Check with a second model

Two copies of one model make the same mistakes at the same time. In the Michigan MiDAS case cited below, one uniform rule set produced tens of thousands of correlated wrongful fraud determinations, one flaw repeated at the scale of the caseload. A second model from another vendor or another design disagrees where the first is wrong in a way the second is not. The research cited below also finds that a model rarely corrects its own answer once it has committed to it. The check therefore has to come from outside the model.

In the Lab, this lever strengthens the pathway by which models cross-check each other's output. It also weakens the pathway by which AI models or agents relay failures to one another. Its side effect raises operator deference drift, because two systems that agree look like proof.