Skip to main content

Boardroom Answers · AI & Data · AI, Data & Analytics

AI Hallucination Liability: Who Is Accountable for the Loss

The question a Chief Artificial Intelligence Officer (CAIO) asks: If a board acts on a hallucinated number from your platform and loses money, who is liable — and can you prove, after the fact, that the number was not invented?

The short answer

A number that cannot be traced to your data is rejected before it is ever saved — and every output that does ship carries a forensic trail of model, sources and reasoning, with a named human owning the decision.

The full executive answer

Let me answer the proof part first, because it is the stronger answer. In our engine, a number that cannot be traced to input data is not published — full stop. Every generation passes through a numeric provenance validator: each figure in the AI’s prose must exist in the data it was given or in its own structured numeric fields. If a figure fails, the system runs exactly one corrective retry; if it fails again, the entire output is rejected and nothing is saved — the user sees a message saying the analysis produced untraceable figures and must be regenerated. There is no code path where an unsourced number reaches a boardroom silently. And after the fact, every insight is stored with its model, its provider, its token counts, its cited sources and its cache key, in an append-only audit log — so yes, we can reconstruct exactly what was generated, from what, by which model, when.

On top of that, the highest-stakes modules — the scorecard, risk register, M&A diligence, boardroom mode and others — get a second opinion from a different provider’s model. The two structured outputs are diffed deterministically; when agreement drops below a threshold, a third model adjudicates each disagreement with written reasoning, and the disagreement plus ruling is displayed to the user. Disagreement becomes visible rigour, not hidden risk.

On liability: Vouli IQ is decision support, and we say so contractually and on every output. The recommendation carries what we call the Vouli Verdict — a proposed board motion that a named human must own. That is the same accountability model regulators already apply to model risk in banking: the model informs, a human decides, and the institution documents both. Under the EU AI Act, decision-support with a human accountable is limited-risk with transparency obligations — which we meet by labelling AI output and showing provenance — not an automated decision system.

What I will not claim: no vendor can guarantee zero hallucination in narrative text — anyone who tells you otherwise is the one hallucinating. What I can show is defense in depth: provenance gate, hard-blocking consistency guardrails, cross-provider consensus, third-model adjudication, framework-citation validation, and a full forensic trail. That is more layers than most banks put around spreadsheets that boards act on today.

Grounded in: Federal Reserve SR 11-7 (model risk management — human accountability, documentation, validation); NIST AI RMF 1.0 GOVERN 1.3 and MANAGE 4.1; EU AI Act limited-risk transparency obligations.

Want this answered live, on your data?