Many Stacks have been discussing ML-generated content on their Stacks in their Meta sites. After reviewing those discussions, I became very curious and wanted to look at the potential impact of ML-generated answers on Music.SE.

As an initial experiment, I have edited one of my lowest scoring answers to add an ML-generated answer from ChatGPT. I present this here as a way to start a discussion about whether we want to control such content at Music.SE.

Are Cello straps and stops necessary to rent for beginner?

Some relevant links:

Is there a list of ChatGPT discussions and policies for our sites?


Ban ChatGPT network-wide

Is attribution required for machine-generated text when posting on Stack Exchange?


  • 1
    Well shucks, that answer deserved an UV already! Dec 20, 2022 at 21:11
  • 1
    Utterly fascinating. I'm going to need to think on this!
    – Richard
    Dec 20, 2022 at 21:38
  • 1
    I'd love to write an answer on this, but I'd need a lot more time to digest the topic as a whole. I've done some work on composing with AI generated motifs in the past and it gives a good starting point for something to dive into fully. So I could see it being used as a starting point to an answer, but some of SO's concerns are valid as not knowing if the answer is correct and just posting it will lead to a lot of low quality content plus the attribution of where ChatGPT is getting the info from makes it hard to properly attribute content on the site.
    – Dom Mod
    Dec 20, 2022 at 22:43

3 Answers 3


(With the disclaimer that I have my head in a hole in the sand and have not experimented with ChatGPT yet): I feel like our needs are different than SO's, and we don't need an explicit ban on AI answers as such, but will have practical reasons that should stand in the way of them.

What I get from a quick glance at SO is:

  • They've been apparently inundated by a high volume of ML-generated activity. If we saw that here perhaps it would be a problem, but we haven't so far.
  • One of the main complaints is that it's "easy to make bad answers, but hard to tell they're bad." If you just paste in a block of code, you need what's essentially a code review to tell whether the answer is useful. A music answer might be easier to tell whether it's kinda good, truly useful, or off base.

It's interesting to look at your experiment. The AI answer is full of good information, and I might upvote it for "this is no replacement for good posture," but it's ultimate answer to "do I need them" is basically "Well, some do ok with them and some without, ask your teacher." Which is kind of a non-answer. The wishy-washiness is cloaked by extra info that wasn't actually asked for.

As it is, this is not a useless answer, and is certainly better than some human-generated ones. But it wouldn't win a ton of votes. I think it would be clear enough that someone a user generated content outside their actual knowledge if they were asked to clarify or challenged on details and couldn't respond. To my mind, if it's a lousy answer, it will get ignored; if it's an ok answer, then the more the merrier, and if it's actually a good answer, then why not!

Now, if we started seeing a flood of mediocre content, such that it overwhelmed more useful content, then we might have to do something. Or maybe some policy about individual intervention: Say a user just goes through dozens of questions in a day (including old ones, I guess, since we don't have that many new ones), dumping AI-generated answers. They're challenged by others to explain, clarify, or correct inadequate answers, but they never respond. None of the answers are so bad as to merit deletion, but they're just not up to snuff, and this user is not responding to reasonable requests that they slow down and give it more thought. In that case maybe some kind of moderator action against the user is warranted, but it's more about the volume than the quality.

As my own experiment, I just tried feeding ChatGPT this question, thinking it's surely a great example of what it wouldn't get right. Its answer was better than I expected:

"Sostenuto" is a term that means sustained or sustained and held. In this context, the composer is instructing the performer to hold or sustain the notes in a manner similar to the way the fourth string of a violin is played.

The fourth string of a violin is typically tuned to a D, and when played it produces a rich, full, and sustained sound. In order to suggest this sound on a keyboard instrument, the performer should aim to produce a sustained and even sound, with a focus on achieving a full and rich tone. This can be done through a combination of touch, pedaling, and dynamic shaping.

One way to achieve a sustained and full sound on a keyboard is to use the sustain pedal, which can be pressed down to sustain the notes being played. This can help to create a sense of continuity and flow in the music. Additionally, using a gentle touch and shaping the dynamics of the notes can also contribute to a sustained and full sound.

It's worth noting that different keyboard instruments, such as the harpsichord or clavichord, may require different approaches to achieve a sustained sound. In the case of a harpsichord, for example, the performer may need to use a technique called "loure" to achieve a sustained effect.

