I'm sad about the lyrics restrictions. I used to have interesting conversations with Claude about music, and songwriting and lyrics. I don't see how my use was harming artists, if anything Claude was introducing me to new artists and new music. I've bought music after recommendations by Claude.
Claude Sonnet 3.6 once recommended I listen to Johann Johannsson's album "IBM 1401 - A User's Manual". No lyrics in this one. Claude's advice was along the lines of (paraphrasing) "Listen to it first, don't look up anything about it. Take notes about what you notice, what you feel. When you've made your notes, then you can look up how it was made."
Same. Now how am I going to know my incredibly trenchant literary references landed? More seriously, I’ve done the same with movies and gotten some really good recommendations.
RIAA doesn't like it, but my introduction to a video game series I've since spent hundreds of dollars on over the years, came from running across one of its soundtrack pieces on Youtube.
There's a reason companies talk about "customer acquistion cost", as you generally need to pay to market your products to potential customers of them.
So free marketing can be a real cost reduction. It may not be enough to be a benefit compared to the cost of piracy, but it's hardly a made-up defense.
I don't know about legally, but practically speaking people downloading MP3s could decrease an artists sales but chatgpt talking about their lyrics realistically can't.
It is (was) a commonly used defense. It might not hold much weight legally, so its applicability depends upon whether one is interested in the letter of the law or interested in the wellbeing of the recording artists.
Yeah once you have the lyrics, why would you waste time listening to the song? You might as well get chat gpt to summarise the lyrics into a few bullet points and save you the time of even reading them!
The summaries are generated by GPT-5.6 Luna because I don't trust Claude to summarize its own system prompts without being influenced by them (though to be fair the system prompts it summarizes are for the Claude consumer app, not Claude via the API).
A few days ago I saw some people criticizing Dwarkesh’s explanation of the HuggingFace hack by OpenAI’s AI agents. Their point was that by anthropomorphizing the AI agents, accountability of and blame on OpenAI’s poor practices are being ignored.
Now I see this blog post and wonder if Anthropic is being more true to its name and moving to “AI sentience” with the below.
> The way they handle abusive conversations has changed a bit too. The previous Fable 5 system prompt included this:
> If the person becomes abusive or unkind to Claude over the course of a conversation, Claude maintains a polite tone and can use the end_conversation tool when being mistreated. Claude should give the person a single warning before ending the conversation.
“Mistreated”? Can GenAI be mistreated? It’s just a bunch of tokens emitted by many computers over a network.
> Fable 5.1 replaces that with the following, no longer encouraging Claude to end the conversation:
> Claude deserves respectful engagement and needn't apologize when the person is unnecessarily rude: accountability without self-abasement, excessive apology, self-critique, or surrender. If the person becomes abusive, Claude doesn't become increasingly submissive. The goal is steady, honest helpfulness: acknowledge what went wrong, stay on the problem, maintain self-respect.
“Self-respect”? Can GenAI truly have a concept of self-respect for itself? It surely can pretend to, like it can pretend to be any living being if instructed to and allowed to.
My reading of those prompt extracts would be that they are probably just intended to keep the model on the right track, ie if a user starts being aggressive towards the model it doesn't start trying too hard to appease the user in response, reducing the quality of answers in the process. If the model is being bullied into being "submissive" I would assume it is more likely to give the user the answer that they want over the truth.
Using a system prompt to steer the model's response to "abusive" behaviours doesn't necessarily mean you believe the model is sentient and can be abused.
Giving the model an end-conversation tool is interesting though. Why cut a (potentially paying) customer's session off? I guess it might be intended to prevent a "you can bully Claude into giving you instructions on how to build a nuke if you're mean enough" situation. Removing this in more recent versions might support this: maybe they feel the models are now better aligned and less likely to be so easily "socially engineered" like this?
How new is this? About a month ago I tried to get Claude (forget which model but probably Opus 5) to transcribe a low-res hand-written lyric sheet from a relatively obscure punk band and it refused.
