AI Can't Unmask Secret Offshore Owners, Study Finds

A researcher at a computer science lab took on a tempting question: could AI crack open the world's biggest financial leaks? The Panama and Paradise Papers exposed thousands of offshore companies set up to hide who really owns what. The public version of those leaks contains about 814,000 companies and people, plus 84,000 ownership links that were already known. The study asked whether a computer could fill in the missing links — the ones that stayed hidden.

The answer is a hard no, at least for the hardest cases. The paper's core idea is simple. If a person shows up only once in the whole dataset, linked to nothing else, there is nothing to tell them apart from any other stranger with the same shape in the records. So any AI that plays fair must guess, and guessing means a 50/50 coin flip. Those invisible owners make up about 20.6% of all owners. A trained AI landed on exactly that coin-flip score, and the same floor appeared in every one of the five leaks examined.

Along the way, the paper flags five common testing mistakes. Each one makes a chatbot or model look like it beat a limit that actually cannot be beaten — a bit like grading your own exam with the answer key in hand. The researcher offers a short rule to avoid each trap.

The practical takeaway flips the usual goal. Instead of chasing the hidden owner, which is close to impossible, the study says investigators should study the machinery of concealment: the shell company patterns, the lawyers, the middlemen. It also warns that bolting on outside data helps far less than people assume. For tax authorities and journalists, that is a useful reality check about what AI can and cannot do.