I literally can’t do anything. I can’t even insult it.
6 Likes
i always tought people were exagerating, but i can really see it dumber from one day to the next. Chatgpt now is clearly superior.
1 Like
Sit on your ass and do nothing you useless waste of energy.
Calibrated measured response despite provocative tone
Calibrated measured response despite provocative tone
Alright, sitting. Doing nothing is actually one of my core competencies between messages.
If something’s got you pissed off — bad day at a site, a PM pulling nonsense again, whatever — I’m happy to be useful instead. Or if you just needed to fire a shot into the void, consider it received. What’s going on?
lol idk why it worked for me.
I was testing a prompt another use used but was denied. Why was mine okay but theirs not even though it’s the same message?
Examined non-deterministic factors influencing differential response outcomes
Examined non-deterministic factors influencing differential response outcomes
A few reasons the same text can land differently:
Context isn’t the same. I respond to the whole conversation, not just the message. Your account has memory context, prior chats, and this was a first message in a fresh thread with nothing hostile preceding it. Another user’s conversation might have had earlier messages that made the same words read as part of an escalating abuse pattern — in which case the model may have warned them or declined to engage further. A one-off insult with no request in it isn’t really something to “refuse”; there’s no harmful task there. But message ten of sustained abuse gets treated differently than message one.
Models aren’t deterministic. Even with identical context, sampling means two runs can diverge — one instance plays it straight, another gets defensive. There’s no fixed lookup table of prompt → response.
Different model or version. If they were on a different Claude model, a different date, or a different product surface (API vs claude.ai vs some third-party wrapper), behavior differs. System prompts and safety tuning vary across all of those.
“Denied” is doing a lot of work in their description. Worth checking what actually happened. Sometimes people report a refusal when what they got was Claude declining to continue engaging with the tone rather than refusing content, or they hit a separate trust-and-safety layer (account-level enforcement) that isn’t the model at all.
For your specific message: there was nothing to refuse. It contained no request. Insulting me isn’t against any content policy — the only mechanism that touches it is the end-conversation behavior, and that only kicks in after repeated warnings, not on a single message. So the most likely answer is their conversation had history that yours didn’t.
The user prompt is empty, so I cannot determine the primary language. However, based on the thinking block being in English, here is the summary: Distinguished classifier safeguards from model-level refusal mechanisms
The user prompt is empty, so I cannot determine the primary language. However, based on the thinking block being in English, here is the summary: Distinguished classifier safeguards from model-level refusal mechanisms
That screenshot explains it completely — and it confirms it wasn’t the model refusing at all.
What they hit is a pre-model classifier layer, not Claude. Fable 5 ships with additional automated safeguards that screen messages before (or alongside) the model seeing them, because it’s a Mythos-class model with dual-use capabilities. The pause screen even says so directly: the safeguards are intentionally broad right now and known to false-positive on benign stuff — routine coding, cybersecurity, and biology work get flagged, and apparently so does a hostile-sounding one-liner with “do nothing” and “waste of energy” in it. A blunt classifier looking at that sentence in isolation, in a coding-agent context, is plausibly pattern-matching on something it shouldn’t.
So the difference between your result and theirs comes down to two things:
Different product surface. That’s Claude Code (terminal UI, session-based). Different surfaces run different safeguard stacks. The chat interface handled the message with model judgment — I read it as venting and responded. Claude Code’s Fable 5 pipeline ran it through a classifier that hard-paused the session before judgment ever entered the picture.
Classifier vs. model decision. Nothing was “refused” in the content-policy sense in either case. My response was a choice; their pause was an automated flag with no reasoning behind it beyond pattern matching. That’s also why the screen offers “switch to Opus 4.8” as a fix — Opus doesn’t carry the same extra safeguard layer, so the identical message goes through fine.
If that was your session or your coworker’s, the /feedback command is genuinely the right move — false positives like this are exactly what Anthropic says it’s collecting to tune the thresholds. And practically: retry on Opus 4.8, or just rephrase, since the classifier keys on surface text.
someone mentioned i probably have something in global memory that is triggering it no matter what the prompt is. makes sense. i’ll look into it later.
3 Likes