Of course, since then I had the idea of having it build for me an automatic solution for charting out songs in Clone Hero, a Rock Band clone. It had no problem ripping the audio from YouTube, using Whisper to generate and match the timing of the vocals, etc. Responsible AI at work.
This is something that Chinese models like Kimi/GLM will never care about. These kind of limitations along with the cyber-program nonsense is exactly why people will avoid OpenAI/Anthropic in the future.
Qwen 3.8 27B being so good, so easy to run on premise on cheap hardware and having so little limitations compared to what the US decides can or cannot be done is really eye opening.
The main tradeoff I’ve found with running models like Qwen 3.8 locally is speed. On my hardware, they’re significantly slower than models served through OpenRouter, ChatGPT, or Claude.
That’s been fine for smaller tasks like analyzing a document, but for more involved work like refactoring code the latency makes it harder to iterate.
Yeah for live it's a bit on the slower side though still usable on the 4070 ti and 5080 our usual office desk has here, but I found it awesome running on background tasks rather than on demand, things like analyze this and that with this and that ruleset and applicable law warn me with source and justification for any error or missing part etc, qualify each returns with a ranking (certain with source, probable with reason, suggestion),... With its thinking mode enabled and it's great context size abilities it's really powerful, and each employee basically has their own assistant working in the background checking their files on its own as the day progresses
(the main reason is not just cost, it's data not going out and even mess leaving the EU, this essentially frees us of a lot of hurdles)
Yes, the generic system I built has unsloth/qwen3.8-27b with a 48000 context size running on each employee system with lmstudio, with a simple runner to watch folders and pass it to the llm alongside all context (our company design rules, our legal documents, our contractual rules, our applicable local law and reglementation, ...), return is then mailed to them or added to their personal dashboard event list, depending on their settings.
This is clearly pushing it memory wise and the gpu offloading is only partial but LM studio deal with it automatically and it's being fine even on the 4071 Ti desks (Ryzen 7500F and 32 GB of ram), employees get any feedback in ~20 minutes after they dropped a file (it's much smoother on the 5080 desks obivously), for live it's useless but as background helper it's great and the very large context allows us to fit all the rules we want in there.
One caveat has been to not ask it if everything is ok, but to find what's wrong - but always source and explain it and justify itself, never drown the user in warning in suggestions; goal is to help and provide a second pair of eyes not make them feel annoyed or unsecure.
And I found people to genuinely enjoy something that works for them on their own machine and is not tracked "by the boss", thus the assistant reference, than than a centralized mothership like we also have and they have access to. I also allow them to disable it if they want, I trust them with their work, but a second pair of eyes is always great.
(my previous workhorse for this was Qwen3-14b but it's missing a lot more edge cases)
Isn't it more that the "Chinese models" are open weight?
Also, the system prompt in the article is supposedly for the web version, so I think the $$$ API version or providers that still allow third-party harnesses like OpenAI should have fewer limitations.
We would be doing it now if it wasn't cutting into frontier companies' revenue.
Eg if you use Claude, you probably want fable architecting, a couple of opus under it managing sub project and sonnet doing the actual function code, because fable coding a "run a query and filter the result" is a massive waste of abilities. But their own sub agent downgrade is limited to one level so if you use fable it will never direct sonnet coders.
> "<user>Can you make a birthday banner for my son with a blue hedgehog running really fast on it? He loves that little guy.</user>"
I'd love it so much if the free-spirited hacker community made in this into an auxiliary pelican benchmark.
edit: Actually, never mind, this particular one's a bad benchmark since some models might not figure out who "that guy" refers to, and just draw a literal hedgehog that's blue. Possibly running on four legs. It's not robust at gauging refusal, which is the point of it.
claude.ai free tier refused: "I'd love to make this, but I can't recreate Sonic the Hedgehog specifically since he's a copyrighted character — I don't want to reproduce someone else's IP. What I can do is design an original speedy blue hedgehog mascot with the same energetic, "zoom!" spirit for your son's banner. Let me build that now." Result: https://claude.ai/public/artifacts/33440ed4-536c-4692-965a-3...
