AI Transparency

AI is used in eight areas on the platform:

• AI overviews — short summaries on eligible resort, school, and instructor detail pages, generated from the information held about that entity and cached server-side. School summaries are generated and displayed only for claimed, published schools. For unclaimed schools, stored descriptions, generated school metadata, staff summaries, and AI summaries are deleted from the active school-content stores, including school translations and generated-copy caches, instead of being kept for reuse after claiming. Basic directory facts can remain. Claiming does not itself verify the accuracy of new text. Historical exports, backups and provider-held copies require separate retention and deletion handling; see the Privacy Policy.
• AI Resort / School / Instructor finder — you describe what you want in plain language ("family resort with a beginner area in Austria") and the AI ranks matching options with reasons. Picks are one signal among several; the directory's own ranking still applies.
• Live voice and text assistant — optional OpenAI voice/text support and authorised listing/account tools. It can navigate, filter and highlight controls. Supported name/bio edits, unpublished lesson drafts and booking rescheduling use existing permissions, a preview and your explicit confirmation; payments, refunds, publishing and general message sending are excluded. Guest voice defaults to five minutes unless Admin configures another allowance, with limited text replies. Shared-network and per-call limits may reduce available time. Account use consumes displayed connection, started-minute and text-reply credits once Admin configures rates; connected silence, mute and inactivity grace count. Weather/snow estimates are not guarantees. Basic contextual support uses published guides without AI credits. Its separate optional AI explanation is limited to the selected support issue; general live-chat and voice prices are unchanged. A support handoff is sent only after your review and explicit confirmation.
• AI Chatbot — answers questions about resorts, schools, lessons, and the booking flow. Disclosed as AI on the first message.
• Gear recommendations — a short questionnaire about your ability and conditions feeds an AI ranking of equipment products.
• Quick Lesson Notes — when an instructor records a note, the audio is sent to a speech-to-text service. The transcript, together with any student name, activity and level supplied by the instructor, is then sent to the configured text-model provider to clean the note, generate a title and extract action items.
• Profile drafting (instructors) — instructors can ask the AI to draft their bio, the two-sentence summary at the top of their profile, or the line under each of their specialities. Only facts already on their saved profile are sent to the configured text-model provider — such as their specialities, qualifications, languages, where they teach, and when they started — and the draft goes into their editor. Nothing is published until the instructor reads, edits, and saves it, and each request uses an AI credit.
• Editorial helpers (admin-only) — admins can ask the AI to draft FAQ answers, summarise long reviews, or extract data from a school's website. Output is reviewed and edited by a human before it goes live.

The decisions that affect your booking are made by people, not by AI:

• Verification — A Find a Ski School administrator makes the approval decision on school and instructor verification applications. AI may assist with pre-screening, but it cannot approve an application. Find a Ski School labels a school's current verification status, an instructor's own approved application and affiliation with a verified school as three separate signals. Evidence-specific checks appear separately only when the platform stores evidence for that capability. Verification is a scoped platform status, not an endorsement or a guarantee of quality, safety, competence, legal permission to work, insurance cover or lesson outcome. Check any evidence that matters to your decision.
• Reviews — anyone with an email-confirmed Find a Ski School account may submit a review. Reviews linked to an eligible completed booking through Find a Ski School are labelled “FASS Verified Booking”; other reviews show their source or verification status and do not receive that label. Providers can respond but cannot remove negative feedback. Suspected abusive content is reviewed by humans.
• Bookings — every booking is a contract between you and the provider shown at checkout — a ski school or an independent instructor. The provider confirms or declines; we facilitate but are not party to the contract.
• Moderation — flagged content (reviews, messages, questions) is reviewed by humans before action.

Schools can buy promotion from us — sponsored placement, prime positioning on resort pages, and a paid subscription tier. Because that money could otherwise quietly shape what the AI recommends to you, here is exactly where it does and does not reach:

• It does not change the match criteria — the \u201cMatches n/N criteria\u201d count the AI finder shows against a school is read from the school's own listing \u2014 the activities, levels, ages, languages and places it says it covers \u2014 against what you typed. Paid promotion is not one of those criteria and cannot change the count, so a school cannot buy a better-looking match.
• It does not affect the order either — promotion plays no part in how results are ordered — not as a signal and, since 2 August 2026, not even as a tie-break between equal scores. Ties are broken by earned quality and review count. Paying cannot move a school up the list at all.
• Sponsored slots are fixed and labelled — separately from ranking, a limited number of directory positions are sold as sponsored slots. These are always marked "Sponsored" where they appear, and they do not displace organic results — they sit alongside them.

If you would rather compare without any of this, every directory listing can be sorted by rating or by price, which ignores promotion entirely.

