Generative AI can invent case law that does not exist. That risk is the single thing keeping consumer legal AI out of real cases. Lawma is built so a licensed attorney owns every legal output — and the AI works under six guardrails designed to make it safe to do so.
The pipeline
Each guardrail is independently testable. Together, they make AI-assisted drafting a tool an attorney can safely sign their name to.
A team of Lawma attorneys controls the legal corpus the AI is grounded on — only pre-approved statutes, rules of court, Judicial Council form instructions, treatises, and case law that has cleared review. The same team monitors new opinions, regulatory updates, and supersession events so the RAG corpus stays current. The model never reaches for public-web case law unless an attorney has put it in front of the system first.
Legal drafting runs under retrieval-augmented generation (RAG) over a verified California legal corpus — statutes, rules of court, Judicial Council form instructions, and case law. The model completes a disciplined template against retrieved authority rather than free-associating. The supervising attorney sees the retrieved set and signs off before anything leaves DRAFT.
Before an attorney can approve a draft, every authority cited is checked for current standing — including with web search against official sources where the supersession status isn't already known. Anything flagged as superseded, overruled, or stale is surfaced in-line so the attorney can see what was considered and rejected.
Statutes link to leginfo.legislature.ca.gov. Forms link to courts.ca.gov. Case law links to CourtListener with full pinpoint citations. Every assertion is one click from the authority that supports it.
The attorney sees the structured trail — facts extracted from intake, issues identified, authority retrieved, current-law check, and the drafted conclusion — not just the polished output. Hallucinations get caught because the reasoning is visible to a licensed human, not because the attorney has to trust a black box.
Nothing drafted reaches the client or the court until a licensed attorney reviews and approves it. The attorney can edit inline, and every edit is captured in a per-document audit trail — when, by whom, and what changed. The DRAFT watermark stays on until approval; CRPC 1.1 (competence) and 5.3 (supervision of non-lawyer assistance) are operationally enforced.
The structural moat
Every direct-to-consumer legal AI platform sits behind the same regulatory ceiling — no attorney-client relationship, no privilege, no specific advice. That's not a marketing gap; it's the line state bars draw against unauthorized practice of law.
Without that relationship formed first, an AI tool has to keep its answers generic. Specifics expose the operator to unauthorized-practice exposure — the same family of ethics rules that closed Avvo Legal Services in 2018 and underwrote the FTC's order against DoNotPay in 2025. Lawma sequences this differently: an attorney accepts the matter first, so the AI works under that attorney's supervision and can deliver the case-specific drafting and analysis the rest of the lane legally cannot. The client's communications are actually privileged.
The competitive landscape and the verbatim disclaimers from every direct-to-consumer platform live in the Competitive Landscape memo.
In the tool
The attorney's case-analysis page exposes the AI's reasoning trail end to end — facts extracted, issues identified, authority retrieved, current-law verification, and the conclusion drafted strictly from the verified record. Every citation is hyperlinked to the official source. Any authority the verification step flagged as superseded or stale is shown in-line, so the attorney can see exactly what the AI considered and rejected.
Five-step pipeline visible: facts extracted → issues identified → authority retrieved → current-law check → conclusion drafted. Each step expands in-line.
Every citation opens the official source. Caution chips appear on any authority the verification step flagged as superseded — like the Davis date-of-separation case now overridden by Family Code § 70.
When an output isn't right, the attorney flags it — the interpretation, the citation choice, the prompt structure — and the next draft adjusts. The supervising attorney is the loop, not a downstream stakeholder.
Attorneys can edit any AI-drafted text in place. Every edit is captured with the attorney, the timestamp, the prior text, and the new text — the complete edit history is visible per document.
See it live in the prototype. The attorney case-analysis page for our sample client walks the full flow — reasoning trail, hyperlinked authorities with caution chips, the “Give Lawma feedback” control, and the per-edit audit trail. Open the live attorney view →
Why this matters
The hallucination problem is not theoretical. It is the reason regulators, judges, and state bars have moved on consumer legal AI in 2024–2026.
FTC Action · February 11, 2025
The FTC finalized its order against DoNotPay on February 11, 2025 (Commission vote January 16, 2025) requiring $193,000 in relief, mandatory consumer notice, and a prohibition on claiming lawyer-equivalence without evidence. The order set the regulatory floor for the unsupervised consumer-legal-AI lane.
State-Bar Discipline · 2026
California State Bar disciplinary charges were filed in 2026 against Dennis Block, a high-volume Los Angeles eviction-firm attorney, over a 2023 eviction-court filing containing fabricated case law generated by AI. The judge called it “an entire body of law that was fabricated.” The matter is a leading example of how AI-as-lawyer use surfaces in regulatory enforcement when no supervision pipeline catches the hallucination.
Federal-Court Sanction · 2023, still cited
In Mata v. Avianca (SDNY 2023), a federal judge sanctioned attorneys $5,000 for filing a brief containing six AI-generated cases that did not exist. The matter became shorthand for AI hallucination risk in legal practice and is now cited by state bars across the country in AI-use guidance.
