14 Comments

PS. Although I included enough info to find the YouTube channel of the "band", please don't harass the creator. While I think they should be more forthcoming in labeling the content as synthetic, it's ultimately up to platforms to give us better control.

I'll also address this HN quip about my HN quip:

>>> Despite some HN quips, AI text detectors are pretty dependable

>> I have never tried AI text detectors, but my impression was that they were considered unreliable?

> They are. The author is wrong.

There is some HN lore rooted in public complaints from people who were sacked from their jobs because someone investigated their writing and it consistently failed such automated checks. The complaints fit neatly into several recurring narratives, so I don't think the reporting gets seriously challenged.

I was quite skeptical of the tooling, but having played with it extensively, I believe the apps tend to be unreliable mostly in the sense that there are prompting tricks to evade detection. I was not able to consistently trigger false positives on human writing - regardless of the writer's language proficiency or style. Further, for the handful of stories where the fired persons' writing was public, it was pretty clear to me that they weren't telling the whole story. In essence: the tech, while not perfect, is pretty good.

[Edit: in a comment below, Ruslan gives an example of a human email that triggered the tool for him, so caveat emptor.]

Expand full comment

On the AI text detectors: to be fair, my experience differs. I've tried https://quillbot.com on a few emails I drafted recently without using ChatGPT (though I still did my best attempt to make them formal), and the detector claimed that the text is 100% AI generated. Even when I tried to type up a less thoughtful email in a fairly casual style, it still reported the same: 100% AI generated.

This could mean only 2 things: either AI detectors are indeed not reliable yet (at least this particular one)... Or I'm an AI! 🤖😬

For real though, it seemed to me like it's biased when it stumbles upon cliche phrases like "we'll do our best to provide a solution most valuable to your company". And if it finds a few of those in the text, it claims it to be AI generated. When in reality it's a regular human brain distorted by years of corporate written communication :D

Expand full comment

However, it did a good job on this comment above concluding that it's 100% human written. Which means I'm NOT an AI :( I guess I can only pretend to be an AI when I write work emails :D

Expand full comment

Interesting - if you ever have a sample of text you're comfortable sharing, I'd love to have a look. Not that I doubt you, I'm just curious about what makes it fail. I'm assuming you had a snippet of a reasonable length? They typically need at least 3 paragraphs or so.

Expand full comment

It's true that I typed the bare minimum required for them to accept the sample. But if that's insufficient for reliable decision, the should probably increase the required length then.

Here's the text I just typed out of the top of my head (and luckily saved a screenshot :D), which was similar to previously sent email, but less formal and thought out, containing tautology, etc:

```

Hello, Joe, nice to meet you!

I'm Ruslan, and I focus on infromation security area here at [redacted].

Thanks for providing additional info, we'll look into it and do our best to come up with the solution most valuable for your org.

Could you please clarify what type of training do you expect for your [redacted]? Knowing your expectations would allow us to tailor the training to your particular needs and expectations.

Thanks for reaching out and we look forward to chatting on monday!

```

Expand full comment

Interesting! I think this is the first real example I'm seeing that reliably trips them. I suspect it's because it's short and ultra-generic - i.e., it's probably pretty close to what Gmail auto-completion would suggest - but yes, you're right that the onus is on them to set up thresholds.

Expand full comment
Jun 12Edited

Hey there, I'm the creator behind your mentioned "band" - I usually call it music project because I don't want to imply anything that it is not! First off, thanks for taking the time to write about this musical experiment. I got a good smile out of your "comically kitschy" description of the lyrics. What can I say? I like it ;-)

Now, about those changing vocals - you're absolutely right. Consistency is a real challenge with the current state of AI audio tools. While we've made leaps in the visual realm (shoutout to Midjourney and of course, Stability AI, who more or less were the first to push image generation to the heights it is today - at least, imo), audio is still playing catch-up. It's a bit like we're in the era of six-fingered images, but for sound.

Fun aside, I want to address a couple of points you raised, because transparency is important to me. As you mentioned, I do disclose the use of AI in my Spotify profile, and I make sure to check the "AI generated content" box on every single YouTube upload. I'm not trying to hide anything from anyone. In fact, I reply to every comment on Instagram, YouTube, and other platforms, confirming that yes, AI was used to generate the foundations of these songs. But I also re-iterate that it does not stop there. I'm not trying to build an AI agent that just spits our music without any human interaction and feeling. 

Again, here's the thing: for me, this project isn't about scaling or mass-producing low effort content. It's a personal form of creative expression. I'll be the first to admit that I can't play a guitar or sing to save my life. But I do know what I like, and I've learned how to use AI tools to bring that vision/feeling to life. My thesis is that people will be (or are already) fed up with low effort trash AI stuff. But when used creatively, it's an insanely powerful tool. 

After the AI generates the basis for a song, I spend hours (for some songs days) in RX11 and other software, reducing noise and refining the audio (the output quality is often times not even at the "six fingers" stage of image generation AI, if you know what I mean).  That's also why I don't put "AI Generated" for example in the title of the Videos or anything. Because in my mind, it would imply that I just hit generate, add the video, and upload it to Youtube - which is simply not true. My goal is to "infuse" (for lack of a better word) each track with genuine emotion, if that makes any sense. It is simply a new way of working with these things. I do truly believe that we're on the cusp of some mind-blowing advancements in AI-assisted music creation (Ai in general but here we are talking music), and I'm excited to be a part of that in some ways. 

