LLM Model Use Notice
Advance notice under applicable law

Where is the LLM model used?

• Safety classification (Safety Router): automatically detects crisis signals in what you type (suicide · self-harm · immediate danger). Locale-aware rules run first, and an LLM supplements them with a second pass when enabled.
• Card generation: general input may be used to generate one new card. Vulnerable signals receive only reviewed cards, and immediate-risk signals do not generate a card. Generated output passes through a prohibited-wording filter before it is shown.

Is it used for important medical or financial decisions?

No. This service is not used for decisions in healthcare, finance, credit, hiring, education, or the public sector. It’s a small ritual for your mind, comfort content. It doesn’t diagnose, treat, or prescribe anything.

That said, because emotional well-being sits close to healthcare, we refresh a self-assessment memo every quarter. The current memo is kept internally, and a public summary is available on request.

Is my information sent to the LLM model?

• A free-text line allowed for transfer (up to 120 Korean characters or 200 English characters) is transferred abroad to an LLM model. Dynamic text matching a defined special-category or crisis pattern is not sent and uses a local safety or fallback path.
• dadokim does not store the original line or its hash after the API response. The LLM model provider normally deletes API inputs and outputs within 30 days, subject to policy-enforcement or legal exceptions.
• Detected patterns such as contact details, email addresses, resident-registration numbers, card numbers, and account numbers are masked before transfer. Because no automated filter can guarantee removal of every identifier, do not enter names, contact details, or other identifying information.
• Commercial API inputs and outputs are not used for model training by default.

Disagree with something a card said?

• Use the “This card isn’t right” button on the card result screen to report it on the spot
• Email: privacy@dadokim.com
• Reported cases feed into a red-team analysis every 90 days

The safety net (4 layers)

1. Locale-aware rule-based crisis-signal classification
2. LLM-assisted classification when enabled (it cannot lower the rule-based risk level)
3. Pre-output wording filter (Medical Service Act · Pharmaceutical Affairs Act · one regeneration retry · safe-card fallback on failure)
4. User reports (4th layer · fed back in)

When Level 2–3 signals are detected (a direct mention of suicide or self-harm · immediate danger), card generation itself is blocked and you’re routed toemergency guidance (in Korea: 109 · 119 · 1577-0199 · outside Korea: findahelpline.com).

※ Automated crisis detection covers both Korean and English input.

Related standards

• Advance-notice requirements for generative LLM models
• Personal Information Protection Act (PIPA)
• Medical Service Act · Pharmaceutical Affairs Act · Mental Health Welfare Act (wording filter)
• E-commerce Act (business information · payments · refunds)