AI Transparency

AI Transparency Policy — Brief (Working Title)

Positioning. Brief is an AI-powered correspondence assistant. It summarizes and translates documents and helps users draft messages based on the facts and communication intentions they provide. Brief does not independently determine legal rights, obligations or legal strategy and is not a substitute for advice from a qualified professional.

Positionierung. Brief ist ein KI-gestützter Korrespondenz-Assistent. Er fasst Dokumente zusammen, übersetzt sie und hilft Nutzerinnen und Nutzern, Nachrichten auf Grundlage der von ihnen angegebenen Tatsachen und der von ihnen gewählten Mitteilungsabsicht zu verfassen. Brief bestimmt nicht eigenständig Rechte, Pflichten oder eine rechtliche Vorgehensweise und ersetzt keine Beratung durch eine qualifizierte Fachperson.

Version: 1.0 · Date: 2026-08-05

1. Purpose and scope

This policy defines how Brief:

labels AI-generated content so users always know what was produced by AI;
communicates the limitations of AI analysis honestly;
technically and editorially prohibits fabricated legal citations;
decides when and how to recommend human professional review.

It applies to every AI-generated surface in the app: the plain-language explanation, extracted deadlines, risk notes, reply drafts, and any future AI feature. It does not apply to static editorial content written by humans (onboarding copy, help pages), which must not be labeled as AI-generated.

2. Labeling of AI content

Principles

Unavoidable: Every screen displaying AI output carries the global AI label (string 1 in in-app-disclaimers.md) in a persistent position — not behind a tap, not dismissible, not below the fold by default.
Unambiguous: The label states both facts: the content is AI-generated, and the user must verify against the original letter.
Consistent: One label style app-wide (icon + short text), so users learn to recognize it.
Exported too: When a user exports or shares an analysis or reply draft, the export carries the AI label and the reply-composer footer (string 5). AI assistance in a document the user sends onward must remain visible in the artifact itself.
Demo separated: Demo-mode output is additionally marked as a fictitious sample case so it can never be mistaken for a real analysis.

Never

Present AI output in a voice or layout implying an authority issued it (no letterheads, seals, official-looking formatting in analyses).
Label AI output as "checked", "verified", or "confirmed" — the app performs no legal verification.
Use the word "advice" (or "Beratung"/«консультация»/„consultanță") for AI output anywhere in UI, marketing, or App Store copy. Approved framing: "explanation", "analysis", "summary", "draft".

3. Model limitations — what we tell users and how we build for it

Disclosed limitations (in the in-app "How Brief uses AI" page and reflected in Terms §2):

AI can be wrong or incomplete, including in confident-sounding ways (hallucination).
AI can misread dates, amounts, and names, especially from imperfect scans; OCR quality bounds analysis quality (hence the low-confidence OCR warning).
AI has no knowledge of the user's full situation — it sees only the submitted text.
AI is not current legal doctrine: laws, fee schedules, and administrative practice change; the model's knowledge has a cutoff and the app performs no legal research.
Language nuance: analyses in ru/ro/en are explanations of a German-language legal context; terminology is approximated and German terms are kept where precision matters.

Product safeguards backing the disclosure

Deadlines and amounts are displayed together with the source snippet from the OCR text ("show in original") so users can verify in one tap.
Detected deadlines ≤ 3 days or past trigger the urgent-deadline escalation (string 3).
Confidence thresholds: below-threshold OCR triggers string 4 and blocks "high-precision" claims in the analysis.
The analysis prompt instructs the model to express uncertainty explicitly rather than guess, and to say "not stated in the letter" instead of inventing facts.

4. Prohibition on fabricated law citations

Policy: The app must never present a fabricated or unverified legal citation (statute, section, case, court decision, official fee schedule) as fact.

Rules

Default: no citations. The system prompt instructs the model not to cite specific statutes, case law, or section numbers in analyses, except when the citation appears verbatim in the user's letter — then it may be referenced as "the letter cites § X …".
Letter-sourced only: Any legal reference shown must be traceable to the OCR text of the letter. References originating from the model's own knowledge are stripped or generalized ("relevant tenancy law" instead of an invented "§ 573c BGB"-style citation) by prompt policy and, where feasible, output post-processing.
No invented case law, ever. Court decisions, file numbers, and precedents must never be generated. Output post-processing treats decision-citation patterns not present in the source letter as violations and removes them.
Reply drafts: Same rules. A reply draft may reference the sender's own file number and the letter's cited provisions, but must not add legal argumentation resting on model-invented citations.
QA gate: The regression test suite includes adversarial prompts (letters inviting citation) and fails the release if fabricated citations appear.
Incident handling: A confirmed fabricated citation reaching production users is a severity-1 content incident: hotfix the prompt/filter, document it, and review this policy.

5. Human-review recommendation policy

Principle: The more consequential the letter, the stronger and earlier the push toward a qualified human — a lawyer (Rechtsanwalt), tax advisor (Steuerberater), recognized advice center (e.g. Verbraucherzentrale, Migrationsberatung), or the issuing authority itself.

Tiers

Tier	Trigger	App behavior
Baseline	Every analysis	Global AI label (string 1); "verify against the original" framing throughout.
Sensitive category	Court, tax, immigration/residence, employment termination, medical	Contextual banner (string 2 / category variants) recommending Rechtsanwalt / Steuerberater / authority contact, shown above the analysis, non-dismissible for the session.
Urgent	Deadline ≤ 3 days or expired	Escalation banner (string 3) prioritized above all other banners; reply-draft screen repeats it.

Combined	Sensitive and urgent	Both banners; analysis header additionally surfaces the professional-help recommendation as the first "next step".

Constraints

Recommendations name professional roles and public bodies, never specific commercial firms; the app takes no referral fees and hosts no attorney advertising.
The recommendation is a recommendation — the app never blocks a user from seeing their analysis, and never claims the user "must" hire anyone.
Category detection errs toward sensitivity: at classification uncertainty between a sensitive and non-sensitive category, the sensitive banner is shown.

6. What the AI is never used for

No automated decisions with legal or similarly significant effects on the user (GDPR Art. 22 posture: the app informs; the human decides).
No sending of replies on the user's behalf — the user always exports/sends manually.
We do not train models on user documents and have no pipeline that could. What the AI provider does with the text it receives is governed by its own terms, which are described in the Privacy Policy.
No inference of characteristics beyond what the analysis requires (no profiling of users from their letters).
No generation of content impersonating an authority.

7. Governance

Change control: Changes to the system prompt affecting citations, disclaimers, or recommendation tiers require a recorded decision by the provider, and a regression run of the citation and disclaimer test suite before release.
Review cadence: This policy is reviewed at least every six months and upon: AI provider/model change, EU AI Act guidance updates, App Review guideline changes affecting AI apps, or any severity-1 content incident.
Public version: A user-facing summary of Sections 2–5 will be published in-app under "How Brief uses AI".