AILegal
A legal AI assistant aimed at people who cannot afford a lawyer; its core design is to store legal norms in a structured form and enforce verification against official sources, preventing citation of nonexistent or irrelevant legal provisions.
Timeline (last 90 days)
Early versions of AILegal used prompting, placing the corpus in the context window and requiring answers based only on the sources, but manual checks found that the assistant would fabricate nonexistent statute numbers (such as 18 U.S.C. § 301) or cite statutes that exist but are irrelevant (such as section 157 of the UK Consumer Credit Act 1974).
AILegal stores each legal norm as a structured Norm object whose fields include a unique ID, statute code, jurisdiction, a human-reviewed plain-language summary, the official source URL, the hash of the version that was checked, the check time, and a verified boolean flag.
AILegal's corpus currently has 53 norms in total, of which 11 are Russian, 26 American, 9 British, and 7 Canadian.
Every three days the AILegal system runs a scheduled task that fetches each official source, requires that the numbers and terms in the summary appear in the fetched text, and stores the hash of the version it checked; if the hash changes, the norm is flagged for manual review.
AILegal's verified flag is never set by a machine; the script can only verify whether the text contains what is claimed, while whether a norm can be used for legal judgment must be decided by a person.
At answer generation time, if a norm has not been verified, the AILegal assistant will not give a deadline or clause number derived from it, but instead clearly states that it cannot confirm it and asks for the missing factual information.
Of the 31 test paraphrases, AILegal's exact-match stage resolved 12; the soft-match stage then added suggestions for another 14, of which 8 were correct, for a precision of about 57%.
AILegal's soft-match stage is used only to generate clarifying questions and never directly outputs legal conclusions.
The AILegal team tried using the local embedding models nomic-embed-text and qwen3-embedding to handle standard paraphrases, and the similarity scores landed at top-3/top-4 with cosine gaps of only 0.001–0.03, at the noise level.
The AILegal team believes the real way to solve the synonym-recognition problem is not to switch to a better model but to have a person write and review the semantic representation (everyday search phrasings) for each norm, which is legal work rather than prompt tuning.
2 more timeline events
All Facts (12)
Early versions of AILegal used prompting, placing the corpus in the context window and requiring answers based only on the sources, but manual checks found that the assistant would fabricate nonexistent statute numbers (such as 18 U.S.C. § 301) or cite statutes that exist but are irrelevant (such as section 157 of the UK Consumer Credit Act 1974).
50%UnverifiedAILegal stores each legal norm as a structured Norm object whose fields include a unique ID, statute code, jurisdiction, a human-reviewed plain-language summary, the official source URL, the hash of the version that was checked, the check time, and a verified boolean flag.
50%UnverifiedAILegal's corpus currently has 53 norms in total, of which 11 are Russian, 26 American, 9 British, and 7 Canadian.
50%UnverifiedEvery three days the AILegal system runs a scheduled task that fetches each official source, requires that the numbers and terms in the summary appear in the fetched text, and stores the hash of the version it checked; if the hash changes, the norm is flagged for manual review.
50%UnverifiedAILegal's verified flag is never set by a machine; the script can only verify whether the text contains what is claimed, while whether a norm can be used for legal judgment must be decided by a person.
50%UnverifiedAt answer generation time, if a norm has not been verified, the AILegal assistant will not give a deadline or clause number derived from it, but instead clearly states that it cannot confirm it and asks for the missing factual information.
50%UnverifiedOf the 31 test paraphrases, AILegal's exact-match stage resolved 12; the soft-match stage then added suggestions for another 14, of which 8 were correct, for a precision of about 57%.
50%UnverifiedAILegal's soft-match stage is used only to generate clarifying questions and never directly outputs legal conclusions.
50%UnverifiedThe AILegal team tried using the local embedding models nomic-embed-text and qwen3-embedding to handle standard paraphrases, and the similarity scores landed at top-3/top-4 with cosine gaps of only 0.001–0.03, at the noise level.
50%UnverifiedThe AILegal team believes the real way to solve the synonym-recognition problem is not to switch to a better model but to have a person write and review the semantic representation (everyday search phrasings) for each norm, which is legal work rather than prompt tuning.
50%UnverifiedThe AILegal team openly invites users to try to break the system: take any norm, compare it with the official source it links to, and try to get the assistant to state something that is not present in the source text.
50%UnverifiedThe AILegal product does not constitute legal advice; the legal content supports Russian and English, and the interface supports more languages.
50%