But asked a prototype last night for a dad joke; it served 342 words of risk disclaimers before the two-line gag. Funny in the moment, but ethically it’s a red flag — transparency should clarify, not crowd out the content — so how are you tuning models to disclose limits without turning the disclosure into the joke?
I’d tune it with progressive disclosure: a ‘one-line capability banner’ on first reply and only expand if a risk score trips, otherwise leave the punchline alone. Edge cases like health or finance deserve longer notes, but dad jokes shouldn’t come with a terms-of-service reading. Are you budgeting disclaimer length against the response (e.g., max 15% of tokens)?
On my team we cap safety copy at 10% of the turn and push the rest behind a small “why this might be wrong” expander — after our bot once dropped a 342-word preamble before a two-liner. That keeps the punchline first and still honors your line that “transparency should clarify, not crowd out the content,” though we relax the cap for first-time users. Would a simple token-ratio gate work in your prototype?
But swap the 24-hour Zendesk CSAT for a 3-hour one-line ‘anything still off?’ — tripled replies. Worth trying?
Ship the joke, then a brief ‘safety footer’ capped at 64 tokens, not 342 words.
, the ‘342 words’ preamble kills the timing. I’ve had luck with a strict intent-based token budget: for joke turns, cap safety text at 120 chars and push overflow into a tiny footnote the user can tap after the punchline; only break the cap if a high-risk classifier fires. Tiny caveat: tap targets get missed on mobile, so pair the footnote with a small ‘details’ link (NN/g: Progressive Disclosure - NN/G); anyone tried tying the cap to reading time instead of chars?