> ## Documentation Index
> Fetch the complete documentation index at: https://help.rytz.com.au/llms.txt
> Use this file to discover all available pages before exploring further.
# Limits and safety
> What the AI assistant will not do, how to recognise when to engage a solicitor, the family-violence safety overlay, and the platform's policy on legal-advice boundaries.
import { CardGroup, Card, Note, Tip, Warning } from '@mintlify/components'
The AI assistant is one of the most-capable surfaces on RYTZ. It is also the surface where being clear about limits matters most. This page covers what the assistant will not do, how to recognise when a question requires a solicitor, and the platform's safety overlay for users disclosing family violence.
## The legal-advice boundary
The most-important boundary:
**RYTZ provides legal information, not legal advice.** The assistant explains the law, walks the framework, and helps you reason about your matter. **It does not give legal advice.**
Legal advice — recommendations specific to your matter, made by someone with legal qualifications and professional indemnity, that you can rely on as professional advice — comes from a family-law solicitor.
The assistant is calibrated to surface this boundary. When you ask a question that requires advice rather than information, it will tell you, and recommend a solicitor consultation.
### Information vs advice — the practical distinction
| Information (the assistant) | Advice (a solicitor) |
| ------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
| "Section 60I requires an FDR certificate before filing parenting applications." | "In your matter, you should apply for an FDR exemption on these grounds." |
| "Stanford requires the court to consider whether it is just and equitable to alter property interests." | "In your matter, the just-and-equitable threshold is likely to be a live issue because \[...]." |
| "The four-step framework applies to property matters." | "Apply for these specific orders on this specific timing." |
| "Affidavits should generally avoid hearsay." | "Don't include paragraph 17 of your draft because it's inadmissible hearsay." |
The assistant gives you the framework. The solicitor applies it to your matter and tells you what to do.
The line is not always clear-cut. The assistant is calibrated to lean conservatively — when uncertain, it surfaces the recommendation to consult a solicitor.
## When the assistant will not answer
There are some questions the assistant will decline to answer. The most-common categories:
"Should I file for interim parenting orders?" The assistant will explain the framework, explain what interim orders do, walk you through the considerations — but the final "should I" is not its question.
"Should I make this final offer?" / "Should I withdraw from FDR?" / "Should I lodge a contravention?" The assistant will explore options; the decision needs human input.
Criminal matters, immigration matters, commercial litigation, employment law, US/UK/other-jurisdiction family law. The assistant is scoped to Australian family law and will redirect to appropriate resources for matters outside that scope.
"What will the judge decide?" / "What's my chance of winning?" The assistant declines to predict outcomes. Family-law decisions are heavily fact-driven; predictions are unreliable and the assistant doesn't perform them.
Drafting content that would mislead a court, suggest a frivolous application, or otherwise harm the integrity of proceedings.
Questions about people who haven't consented to being part of your matter (children's friends, partners' new partners, employers). The assistant treats third-party privacy carefully.
## How to recognise when to engage a solicitor
The assistant flags this directly when relevant. A few signals to watch for:
| Signal | What it means |
| ----------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------- |
| The assistant says "this requires legal advice" | Engage a solicitor |
| The matter has unusual features (international elements, business-valuation disputes, contested mental-health considerations) | Engage a solicitor |
| The other party is represented by a solicitor with experience in your matter type | Engage at least a limited-scope retainer |
| You are about to file or sign anything binding | Have a solicitor review |
| You are facing court without representation in a contested matter | Strongly consider retaining counsel for the hearing day |
| Stakes have escalated (children's safety, substantial property, FV-FVO interaction) | Engage a solicitor |
[Legal aid eligibility](/learn/legal-aid-eligibility) is the first stop if cost is the barrier. For ineligible parties, community legal centres and unbundled legal services are typically more accessible than full retainer.
## Family-violence safety overlay
For users with family violence disclosed in their Master Case File, the assistant applies a safety overlay:
**Safety-first responses.** When a question touches FV-related matters, the assistant prioritises safety above strategic optimisation. If a strategic option puts the user at risk, the assistant surfaces that risk before discussing the strategy.
**Crisis resources surfaced when relevant.** If a conversation suggests acute risk (current threat, escalating violence, child-safety concerns), the assistant surfaces 1800RESPECT, local crisis lines, and platform safe-exit features alongside any other content.
**Safe-exit feature available.** The platform's safe-exit button (a "leave site without trace" feature) is enabled by default for FV-pathway users. The assistant references it in conversations where device-monitoring is a concern.
The overlay can be turned off if not appropriate (you've moved past the high-risk phase and find the safety prompts excessive); the default is on for FV-disclosed users.
## Recognising AI errors
The assistant is calibrated but not infallible. Recognise:
* **Citation errors** — occasionally the assistant cites a case incorrectly or attributes a holding to the wrong case. Always verify citations against the [Legal Research Library](/learn/legal-research-library) or AustLII before relying on a cited authority.
* **Statutory currency errors** — the platform's training data has cutoffs; for very recent legislation or case law, the assistant may not know about it. The applies-to footnote on legal-substance content shows the platform's knowledge horizon.
* **Hallucinated procedural detail** — the assistant occasionally confabulates specific procedural details (form numbers, fee amounts, timing). For procedural questions, cross-check against the [Forms Library](/forms/overview) or FCFCOA's official resources.
* **Strategic over-confidence** — when the assistant gives a strong strategic recommendation, it should also note alternatives and uncertainty. If a recommendation lands without uncertainty caveats, treat it as preliminary.
## Privacy under AI processing
A reminder of the privacy boundary:
* **What you tell the assistant** — processed via Anthropic's API under contractual data-protection terms. Not used to train models.
* **Files you upload** — same. Not used to train models. Encrypted at rest.
* **Conversation history** — stored on the platform, encrypted at rest, scoped to your account, deletable on demand.
* **What the assistant tells you** — generated from grounded sources (Education Portal, Legal Research Library) plus the underlying model. Not stored as part of any training corpus.
See [Privacy and data](/account/privacy-and-data) for the full data practices.
## What's next
Step back to the framing.
Architecture-level — what's grounded, how answers are produced.
First stop when a solicitor is needed and cost is the barrier.
How AI processing fits the platform's data practices.