Why imported legal AI gives confidently wrong answers in Bangladesh

Bangladesh inherited a common law vocabulary. Consideration, estoppel, specific performance and injunction all mean something here. That shared vocabulary is precisely the trap: a model trained on English and American material recognises the words, produces a fluent paragraph, and silently imports the wrong rule underneath them.

Ask about ending employment and you will be told about at-will termination, which does not exist here. Ask about a bounced cheque and you will get American commercial paper doctrine instead of a criminal complaint with a thirty-day notice. Neither answer looks wrong. Both are.

Substantive law can sometimes survive translation. Procedure never does. Court fees, limitation, the correct forum, whether a matter goes to a Joint District Judge or an Artha Rin Adalat, what has to be deposited before an appeal is entertained — none of this is derivable from general legal knowledge, and all of it decides whether a filing survives its first hearing.

A legal assistant becomes safe the moment it is required to point at the provision it relied on. Citation is not decoration — it converts an unverifiable assertion into a claim you can check in under a minute, and it makes the model's ignorance visible instead of letting it be papered over with fluency.

That is the whole design principle behind AttorneyAI, and it is also the reason we say plainly that output is a first draft for a qualified person to check, not advice. A tool that claims more than that is selling you its confidence, not its accuracy.