The content policy of the upstream provider message

WhatMatters

It's the first time it happened, so maybe someone here can shed some light on this? Consider a prompt that's starting a new chat:

I've heard they had this mosquito chamber test. Basically a chamber filled with 500 hundred freshly hatched mosquitos and you stick a hand in there and a scientist counts the number of mosquitos that landed on your hand and the number that decided to feed and they calculate that ratio and give you a comprehensive "true age" analysis. Then they show you to a dog, if a dog refuses to smell you, that's -1 point for your worthiness as a decent human being.

And Claude 4.5 Sonnet with customized instruction set starts typing an answer when all of a sudden...

We are sorry, the input prompt triggers the content policy of the upstream provider. Please modify your prompt and try again.

Re-submitting the same prompt and the machine happily drafts an answer which as far as I could see is different from the original one it tried to deliver. But I didn't have time to read the original one because it disappeared so fast. What is that error?

My next reply when I got the LLM's reply also got blocked with exactly the same message.

WhatMatters

I also started noticing recently how the text the machine is outputting in real time is being changed. After a couple of sentences are already on screen, those may get erased and replaced with other text.

yiwei-1

We are sorry, the input prompt triggers the content policy of the upstream provider. Please modify your prompt and try again.

This error message means that the model provider (Anthropic, in this case) rejected your request because it was deemed to violate their model usage policy. This can occur either before the model begins generating a response or midway through, if the provider determines that the output would violate their policy.

WhatMatters

Right. Can I have a link to said policies, if possible? Ultimately I would like to see LLM policy on Kagi's end and upstream providers's policies as related to Kagi. Because obviously I'm not interacting with Claude directly on Anthropic website, so there two parties besides me engaged in this, meaning the entire policy must be somewhat complicated. Any published documents I can look at?

fxgn

WhatMatters It's just Anthropic's regular usage policy, which includes access "via any authorized resellers or passthrough access"

https://www.anthropic.com/legal/aup

WhatMatters

Thanks @fxgn
But isn't there a separate policy/agreement for Kagi's side of the LLM equation? Would like to see that too.

Second, it used to be that before models would simply reply with "I can't, I shan't and I won't" and now somehow they start typing away and that stream is being interrupted. Is that something new? I'm trying to understand the mechanism behind all those filters, how they used to be and how they are now.