A model asked something it does not have reliable information for will very rarely say so plainly. Instead it produces an answer with exactly the same tone, pacing, and structural confidence as a fully grounded response, and a caller has no way to tell the difference just from how it sounds. This is one of the most consistently underestimated failure modes in voice specifically, because in a text interface a hedge or a citation gap is at least visually inspectable. On a phone call, tone is the only signal a caller has, and an ungrounded guess delivered smoothly sounds exactly as trustworthy as a verified fact delivered the same way.
The mechanism behind this is not unique to voice, it is the same underlying tendency any language model has toward generating plausible sounding completions regardless of whether the underlying information is actually solid. What makes it a bigger problem specifically in a call bot context is that a caller cannot pause and independently verify anything mid conversation the way someone reading a chat response might quietly do on their own. Whatever gets said on the call is what the caller walks away believing, often immediately acted on, before there is any real opportunity to catch a subtly wrong answer.
The fix requires treating confidence as something the system has to earn explicitly for each response, rather than something it defaults to automatically. This means building an internal distinction into the system prompt between information the model actually has grounded access to, retrieved directly from a knowledge base or confirmed earlier in the call, versus anything it would otherwise be inferring or generating from general pattern completion. Anything falling into that second category needs to trigger a different response posture entirely, not a confident direct answer, but an explicit acknowledgment of uncertainty paired with a path forward, something like I don't have that confirmed on my end, let me connect you with someone who can verify that for you, rather than a smoothly delivered guess dressed up in the same tone as everything else.
Getting this right in practice means being unusually explicit in the instructions about what counts as grounded versus inferred, because left ambiguous, a model will often treat its own general world knowledge as functionally equivalent to information actually provided for that specific deployment, especially on questions that sound like they should have a standard, common sense answer. The instruction has to draw that line clearly rather than trusting the model to intuit it, specifying exactly which categories of question must be answered strictly from provided source material and explicitly forbidding the model from filling any gap in that material with plausible sounding general knowledge instead.
The tone of the uncertainty acknowledgment matters almost as much as the fact of including it at all. Delivered flatly or apologetically, an admission of not knowing something can read as the system being broken or unhelpful. Delivered as a brief, matter of fact redirect toward the right resource, it reads instead as a competently handled boundary, and callers generally respond far better to a confidently delivered I don't have that, but here's how we get it than to either a confidently wrong guess or an awkward, hesitant non answer.
Written by Mohammad Farhan Habib Faraz
Senior Prompt Engineer and Prompt Team Lead at PowerinAI
www.powerinai.com
Top comments (0)