How we use AI

A working methodology for using AI in legal services without committing the unauthorized practice of law. We publish this because the ground rules are still being written, and we'd rather be transparent about ours than pretend the question doesn't matter.

The principle

Family law is fact-specific. Two clients with similar-sounding situations regularly need different advice because of small differences in domicile, asset structure, marriage date, country of citizenship, or the way a will was executed. This is not a market failure of legal services — it's the nature of the work.

That truth has a corollary: a website cannot give you legal advice about your case. Not even a sophisticated, AI-powered website. The moment a tool tells you what to do in your specific situation, it is doing something only a licensed attorney is allowed to do, and it is doing it badly because it doesn't have the facts.

So the first rule of how we use AI on this site is: we don't use it to give advice. Everything here is educational. Everything here is general. Everything here is meant to make a conversation with an attorney better-informed — not to replace that conversation.

What this site does

• Provides bilingual educational definitions of family-law terms with explicit jurisdictional tags.
• Explains how a concept works under Massachusetts family law, under PRC family/civil law, or in the cross-border interaction between the two.
• Names the AI model that produced each piece of content and the date it was generated.
• Identifies which entries have been reviewed by an attorney and which are AI-generated and pending review.
• Marks limits explicitly: where the law is unsettled, where treatment varies by case, where the cross-border picture differs from the single-jurisdiction one.
• Routes every page to a clear path for talking to an actual attorney about an actual situation.

What this site does not do

• It does not give advice about your case. Not in the encyclopedia, not in any chatbot, not in any future feature.
• It does not predict what a court will decide. Litigation outcomes turn on facts and judges, neither of which an AI can know.
• It does not tell you what to file, when to file it, or what to say. Those are decisions made between a client and a lawyer.
• It does not generate legal documents for use in your case. We do not draft AI complaints, AI motions, AI prenuptial agreements. Document drafting is legal work, not educational content.
• It does not create an attorney-client relationship. Reading or interacting with this site is not a retention.

How AI is used

The encyclopedia entries are generated using Claude Haiku 4.5, an AI model from Anthropic. Each entry passes through a tightly constrained prompt that:

• Establishes the role as an educational reference writer, not a lawyer.
• Specifies the jurisdictional frame (MA, PRC, cross-border, or general) so the entry is grounded in a particular legal system rather than a generic average.
• Forbids the model from using "you should," "in your case," or other phrasing that would drift from definition into advice.
• Forbids fabricated statute numbers or case citations — if specific authority is needed, the entry says "under Massachusetts law" rather than naming a section that may not exist.
• Requires real Mandarin (Simplified Chinese) for the bilingual portion, not transliteration.

Each entry is stored with the model name and generation date, visible at the bottom of the entry. When an attorney reviews an entry, that review and the reviewer are recorded too. The provenance trail is part of the content, not metadata hidden behind it.

What we do when the model is wrong

AI models can produce confidently wrong text. Our defenses are layered:

• Constrained prompts that limit the surface area for hallucination.
• Visible jurisdictional tags so a reader who notices a mismatch can flag it.
• Attorney review for entries before they are marked reviewed.
• An open feedback channel — if an entry is wrong, we want to know and we want to correct it.
• Public model and date attribution, so a reader can decide how much weight to put on a particular entry.

None of these are perfect. Together they're better than the realistic alternative, which is either having no educational reference at all, or pretending an opaque legal-content site has reviewed everything when it hasn't.

The bilingual problem — and why we still publish bilingual content

Translation between English and Mandarin in a legal context is not symmetric. Many family-law concepts in MA do not have an exact equivalent in PRC family law, and vice versa. "Equitable distribution" is not the same as "夫妻共同财产 community property"; "best interests of the child" is not exactly "儿童最佳利益" in the way Chinese courts apply it.

We publish bilingual entries anyway because the alternative — English-only content, or a placeholder Chinese page that's just an MT pass over the English — serves Chinese-speaking families worse. Where translation introduces a real conceptual mismatch, the entry's "Cross-jurisdictional notes" section explains it.

If you read an entry in Mandarin and feel something is off, it might be us, it might be the model, it might be the irreducible difference between two legal systems. Talking to a lawyer fluent in both is the right move regardless.

Why we publish this methodology

Two reasons.

First, because the practice of using AI in legal services is moving faster than the bar associations are responding. The Massachusetts Board of Bar Overseers has not published comprehensive AI guidance as of mid-2026. The American Bar Association's Formal Opinion 512 (July 2024) on generative AI is a reasonable starting point but doesn't cover the structured-content patterns we use here. Practitioners are essentially writing their own rules in public, and we'd rather do that visibly than invisibly.

Second, because we think the standard matters more than the implementation. The way we approach UPL and AI here — jurisdictional tagging, model attribution, review trails, explicit refusal to give case-specific advice — is portable. If another firm wants to copy the pattern wholesale, that's fine. The only thing we'd ask is that they hold the line on the principle: the AI is for education, the lawyer is for advice.

Open questions we're still working on

• Topic chatbots. When we ship topic-specific chatbots in the next iteration, the same rules apply, but the failure modes are sharper — conversational AI is harder to keep on the educational rail than static encyclopedia entries. We're being deliberate about scope and disclaimers.
• Court scenario walkthroughs. Generative narrative is inherently imaginative. We need to keep these clearly marked as illustrative fiction, not as forecasts of what will happen in any reader's case.
• News curation. When we surface external news and add AI-generated context, we'll cite the source and clearly separate the source's reporting from our model's framing.
• Chinese-language UPL. The bar rules we're navigating are written in English by U.S. regulators. A reader in China reading our Mandarin content is in a different regulatory frame entirely. We're working with practitioners on both sides to think this through.

If you find a mistake

The encyclopedia is a living artifact. If an entry is wrong, misleading, or outdated, please tell us — talk to RCXLaw or contact us directly. We will correct, re-attribute, and timestamp the change.

This page itself is a working draft. The methodology will evolve as the regulatory environment, the AI models, and the practice itself evolve. The version you're reading was last meaningfully updated in May 2026.