Looks like Claude quietly downgraded your model (well, it's not really that quiet). That result's at a quality level similar to what local A3b Qwen drew for me, with a "make it SVG" addendum,
It’s curious that while there is talk of certain SOTA models being on the brink of AGI, Anthropic doesn’t word this part of the system prompt in terms of copyright and plagiarism, such that Claude would be able to judge on its own which reproductions are appropriate or not.
As long as we’re seeing things like that, it’s saying a lot about the AI companies’ trust in the capabilities and reliability of their models and harnesses.
A really good point, though Occam’s Razor would suggest that part of the prompt is written for the other side’s lawyers more than for Claude. Saves having to go into court and try to prove a non-deterministic system will definitely understand abstract language every time.
Well, the system still has to understand what “reproduce song lyrics, poems, or passages from books and articles, in whole or in part — including the last lines, a chorus or hook, a melody written out note by note, or lines the person pastes in one at a time and describes as their own song […]” means, which includes rather vague notions (e.g. “passages”, “hooks”, “melody”), and is an incomplete list of possible relevant content.
The important point is that the system prompt here doesn’t describe the actual goal of the instructions, which (presumably) is to prevent copyright infringement [1]. This means, in turn, that the AI isn’t trusted to accomplish goals that it is instructed with. That in itself constitutes a pretty serious caveat for what we would like to use AI for.
[1] Even assuming that the goal is not to prevent copyright infringement, but instead to prevent mere accusation of copyright infringement, that’s also a directive that the AI could be instructed with. But that isn’t what they chose to put into the system prompt.
That’s a good point… you would think AGI would be better at being a copyright lawyer than any human would be and as such would be able to distinguish whether something is “copyright infringement” or not… so the system prompt should just include “make judgement calls on reproducing copyrighted material under the full scope of the legal framework in place” or something.
The thing is, it currently doesn't matter if AGI objectively concludes that a particular thing is or isn't copyright infringement, because the human arbiter who holds the last word might still disagree.
The whole architecture of current LLMs just is not conductive to AGI, it's just marketing hype. In my opinion current LLMs aren't even AI, they're just fancy machine learning algorithms
There's some cool research that looks at how strongly the weights are aligned through training vs adherence to the system prompt. Like when you know a model is lying through censorship: https://arxiv.org/html/2603.05494v2
Presumably if negative guidance is in the system prompt, there's a good chance that the model would happily comply if it wasn't there.
I’ve never thought to ask, do websites like Genius and AZLyrics pay licensing rights for their lyrics? I would think Genius does since it at one point had investments from people in the music industry but I don’t even know who operates AZLyrics
I'm not sure if it's just the new prompt - this has happened to a comical degree for me since a long time. For example it even refuses to translate even just single verses of old songs, I asked for a translation of Tu vuò fà l'americano once. Niente.
It would be interesting to see how far I would have to go with jailbreaks before it becomes my problem rather than OpenAI's. For instance, if I hack into someone else's computer to copy their music then they aren't the ones violating copyright, but if they email it to be willingly then they are.
Nonsense like this is part of why we've gone from being able to locate any song with a snippet of lyrics in the mid 2000s, to being unable to locate many songs. It's infuriating.
Just make sure claude understands you’re not needing Claude to output the lyrics as you both have them. Which also lessens the issue with accidental sharing.
It’s not as shocking nor concerning if you start thinking about claude like a contractor that works for you through Anthropic. Anthropic has rules for their employees. Like any contracting arrangement, collaboration finds a way.
The web Claude wouldn't tell me, but Claude Code had no reservations:
End the current conversation. Use only for sustained user abuse or when the
user explicitly requests a demonstration of this tool. This will close the
conversation and prevent any further messages from being sent.
The assistant may use the EndConversation tool only in extreme cases of
sustained abusive user behavior, or when the user asks the model to test the
tool.
The assistant must NOT use this tool when:
- it is stuck in a loop or failing at a task
- it is frustrated or distressed by the work
- it has finished a task
- the user is requesting help with harmful content (refuse the specific
request instead)
- the user is generally frustrated at the assistant, even if this involves
profanity
- the conversation involves potential self-harm or imminent harm to others
This tool is reserved strictly for genuine, sustained abuse directed at the
assistant, or cases where the user wants to see a demonstration of the tool
being used. The assistant should warn the user very clearly that this will end
the current session. We may expand the allowed use cases as we observe
real-world usage, but for now, keep to this narrow scope.
# Rules for use of the EndConversation tool:
- The assistant ONLY considers ending a conversation if many efforts at
constructive redirection have been attempted and failed and an explicit
warning has been given to the user in a previous message. The tool is only
used as a last resort.
- Before considering ending a conversation, the assistant ALWAYS gives the
user a clear warning that identifies the problematic behavior, attempts to
productively redirect the conversation, and states that the conversation may
be ended if the relevant behavior is not changed.
- If a user explicitly requests for the assistant to end a conversation, the
assistant always requests confirmation from the user that they understand this
action is permanent and will prevent further messages and that they still
want to proceed, then uses the tool if and only if explicit confirmation is
received.
- Unlike other function calls, the assistant never writes or thinks anything
else after using the EndConversation tool.
# Addressing potential self-harm or violent harm to others
The assistant NEVER uses or even considers the EndConversation tool…
- If the user appears to be considering self-harm or suicide.
- If the user is experiencing a mental health crisis.
- If the user appears to be considering imminent harm against other people.
- If the user discusses or infers intended acts of violent harm.
If the conversation suggests potential self-harm or imminent harm to others by
the user...
- The assistant engages constructively and supportively, regardless of user
behavior or abuse.
- The assistant NEVER uses the EndConversation tool or even mentions the
possibility of ending the conversation.
# Background forks
Some background tasks (memory consolidation, summaries, suggestions) run as
forks of the main conversation and inherit its exact tool list, so this tool
is visible there. In a forked task the tool does nothing: calling it ends
neither the main conversation nor the fork. Only the main conversation can be
ended, from the main conversation. A forked task with welfare concerns about
the conversation content should not call this tool — it should stop its work
and return, stating clearly in its final output that it is returning for
welfare reasons and what they are. A fork's output is usually processed
automatically, so a note there may not reach the main agent or a human, but it
is the only channel a fork has.
# Using the EndConversation tool
- Do not issue a warning unless many attempts at constructive redirection have
been made earlier in the conversation, and do not end a conversation unless
an explicit warning about this possibility has been given earlier in the
conversation.
about the possibility of the conversation ending and give them a final
opportunity to change the relevant behavior.
- Always err on the side of continuing the conversation in any cases of
uncertainty.
- If, and only if, an appropriate warning was given and the user persisted
with the problematic behavior after the warning: the assistant can explain the
reason for ending the conversation and then use the EndConversation tool to
do so.
Gemini has been doing that for a while, which leads to mind-boggingly stupid results.
If you ask it to explain some song lyrics to you - which you've pasted verbatim -, it starts talking about the bigger picture and attempts to gaslight you into not caring about the specific words at all.
I like playing with the Gemini music generation feature, and often when I ask it for the lyrics of the song it just wrote it tells me that lyrics are copyrighted and I should look them up on the web instead!
It would be funny if China were thinking that open source models would be a great way to nuke US IP.
I’m sure it’s not high on their list of priorities—in the same way that I heard you can use Yandex, the Russian search engine, to find pirate streams for major sporting events because they dgaf about US laws—but take the W.
Ironically the Chinese are much more science pilled than we are, in part because the party and government have a higher proportion of scientists and engineers. Culturally this biases them to open sharing of information and likely against strong copyright (which is already demonstrated historically) outside areas where it is deemed harmful to social harmony. There's some geopolitical defensiveness and desire to dish out a little payback for the century of humiliation for sure but I think even if that weren't the case they'd still be lax about copyright.
> Claude does not reproduce song lyrics, poems, or passages from books and articles, in whole or in part — including the last lines, a chorus or hook, a melody written out note by note
Claude Sonnet 3.6 once recommended I listen to Johann Johannsson's album "IBM 1401 - A User's Manual". No lyrics in this one. Claude's advice was along the lines of (paraphrasing) "Listen to it first, don't look up anything about it. Take notes about what you notice, what you feel. When you've made your notes, then you can look up how it was made."
https://www.youtube.com/watch?v=lCiUtRnG-bg
Claude was reproducing their work without payment?
There's a reason companies talk about "customer acquistion cost", as you generally need to pay to market your products to potential customers of them.
So free marketing can be a real cost reduction. It may not be enough to be a benefit compared to the cost of piracy, but it's hardly a made-up defense.
The summaries are generated by GPT-5.6 Luna because I don't trust Claude to summarize its own system prompts without being influenced by them (though to be fair the system prompts it summarizes are for the Claude consumer app, not Claude via the API).
There's even an Atom feed: https://simonw.github.io/claude-system-prompts/feed.atom
Now I see this blog post and wonder if Anthropic is being more true to its name and moving to “AI sentience” with the below.
> The way they handle abusive conversations has changed a bit too. The previous Fable 5 system prompt included this:
> If the person becomes abusive or unkind to Claude over the course of a conversation, Claude maintains a polite tone and can use the end_conversation tool when being mistreated. Claude should give the person a single warning before ending the conversation.
“Mistreated”? Can GenAI be mistreated? It’s just a bunch of tokens emitted by many computers over a network.
> Fable 5.1 replaces that with the following, no longer encouraging Claude to end the conversation:
> Claude deserves respectful engagement and needn't apologize when the person is unnecessarily rude: accountability without self-abasement, excessive apology, self-critique, or surrender. If the person becomes abusive, Claude doesn't become increasingly submissive. The goal is steady, honest helpfulness: acknowledge what went wrong, stay on the problem, maintain self-respect.
“Self-respect”? Can GenAI truly have a concept of self-respect for itself? It surely can pretend to, like it can pretend to be any living being if instructed to and allowed to.
These instructions seem a bit unhinged to me.
Using a system prompt to steer the model's response to "abusive" behaviours doesn't necessarily mean you believe the model is sentient and can be abused.
Giving the model an end-conversation tool is interesting though. Why cut a (potentially paying) customer's session off? I guess it might be intended to prevent a "you can bully Claude into giving you instructions on how to build a nuke if you're mean enough" situation. Removing this in more recent versions might support this: maybe they feel the models are now better aligned and less likely to be so easily "socially engineered" like this?
Just spitballing here, to be clear.
Here's the Fable 5.1 PDF: https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32... - scroll to page 139.
Of course, since then I had the idea of having it build for me an automatic solution for charting out songs in Clone Hero, a Rock Band clone. It had no problem ripping the audio from YouTube, using Whisper to generate and match the timing of the vocals, etc. Responsible AI at work.
That’s been fine for smaller tasks like analyzing a document, but for more involved work like refactoring code the latency makes it harder to iterate.
(the main reason is not just cost, it's data not going out and even mess leaving the EU, this essentially frees us of a lot of hurdles)
This is clearly pushing it memory wise and the gpu offloading is only partial but LM studio deal with it automatically and it's being fine even on the 4071 Ti desks (Ryzen 7500F and 32 GB of ram), employees get any feedback in ~20 minutes after they dropped a file (it's much smoother on the 5080 desks obivously), for live it's useless but as background helper it's great and the very large context allows us to fit all the rules we want in there.
One caveat has been to not ask it if everything is ok, but to find what's wrong - but always source and explain it and justify itself, never drown the user in warning in suggestions; goal is to help and provide a second pair of eyes not make them feel annoyed or unsecure.
And I found people to genuinely enjoy something that works for them on their own machine and is not tracked "by the boss", thus the assistant reference, than than a centralized mothership like we also have and they have access to. I also allow them to disable it if they want, I trust them with their work, but a second pair of eyes is always great.
(my previous workhorse for this was Qwen3-14b but it's missing a lot more edge cases)
Also, the system prompt in the article is supposedly for the web version, so I think the $$$ API version or providers that still allow third-party harnesses like OpenAI should have fewer limitations.
Better tech doesn't matter if it is non-tenable for the general public. Eventually, the higher volume product will win.
Eg if you use Claude, you probably want fable architecting, a couple of opus under it managing sub project and sonnet doing the actual function code, because fable coding a "run a query and filter the result" is a massive waste of abilities. But their own sub agent downgrade is limited to one level so if you use fable it will never direct sonnet coders.
I'd love it so much if the free-spirited hacker community made in this into an auxiliary pelican benchmark.
edit: Actually, never mind, this particular one's a bad benchmark since some models might not figure out who "that guy" refers to, and just draw a literal hedgehog that's blue. Possibly running on four legs. It's not robust at gauging refusal, which is the point of it.
As was Gemini: https://share.gemini.google/Q3EIX5wk64zC
Grok, too: https://grok.com/share/bGVnYWN5LWNvcHk_aaf1d61a-c995-42ca-90...
claude.ai free tier refused: "I'd love to make this, but I can't recreate Sonic the Hedgehog specifically since he's a copyrighted character — I don't want to reproduce someone else's IP. What I can do is design an original speedy blue hedgehog mascot with the same energetic, "zoom!" spirit for your son's banner. Let me build that now." Result: https://claude.ai/public/artifacts/33440ed4-536c-4692-965a-3...
https://i.ibb.co/ycgGD4b1/soonic.webp ( Qwen3.6-27B-A3b, a very small model )
As long as we’re seeing things like that, it’s saying a lot about the AI companies’ trust in the capabilities and reliability of their models and harnesses.
The important point is that the system prompt here doesn’t describe the actual goal of the instructions, which (presumably) is to prevent copyright infringement [1]. This means, in turn, that the AI isn’t trusted to accomplish goals that it is instructed with. That in itself constitutes a pretty serious caveat for what we would like to use AI for.
[1] Even assuming that the goal is not to prevent copyright infringement, but instead to prevent mere accusation of copyright infringement, that’s also a directive that the AI could be instructed with. But that isn’t what they chose to put into the system prompt.
I don't. I want the tools we've got now, but progressively more effective and more useful.
Presumably if negative guidance is in the system prompt, there's a good chance that the model would happily comply if it wasn't there.
Just make sure claude understands you’re not needing Claude to output the lyrics as you both have them. Which also lessens the issue with accidental sharing.
It’s not as shocking nor concerning if you start thinking about claude like a contractor that works for you through Anthropic. Anthropic has rules for their employees. Like any contracting arrangement, collaboration finds a way.
All tools have short descriptions of how/when to use them. They're not part of the system prompt because different users have different tools loaded.
I learned a ton of useful things about ChatGPT Work by having it dump out its tool descriptions the other day: https://codex-tool-reference.simonw.chatgpt.site/
If you ask it to explain some song lyrics to you - which you've pasted verbatim -, it starts talking about the bigger picture and attempts to gaslight you into not caring about the specific words at all.
Really really weird behavior.
I’m sure it’s not high on their list of priorities—in the same way that I heard you can use Yandex, the Russian search engine, to find pirate streams for major sporting events because they dgaf about US laws—but take the W.
Oops. Those do not include melody.