A new preprint argues the reason chatbots sound certain and still get it wrong is not a glitch, it is missing plumbing.
The paper, posted August 17 on arXiv, proposes a framework for when an AI system's answers can actually be checked and corrected, rather than just sounding plausible. It argues an output only has real "answerability" when a live, independent channel, like a database lookup, a tool call, or executed code, can catch and fix a mistake as it happens. Text-only language models mostly work off patterns of correction absorbed during training, which the authors call inherited rather than live answerability. Based on that split, they predict models will produce fluent, wrong answers specifically on tasks that need direct access to current facts, and that fixes like retrieval, self-consistency checks, or code execution will only help on the tasks where they actually supply that missing live check.
That is a more precise way to explain something builders have argued about for years: why retrieval-augmented generation and tool use cut hallucinations in some products but barely move the needle in others. Instead of treating "grounding" as a single fix, the framework says to ask which specific correction channel a given task needs, and whether the model actually has it.
It is a conceptual framework, not a benchmark, so the real test is whether it predicts new failures better than the educated guesses teams already make about where their chatbots will go wrong.