So I guess we see some of SO's concerns here: it's so rational that it gives a convincing impression of being written by a person, and even delivers a pretty reasonable "bad answer": It understands that we're telling a piano to imitate the tone (I even like its word choice of "... in order to suggest this sound"!) of a certain violin string. It handled the Italian translation for us. It's right that "tone" for a piano is a matter of "a combination of touch, pedaling, and dynamic shaping." But it has "the wrong end of the stick" in general, talking only about tone rather than timing, and of course missing the point that most of the humans focused on, the fact that this marking came from an editor rather than the composer. And it has a few glaring things just plain wrong: The "fourth string" is G, not D, and as far as I can tell, louré is only a bowed-strings technique, and not a harpsichord one. (Either ChatGPT is showing its ignorance or I am!)

Other users would have jumped down the throat of this answer with these mistakes (or maybe some kind soul would have edited the "D" to "G" assuming a typo). There would be little reward to the spammer for posting it.

EDIT: I clicked "Regenerate response" and ChatGPT tried again; this time there are no glaring errors and the answer is even more reasonable:

The word "sostenuto" in this context is an Italian musical term that means sustained or held out. It is often used to indicate that a musical line or phrase should be played in a sustained or legato style, with the notes connected smoothly and without interruption.

The instruction "come IVa corda di violino" means "like the fourth string of a violin." The fourth string of a violin is the lowest-pitched string on the instrument and has a rich, full sound. It is often used to play the lowest, most sustained notes in a piece of music.

To suggest this property on a keyboard, the player should aim to produce a sustained, legato sound by using the sustain pedal to hold the notes for a longer duration and by using a smooth, connected touch to play the notes. This can help create a sense of continuity and flow in the music, evoking the rich, full sound of the violin's fourth string.

  • 5
    That's the trouble, though. It manages to produce such a well-rounded, 'nearly' answer that it would be tough for the OP to differentiate well-written tosh from accurate information. I've tried asking it questions which I know to have impossible solutions, yet it blithely tells me how to do something I know from experience will not work.
    – Tetsujin
    Dec 30, 2022 at 19:26
  • 3
    @Tetsujin It is possible for humans to answer with well-written tosh also, but that’s not a new problem for SE. Jan 6 at 14:13
  • @ToddWilcox One might even note that such answers, being curated through ML, are just human- (and well-)written stuff mashed up. If it's clever enough to use "suggest the sound" it's only because some uncredited human was clever enough to do so and it happened to find the right snippet... Feb 9 at 15:55
  • 1
    ChatGPT: "The fourth string of a violin is typically tuned to a D". You know, Andy, that that is complete nonsense. If ChatGPT is as clearly factually wrong as this then there is little room for it here. Apr 6 at 13:14
  • The problem with AI answers is precisely that you cannot predict when they will be complete hallucinations. Anything which requires fact checks will be susceptible to facts the AI drew out of ... let's not go there. A recent example is in Donald Knuth's question about a novel; the AI gave a detailed but completely fabricated account of the chapters of Leon Uris' novel (or maybe the facts are actually true about some other book; haven't checked). See e.g. thenewstack.io/…
    – tripleee
    Jun 7 at 12:14
  • @tripleee Good point; my "quality is self-evident" point above breaks down when the topic is so arcane that fact-checking requires more than common top-of-our-head knowledge (like "Did Jelly D'Aranyi introduce Ravel to Bartok?"). Jun 7 at 12:30

I haven't seen a lot of ChatGPT answers to have a thoroughly informed opinion, but I am certainly wary.

While there is a potential for some good to come from ChatGPT, there is definitely some bad that come from allowing/encouraging it.

Now, if you'll allow me to reframe the question:

Do we (Music: P&T) have a need for ChatGPT?

  • Is it likely to solve any actual problems that we have?
  • Are there a significant number of questions here that don't get good answers, that ChatGPT is likely to provide good answers to?

I don't think so. In my experience almost all "good" (and not-so-good) questions get (often several) good answers here.

Questions that ChatGPT is likely to be able to answer well, with and actually useful answer (questions that don't require any real experience or nuanced interpretation) are generally not likely to be great questions for this place anyway.

At the very least, I think all ChatGPT answers should be very clearly marked as artificially generated.