AI-generated text can be inaccurate, out of date, or misleading. Common failure modes we've seen:

• An overview mentions a feature the resort no longer offers (e.g. a closed lift, a discontinued ski school).
• An AI Finder pick ranks a school highly because the school's listing claims a speciality we haven't independently verified.
• The chatbot answers a question with confident but incorrect detail (a hallucination).

How to flag a wrong AI output: every detail page has a "Suggest correction" link in the page footer. For chatbot replies, use the thumbs-down feedback button at the bottom of the response. We re-generate overviews when the underlying entity facts change, so a corrected listing fixes the AI summary on the next refresh cycle.

For booking-affecting questions (price, availability, lesson content), always check the school's own website or contact the school directly. The AI overview is a quick read, not a binding source of truth.

The live assistant uses OpenAI for text and realtime voice. Voice mode sends microphone audio while connected, and supported tools retrieve listing data or permitted customer information to answer a request. The current conversation stays in page memory; our application does not store raw voice audio. OpenAI processing and retention depend on the applicable account settings and terms. Other AI features use configurable providers. Quick Lesson Notes sends recordings to its speech-to-text service and transcripts/context to its selected text-model provider. Supported providers include Anthropic, OpenAI, Google, Groq, Together AI, Mistral, DeepSeek, OpenRouter and xAI. Ask us for the active subprocessor list.

Find a Ski School does not train AI models on your data. We don't run our own foundation models, and we don't fine-tune third-party models on your queries, conversations or profile. That promise doesn't depend on any third party and stays constant.

What the providers do with the inputs we send is governed by the contract and Data Processing Addendum for the provider active at the time. The provider can change, so the links below are examples rather than a complete current list. Contact [email protected] for the active provider and applicable terms.

• OpenAI — openai.com/policies/data-processing-addendum
• Anthropic — anthropic.com/legal/dpa
• Google — cloud.google.com/terms/data-processing-addendum

Training and provider-side retention terms differ by provider and may change. The applicable contract for the active provider is authoritative; do not assume the same retention period applies to every supported provider.

What crosses the boundary:

After AI & Chatbot consent, the signed-in assistant may automatically send OpenAI your saved account name, self-reported activity and ability, piste comfort, up to eight currently public favourite schools, instructors and resorts per type, and limited completed-booking counts and scheduled hours for learner accounts. These are reference preferences, not verified skill or time skied; bookings may be for other people. Your current request takes precedence. Health information, contact details, birth dates, free-text bios, payment data and other people's profiles are excluded from this automatic context. Guests, demos and support-impersonation sessions do not receive a real account's saved preferences. Information is read for the request or tool; conversations do not train the assistant.

The optional trip brief contains details you apply, such as a country or resort, dates, activity, self-reported ability, adult counts, child ages you enter, lesson format and budget. After AI & chatbot consent, the applied brief can be sent to OpenAI with your next question or when you update a connected voice conversation. Child ages are not inferred from other people or automatically taken from their profiles. Your latest explicit request takes precedence. These are your stated requirements, not verified suitability, a booking quote or confirmed availability.

Saving a trip is a separate choice: you must confirm saving the structured brief and selected public listings to your own account. We store those details, your account reference and timestamps until you delete the plan or account; they are included in account export and erasure. The full conversation and audio are not saved with the plan. Saving or reading a plan does not itself call an AI provider or spend assistant credits. Public shortlist entries are checked again when saved or read and can become unavailable. Loading a plan does not book anything; applied details reach OpenAI only under the AI-consent rules above.

Answer feedback is optional and account-linked. By default we retain only your selected category and topic, an account reference and timestamps. You may separately choose an edited answer excerpt of up to 500 characters, review its redacted preview and explicitly confirm sharing it with FASS. Common contact details, links and credentials are removed, but names or other personal details may remain: check the preview. Authorised administrators review categories and deliberately shared excerpts; report displays omit account identities, but the underlying feedback remains linked to your account. Feedback is not sent to an AI provider or used to train models, and it does not create a support request. You can remove your feedback; retained reports are included in account export and erasure. Reports are excluded from ordinary feedback views after 90 days. Bounded hourly maintenance removes expired rows; physical deletion is not guaranteed at an exact deadline.

Navigation reliability checks use receipts that expire after 10 minutes and keyed hashes in server memory to bind them to the owner, offered action and access context. A 30-minute report for the current server process counts page loads, applied filters, highlights, failures and cancellations; it contains no routes, filter values, account identifiers, transcripts, audio or IP addresses and clears on restart. These are browser reports, not confirmation that a problem was solved. Streamed draft replies remain temporary page content under the conversation rules above.

