AI Policy

moned exists to connect people building a life abroad with accountable human knowledge from Professionals, Content Creators, and Guides. AI is used only where it can reduce search or administrative effort without becoming the authority for a sensitive marketplace decision.

This policy describes behavior evidenced in the current code. A possible future use is not presented as active. moned should update this page and the technical inventory together when an AI surface changes.

In controlled conversational discovery, Gemini may interpret a request, ask a useful clarification, and choose from a small server-owned set of read or reversible-selection tools. Tool arguments are bounded and validated, the backend rechecks ownership and current conversation state, and public expert facts come only from moned's governed services. The model never supplies the content of expert cards, comparisons, calendars, prices, or status surfaces.

The legacy search path uses Gemini to propose a versioned structured search patch and keeps deterministic search rules authoritative. Manual deterministic search remains available regardless of provider health.

If the provider is unavailable or returns an empty, invalid, unsupported, or ungrounded result, the conversational surface identifies that guidance is temporarily unavailable and offers deterministic/manual discovery. Provider failure is not presented as successful AI reasoning.

AI-assisted discovery is not professional advice. The composer asks users to describe only the kind of expert they want. Recognizable card or payment secrets and first-person health, religion, ethnicity, sexuality, political, union, genetic, or biometric disclosures are rejected before assistant persistence or provider use. Because automated detection cannot recognize every possible wording, users should not enter emergency information, passwords, credentials, or sensitive case facts.

Session artifacts are conditional product features. Each participant's microphone audio can be captured separately only after every required participant has consented and published a microphone track. Camera video and screen sharing are never included. Consent can be declined or revoked without ending the call.

Revoking consent stops further capture, prevents a new summary, and schedules captured raw audio for deletion. Raw microphone audio is deleted after successful transcription and is kept for no longer than seven days only when recovery is needed. Internal speaker-labelled transcripts are retained for up to 30 days and shared summaries for up to 365 days unless erased earlier.

Google Cloud Speech-to-Text V2 produces the transcript when configured. A versioned summary prompt and response schema constrain the Gemini summary. Capture, transcription, or summary failure does not change payment, booking, session outcome, or dispute state. Participants see that a summary could not be created.

A shared summary is designed for continuity. It may omit context or contain an error. Participants should rely on the conversation and any separately issued professional document rather than treating the summary as advice or an official record.

Deterministic domain rules and governed records remain authoritative for service publication, availability, booking lifecycle, server pricing, supplier identity, tax classification and decisions, documents, marketplace reporting, withholding, payments, payouts, refunds, permissions, and account state. Generative AI does not approve a tax corridor or produce the canonical tax calculation.

Expert credential review, experience or profile review, capability approval, regulated-service scope, enforcement, appeals, and sensitive administrative actions are not delegated to a generative model. Where those workflows require an authorized reviewer, the reviewer records the decision and reason through the governed administrative path.

This does not mean every moned operation receives individual human review. Routine validation, rate limits, lifecycle transitions, and other deterministic rules operate automatically. Human accountability is concentrated where judgment and authority are required.

Current AI provider categories are cloud model generation and cloud speech transcription. The implemented providers are Google Vertex AI Gemini for search interpretation and summaries, and Google Cloud Speech-to-Text V2 for transcription. Provider configuration and retention must remain aligned with the Privacy Policy and approved operational settings.

Conversational discovery is limited to 1,200 characters per turn, bounded recent context, compact tool facts, and an EU Vertex endpoint when enabled. Anonymous raw wording is retained for up to 24 hours and bounded structured state for up to seven days. Signed-in raw wording is retained for up to 30 days and bounded structured history for up to 90 days. People can delete individual conversations or all assistant history.

Google does not use moned's prompts to train or fine-tune its models without permission. Google approved moned's project-level exception from prompt logging for abuse monitoring. This approval concerns the provider's abuse-monitoring prompt logging. It does not change moned's own retention rules or processing required to operate and secure the service.

moned should not use support reports, security disclosures, private evidence, credentials, or internal reviewer notes as general model prompts. Sensitive case content is not included in ordinary logs or notification payloads.

AI-assisted search identifies itself as AI-assisted and explains that it is narrowing expert discovery. Conversational tool phases describe application actions such as searching public profiles; hidden reasoning and raw tool arguments are never shown. Provider degradation appears as an explicit notice. Transcript and summary surfaces describe the consented microphone-audio source and distinguish a summary from camera recording and professional advice.

Generated text should never imitate a verification label, administrative decision, or provider-authored professional document. Review labels come only from the governed verification system.

AI systems can misunderstand a request, miss context, reproduce bias, or generate incorrect text. Structured schemas, deterministic fallbacks, consent, access control, and human review at authority boundaries reduce risk; they do not make errors impossible.

Report a search or summary problem through product support, a privacy concern through the privacy route, unsafe conduct through conduct reporting, or a vulnerability through security disclosure. moned remains accountable for the product decisions it makes around AI-assisted systems.

• NIST AI Risk Management Framework and Generative AI Profile
• OECD AI Principles: human rights, transparency, robustness, and accountability
• European Commission overview of the EU AI Act’s risk-based framework