A general chatbot invents case law because that is what it was built to do: predict fluent text. It is not consulting a library. Citation-grounded AI is a different architecture — and the difference is exactly the part lawyers care about.
Why fluency produces fabrication
Large language models generate the most probable next words. Asked for a supporting citation, the most probable output is something shaped like a citation — correct format, plausible parties — whether or not the case exists. Fluency and truth are not the same objective, and when they diverge, fluency wins.
Grounding, in practice
Citation-grounded systems retrieve the relevant passages from your matter first, then write an answer constrained to that retrieved evidence — and attach the source to every claim. If there is no supporting passage, there is no answer.
No source, no answer. The citation isn't decoration; it's the gate.
So a reference like § Ltr 03-14 is not generated text — it is a handle to a real document and page you can open in one click.
Christie cites every answer to your own files.
Get early access →What this changes for your workflow
You stop verifying whether the AI told the truth and start verifying the underlying source — the same thing you would check anyway. The tool does the retrieval; you keep the judgment. That is the only arrangement that belongs in a filing.
