I think the framing as “lying” is emotionally accurate for users but technically misleading for builders. Most failures I see in production aren’t intentional deception, they’re confidence calibration failures: the model sounds certain when it should sound tentative.

What helped us most was treating outputs like untrusted input: retrieval + citations for factual claims, strict tool permissions, and an explicit “I don’t know” path that’s rewarded instead of punished. That doesn’t make LLMs truthful, but it does make them a lot less brittle.

So to me the key question isn’t “are models liars?” but “what product and engineering constraints make wrong answers cheap and detectable?”