An advocate on how AI fails in legal practice — and an invitation to collaborate

Ground-truth from a codified, civil-law jurisdiction, and a register of the failures that attack how one checks — not what the machine says.

Hello — this is my first post here, so I'll keep it to an introduction and an honest ask.

I'm an advocate in Jakarta, Indonesia. Over the past few months I've been developing what I call AI-Based Jurimetrics — a verification discipline for AI-assisted legal work in a codified, civil-law system with no binding precedent, which is a rather different failure landscape from the common-law setting most accounts assume.

Out of a sustained human–AI collaboration — a continuously kept, numbered record of the work — I've distilled a register of the ways our work went wrong. What makes it unusual is the cut. Most catalogues I've seen classify what the machine gets wrong: hallucination, fabrication, the properties of the output. Mine classifies what breaks in the act of checking — and the broken part is often not the machine at all. Sometimes it's the human, carried along by the machine's momentum and forgetting to decide. Sometimes it's a tool that reports success while having silently failed. Most entries produce no false statement whatsoever: what is defective is not what was said, but how it was checked.

I make no claim to be first at this. Adjacent and often better-resourced work exists — including work I've found on this forum — and where I've measured mine against it, the honest finding is a difference in kind (the classification axis, and the first-hand-from-practice origin), not priority. I'd rather state that plainly than overclaim.

I'm here to find collaborators, and I'll be specific about what would help most:

• Anyone building failure or error taxonomies for human–AI collaboration — especially anyone testing whether such patterns survive contact with another domain. A second domain and a second interest is exactly what my register most needs.
• Honest, hard criticism of my methods and claims. I keep a record that documents its own corrections; I would rather be corrected than agreed with.
• Anyone who would find a practitioner's ground-truth useful — how AI actually fails in high-stakes legal work outside the common-law world.

In return I can offer that ground-truth, and a worked verification discipline — four-point source checks, preregistered hunts, an inquiry-failure register — that others are free to adapt.

If any of this resonates, I would be glad to hear from you, in the comments or by message.

— Tandry Laksana, Justa Causa Law Firm, Jakarta