substitute for accountable human decisions.

Start With the Task, Not the Model

Each AI use should have a named purpose, permitted inputs, expected output, responsible owner, review requirement, quality measure, fallback, and stop condition. A general claim that “AI coordinates emergencies” is not specific enough to govern safely.

Potential administrative uses may include:

structuring information reported by an authorized user;
identifying missing fields or contradictory administrative details;
drafting a summary for review against the original source;
supporting search across approved provider or destination information;
preparing questions or requests for a provider, payer, or partner;
assisting with low-risk language access while labeling limitations;
classifying workflow state or routing work for authorized human review;
checking whether required administrative evidence is present.

Whether any function is available depends on the product, deployment, data, jurisdiction, and accepted arrangement. A possible use case is not a claim that it is live everywhere.

Keep High-Consequence Decisions Outside the Model

AI output should not independently diagnose, prescribe, determine clinical urgency, clear a traveler for transport, accept a patient, dispatch an emergency resource, authorize treatment, issue a guarantee of payment, approve a claim, or exercise public authority. The appropriately qualified and authorized person or organization retains each decision.

Administrative urgency flags can help prioritize human attention, but they are not clinical triage and must not suggest that it is safe to wait. The absence of a flag does not rule out a serious condition.

Human Review Must Be a Real Control

“Human in the loop” is meaningful only when the reviewer has the necessary role, context, time, authority, and source evidence. Reviewers should be able to see what the model received, what it produced, which parts are uncertain, and how to correct or reject the output. A reviewer clicking approve without those conditions is not a reliable safeguard.

The workflow should identify when review is mandatory, what happens if a reviewer is unavailable, and which output may never proceed automatically.

Preserve Source, Provenance, and Uncertainty

AI can produce fluent text that is wrong. Important summaries should remain linked to their sources. Provider and destination facts need evidence and verification dates. Patient-reported information should remain labeled as reported until the responsible professional verifies it. Conflicting information should not be silently collapsed.

The World Health Organization has warned that large language model responses can appear authoritative while being inaccurate or incomplete. Its statement on safe and ethical AI for health calls for transparency, expert supervision, rigorous evaluation, and protection of sensitive data.

Language Access Needs Its Own Safety Model

Machine translation may help a person understand basic navigation or prepare low-risk administrative text. It can also mistranslate symptoms, medication, consent, exclusions, or urgency. A program should identify which content has been professionally reviewed, which is machine-assisted, and which requires a qualified interpreter or the provider's approved language service.

Never advertise language coverage merely because a model can generate text in that language. Test meaning, script rendering, accessibility, escalation, and the real human support route.

Protect Data Before Asking the Model

A user's ability to paste information into a tool does not establish permission to process it. A deployment must define purpose, lawful basis or other required authority, minimum necessary input, model and vendor access, retention, training use, geography, security, logging, deletion, and incident response. De-identification and synthetic data should be used when the task does not require real personal information.

WHO's health data governance policy brief highlights the role of governance, quality, privacy, equity, and human rights in trusted AI-enabled health systems. Deployment-specific law and agreements still control.

Evaluate Performance in the Actual Context

Test representative languages, locations, user roles, and information quality.
Measure false confidence, omissions, unsafe suggestions, and reviewer correction.
Test model, vendor, network, and integration outages.
Review performance after material model, prompt, data, or workflow changes.
Provide a route to contest outcomes and report harm or misleading content.
Pause or narrow a function when evidence does not support continued use.

WHO's guidance on large multi-modal models recommends well-defined tasks, stakeholder involvement, regulation, auditing, and impact assessment. Its broader principles—autonomy, safety, transparency, accountability, inclusion, and sustainability—provide a useful test for administrative AI as well as clinical applications. WHO does not endorse Tourist SOS or Terra.

The Tourist SOS Commitment

Our direction is AI-assisted coordination with explicit human and institutional authority—not AI replacing the people responsible for care, dispatch, coverage, payment, or public decisions. We aim to show users what a result means, what it does not mean, and what confirmation is still required.

Public descriptions will continue to distinguish demonstrations and proposed capabilities from available services. Current limitations are set out in our Medical and AI Disclaimer, Terms, and Trust Center. Responsible AI is not a feature we complete once. It is an operating discipline that must follow the product as its users, models, evidence, and environments change.