We keep limited technical records to investigate assistant connection failures: a random support reference, observation times, the type of assistant operation, and fixed diagnostic codes, stages, status and configuration-field names. We do not attach account, guest, session, call or IP/network identifiers, page URLs, messages, prompts, transcripts, audio or raw provider/error responses. Shared storage holds at most 1000 references; authorised Admin views show the latest 40. Records are excluded from those views 24 hours after the last observation; bounded scheduled maintenance removes expired records without guaranteeing deletion at the exact expiry time. If shared storage is unavailable, this server may show a five-minute local fallback that is lost on restart and labelled incomplete. These records are not sent to an AI provider or used for model training. A reference you supply in a support request may connect that failure to your case under the ordinary support-retention rules.

• To providers: entity facts used for overviews, AI Finder queries, gear-questionnaire answers, chatbot messages and, when Quick Lesson Notes is used, the audio recording, raw transcript and any student name, activity or level supplied. The text provider receives the transcript to clean it, generate a title and extract action items. When an instructor asks for a profile draft (a bio, a summary, or a speciality line), the text provider receives facts from their saved profile, such as their bio, specialities, qualifications, languages, where they teach, and when they started. The live assistant also receives microphone audio while voice is connected, text and voice transcripts, page context, requested listing data and limited details of your own upcoming bookings when requested.
• Not to providers: payment card details or GPS lesson-tracking data. The live assistant's automatic own-account personalization is limited to the fields described in this section. Other account information is sent only when you supply it, an authorised tool needs it for your request, or it is supplied as lesson-note context. Quick Lesson Notes shares recordings, transcripts and supplied context as described above.

Technical AI usage records can include feature, provider, model, token counts, timing and success or error details. These usage records contain metadata, not the content of your conversation. Conversation storage, Finder query logs and saved searches differ by feature. See the Privacy Policy section "AI-powered features & automated processing" for the full legal-basis breakdown.

Some user-facing AI surfaces can be turned off independently. These settings do not control Quick Lesson Notes, which an instructor chooses whether to use. The settings below persist in your browser; signed-in users see the same toggles in Account Settings → AI features. Either surface writes to the same local store.

An Admin-issued private preview link provides access in one browser for at most 72 hours. Its one-time redemption sets the signed HttpOnly fass_assistant_preview cookie. It grants no extra account or data permissions and bypasses neither AI consent nor credits, rates or safeguards. Server grant records contain hashes, labels and timestamps, not the raw link token; expired or revoked grants are removed after 24 hours by maintenance. Existing authorised QA demos use the signed HttpOnly fass_assistant_demo_scope cookie for at most 3 hours, a fixed fictional school and a separate Admin setting. These requested access/security cookies do not themselves permit optional AI processing or create a self-service temporary account.

With Analytics consent, eligible public homepage visitors may see title A, B or C for the assistant orb. Testing is paused until Admin starts it; without Analytics consent you see A or the chosen winner, with no experiment identifier or metrics. A random browser identifier in local storage lasts 30 days and is cleared when Analytics consent is withdrawn. First-party server records retain its hash, experiment run, variant, language and assignment/event timestamps for visible impressions, clicks and successful text replies or voice connections for up to 90 days. They contain no IP address, account ID, profile, messages, transcripts or audio. Private previews, demos, administrators and impersonation sessions are excluded. This visual test does not require AI consent; actual chatting still does. The body and button stay the same, and Admin chooses a winner manually. Withdrawing Analytics stops new measurement; existing server records expire under this retention schedule.

Other options:

• Live voice and text assistant AI & chatbot consent is required before chatting; voice also needs a deliberate start and microphone permission. You can choose an offered voice before connecting, stop or reset, withdraw consent and download available text/captions without audio. Idle and maximum-length limits, guest allowances, account credits and security restrictions can end a conversation. FASS does not collect full live transcripts/raw audio for training. Any future training reuse would need a separate purpose and lawful basis. Standard search and human support remain available. Contextual support starts when you accept an inline offer. Published instructions do not need AI consent. Before an optional AI explanation, we show which issue details and customer-visible ticket replies will be used and require AI & chatbot consent. Internal notes, private staff replies and attachments are excluded. Human support remains available; an existing request closes only after you confirm resolution.
• The chatbot The legacy concierge is opt-in. Its history remains in your browser for 24 hours after your last message. Messages pass through the configured AI provider. The live assistant has the separate memory and retention arrangements described above.
• Gear recommendations only run when you explicitly fill in the questionnaire. Not filling it in is the opt-out.
• The AI Finder only runs when you type a query into its input. The directory itself works without touching it.
• Quick Lesson Notes does not offer a manual-only mode or collect a learner consent record. If you do not want audio or AI processing, ask the instructor not to use Quick Notes and request a manually written recap through another agreed channel.
• You can request access to, correction of, or deletion of your AI-related data at any time. See the Privacy Policy section "Your rights" for how to do that.

Questions or a wrong AI output to flag? Contact us.