Researchers have built a tool that guesses how likely a chatbot's answer is to be correct - without ever looking inside the model.
The system, called Pinocchio, is an external calibrator trained on outputs from seven large language models. It reads a model's response and, in a single forward pass, produces a confidence score for whether that response is correct. The researchers report an AUROC of 0.862 on held-out responses from the same seven models, and say the calibrator transfers with zero additional training to thirteen other models from eight organizations it never saw during training. A stripped-down, text-only version with just 0.8 billion parameters matches the accuracy of their largest checkpoint, and the team says it can be added to an existing pipeline in two lines of code.
Most uncertainty-estimation methods need access to a model's log-probabilities or the ability to fine-tune it - both routinely blocked by closed API providers like OpenAI, whose GPT models don't expose token probabilities. Pinocchio sidesteps that by treating the target model as a pure black box, which matters for any product built on rented API access rather than an in-house model. That is a real gap: plenty of companies now bolt LLMs onto decisions - approving claims, flagging fraud, drafting legal text - with no reliable way to know when the model is guessing.
An AUROC of 0.862 is decent, not oracle-grade - it will flag some right answers as risky and wave through some wrong ones, so treat it as a second opinion, not a lie detector.