So while I appreciate the critical lens through which you viewed my work, I want to assure you and your blog readers that my intentions are good. This is a fun, challenging, and deeply rewarding project for me - a way to push my personal boundaries and a test of what's possible when human creativity and AI join forces. Okay, this sounded more pompous than planned, but I guess you get the point. Also, it does not mean that everybody likes it, nor that the song structure might not be "weird" at times due to a glitching AI. Sometimes I can repair it afterwards by glueing the song differently together in FL Studio - sometimes I keep it. Music is something that is very subjective, and while it might sound cheesy to some, others might like it and call it catchy. Just as a thought.

In any case, thanks again for sparking this conversation, and for encouraging your readers to engage with my content in a respectful manner - because there is already a lot of hate about it out there (which I can partially understand, given the state of the music industry, but this is another bigger discussion that goes beyond the scope of this comment). 

Cheers,

Jay

Expand full comment

Thank you for the thoughtful reply! I wasn't expecting this, and I really appreciate you taking the time.

I have no way to judge your process or intentions, which is why I didn't make it a focus of the post. I can judge the output, which feels awfully like Suno tracks with ChatGPT lyrics. But as noted in the pinned comment, my main concern isn't with the creator - it's with platforms not giving us choice. I wasn't upset that you or others are experimenting with generative music. I didn't like that it showed up on my Spotify playlist unannounced, and looked just like a human band.

I will say one thing, though. You went out of your way to make this look like a real band - complete with what seems like 100% AI-generated cover text for your Spotify description and YouTube channel. You didn't style it as a one-person shop experimenting with AI tech. The AI lede was buried and ambiguous (Spotify) or non-existent (YouTube description & videos).

I can't help but suspect that part of you wanted to see how it goes - how many people can't tell the difference. Maybe I'm being unfair, but that part didn't give me good vibes.

But again - I think you're well within your rights to have fun with this tech!

Expand full comment

[The author had a followup which ended up in a separate top-level response with confusing ordering, so I'm manually moving it here.]

Jay: "short addition to the longer comment: After checking your mastodon and thinking about it, I will rewrite the Spotify Bio to make the "AI based music project" and "not real band" aspect more clear.

Also, I would like to add that I do mean it when I say I have an idea/feeling in mind and then use the AI tooling to get to that (because yes, I could not play the instruments myself alone). Hence, I have huge tables with thousands of rows like this https://ibb.co/FXg3MS0 to incrementally refine the input so that it matches the thought in my head (which also takes a lot of time and is not about make scalable music fast and easy)"

Expand full comment
Jun 12Edited

You're right, part of the branding is indeed deliberate. The aim was/is to set this project apart from the "I generated my top song in 5 minutes" crowd on YouTube. But I can see how, from an outsider's perspective, the AI aspect might not have been as clear as it seemed to me on a subjective level.

For example, the header image on Spotify of "the band" that you mentioned at the beginning of the article - in my mind, it was so strikingly AI-generated by design, because I wanted to express right from the start that "this is not a real classical band". I mean, with tools like Midjourney, you can easily create images that are much harder to distinguish as AI or not. But that wasn't my intention at all.

In any case, I really appreciate your response here. Btw., based on this feedback here, I've made some changes:

- added a lot more "AI" remarks in my social bios to make it clearer.

- rewritten the Spotify bio to emphasize the AI aspect much more prominently.

- added an AI remark to the YouTube channel description.

I hope this makes things clearer moving forward.

Now it's up to the platforms to allow people a switch like "I don't want to see AI generated content in my suggestions" or something like that. The pathway is there, and the data is being collected (for example, with the checkmark about AI generated content when you upload stuff to Youtube).

Thanks again, and have a great day!

Expand full comment

It’s over.

Expand full comment

Yes AI has its limits, but the real limit is still the user as AI systems that require prompting are just mind amplifiers.

Surely this (link below) isn't that bad?

https://www.udio.com/songs/1ijngT2A1rP9u7ahQtdDeo

Expand full comment

I think that's a simplistic description. I can prompt image or music generators in very laconic and frankly thoughtless ways, and get results that I wouldn't be able to replicate myself - but that draw from the works of more talented people elsewhere.

Don't get me wrong, it's cool - but ultimately, if the bulk of the creative heavy lifting is done by the machine, I don't think it's fair to take credit for the result. When I order food in a restaurant, I don't pretend that I cooked it myself. I made choices, but that was the easy part.

Of course, there's plenty of cases where the creative labor is more evenly divided, or where the AI is just used to automate the grind (e.g., dust removal or color correction of a photograph). I don't think this is nearly as controversial.

Expand full comment

When you use AI to be creative you are acting as an Art Director, Producer, etc. not so much an artist, and certainly not yet as a performer. Live music will be more in demand for that reason now. One way to test this is to see what different people can produce consistently compared to others with exactly the same tools and team of AI minions. I posit that it simply shifts the playing field down to the level of anyone with the ability to access the AI, those people will then be the limit on the "value" of what is produced. The machine is not being creative, the user is, the machine empowers the user to achieve at a level of somebody with the privilege to have the requisite education, experience, and equipment etc. Those talented people the AI is "stealing from" will still produce more sophisticated work than you can with the same AI tools. It is a baseline shift, not a paradigm shift, in fact I think the really talented people will benefit more than everyone else as the "mind amplification" effect is nonlinear, the more talent you have the better the results with the top people being way ahead of everyone else once they master the tools and adapt their modus operandi to suit.

e.g. One person gives you cheesy metal, and I give you a song in Latin about Plato's allegory of the cave. Both fit roughly into the same style, which has greater cultural value?

Expand full comment