  • 1
    The trouble with being 'clearly marked as artificially generated' is that the people who are posting these as answers are doing it in the hope of a quick rep gain; all tend to be newbies, oldbies tend to know better. They don't give a monkeys that it's screwing up the site. I've seen ones recently that are re-typed rather than copy-pasted, so you can only tell from the overall style [& sometimes hard-to-spot factual errors] not the word-perfect grammar that they are in fact ChatGPT.
    – Tetsujin
    Jan 21 at 16:09
  • @Tetsujin I mean to say that I think answers generated with ChatGPT that aren't clearly marked should either be edited to be so (if they are good), or just removed. (Other than the above) I don't think I've actually seen any ChatGPT answers on Music: P&T. Jan 21 at 22:03
  • The writing style is pretty easy to spot once you've seen a few. The issue is that unless you know the exact answer & all potential pitfalls it's very hard to tell whether it's absolutely correct or just so much horse-apples. See this - apple.stackexchange.com/a/453679/85275 - it has the writing style, but also it has typos which ChatGPT never does [hence my comment underneath]. As an answer to the question, it's basically garbage.
    – Tetsujin
    Jan 22 at 9:20
  • Here is an answer from ChatGPT, generated from the question title, copy/pasted verbatim… "It is possible to determine the amount of memory being used for graphics on Apple Silicon devices using the Activity Monitor application. In the Activity Monitor, you can view the memory usage of the different processes running on the device, including the GPU process. Additionally, you can also use the command-line tool "vm_stat" to see the memory usage of the system, including the memory used by the GPU." It is equally garbage.
    – Tetsujin
    Jan 22 at 9:27

Disclaimer: I'm not attached to any particular answer, but I would like to improve my understanding of the issue.

Topic for discussion: Banning ChatGPT would be like banning copying Google search results; that is, if a tool is run on the site to detect and eliminate ChatGPT or Bing Chat results, then a tool should also be run to detect and eliminate any plagiarism of text already on the Internet.

Reasoning: Both Large Language Models and AI searches such as Google or Bing are generating answers by user prompts on a statistical database, but the user is not generating original content, only providing prompts and using that content. The US copyright office has recently released guidance on copyright of AI-generated content: https://copyright.gov/ai/ai_policy_guidance.pdf The implications of its answer are interesting; particularly how it interprets the human-provided content and human manipulation of machine-generated content.

It could be said that there is only one criterion for an answer: it is either useful or it is not. Almost none of our answers here are original content; one could say we are curators, not creators, of knowledge that we present in an easily searchable form. Curation has its elements of judgment and refinement and creative content generation.

I look forward to reading your answers to this issue in the comments below.

  • 3
    You need to research Chat GPT a bit more. This answer is very naïve, sorry. ChatGPT has a 'flip of a coin' chance of being right or wrong. The problem is people use it to try gain 'quick interwebz points' rather than provide good, rational answers, based on their own knowledge, with appropriate citation where neccessary. This means someone else has to check the validity of every answer - as the poster didn't. Signal to noise ratio becomes high, additional useful information goes up by a negligible amount.
    – Tetsujin
    Apr 7 at 17:45
  • 2
    Try it yourself. Ask it a question to which you are either certain there is no answer or to which there is a clear but complex answer.
    – Tetsujin
    Apr 7 at 17:45
  • @Tetsujin could not the same be said of a naive Google or Wikipedia search and the user copy-pasting the answer they find? As for "naivety," I have a career in machine learning. ;-)
    – empty
    Apr 7 at 17:51
  • 3
    It could - but presumably by floating to the top of a search engine, it would have some traffic backing it up, or the user may have gone so far as to check two or more search results, not just one. I have a career in the music industry, but don't ask me to write you a symphony.
    – Tetsujin
    Apr 7 at 17:54
  • I can certainly see the reasoning behind SO's banning of ChatGPT. For the moment, they say. SO has an enormous active user base with thousands of questions per day. Very often the code is more or less original. However, many software developers and sysadmins make a career out of copy-pasting SO answers into corporate software, when they aren't using open-source software from GitHub.
    – empty
    Apr 7 at 17:56
  • No personal attacks going on here. Have a nice day.
    – Tetsujin
    Apr 7 at 18:00
  • 1
    I'm going to point out if you should not just copy answers you find from google. Copying without attribute is plagiarism and is not allowed here regardless of if it is useful or not. If you find plagiarized content, please flag it. This is also one of the side effects of Chat GPT as getting an answer generated from it does not have sources and as pointed out by others on this topic previously.
    – Dom Mod
    Apr 8 at 22:18

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .