🤖 BubbleGPT - An AI powered Bubble at what cost?


As someone developing my own apps, these tools would be phenomenal for rapidly putting a new MVP together or, of course, improving an existing product. But as someone who develops apps for others, this tech threatens the income that affords me time to try building new businesses on Bubble. So the calculus becomes:

  • I can’t bill as much for my work because… everyone knows it’s easy.
  • Even though I can launch/improve MVPs faster than ever, 7 out of 10 business ideas are still going to fail to generate any meaningful income.

There will likely still be developers but, perhaps we’ll have to juggle an ever increasing number of clients to maintain earnings. But then, of course, now the rate of competition for picking up clients is higher for everyone. Plus, it takes very little skill in making apps so the market gets flooded with them and what was a 7/10 rate of failure is now 8/10 or 9/10. Rip offs left and right… I don’t know… makes the mind spiral a bit.


Before asking an AI to build an app from a prompt, I guess one of the first move for Bubble would be to feed an AI with the tons of valuable contents shared in the Bubble forum for years, so you could ask this “copilot” for help whenever you’re stuck in the dev process.
Maybe an AI could also help the Bubble team to rewrite the knowledge base :wink:
But please, don’t just add AI for marketing or hype purpose…


Turns out Bubble is looking to pay two engineers $10k each per month to, per the job post: “train LLMs to output Bubble’s no-code language”

So the prediction was correct and they necessarily will need lots of training data which is, crucially, a requirement our app data satisfies.

The reality is, whether or not the Terms of Service we’ve already agreed to “by the word” allow Bubble to use the business logic we’ve put into our apps over years of building on the platform, such terms are not something we’ve agreed to in Spirit.

@josh defends my point perfectly in today’s blog post as he writes:

Conversational interfaces for building no-code apps is the next big transformation of the space. A few years ago, a conversational app builder might have sounded outrageous. Today, it’s what we’re hiring people to build.

That’s right, the notion that I could speak a website into existence was so outrageous even 4 months ago that it would be false for Bubble to argue now or in the future that, by agreeing to their TOS, we’ve already given them the right to use our apps to train an AI. To be sure, I would have never agreed to something like that. Nobody saw this coming and I imagine when users of this platform really start to think about the implications of their work being used to make an AI that can create sites that compete with every site we’ve ever made (in a literal sense, having our own work compete against us) - Bubbler’s will decline that in droves.

Look, team Bubble - if you want to make BubbleGPT, that’s fine but do not use our Bubble apps as training material without our explicit consent. Generative AI may have “crossed a threshold” but you gotta knock first and ask for permission before you cross ours. @emmanuel


Maybe an opt out would be fine. But first let them build something…

1 Like

if the entire context cant be understood from only docs and api, gpt 4 fails. But if you prompt it good enough (by giving it lots of info), results become better

If they build something using our app data then the opportunity to opt out has been missed.


I totally understand everyone’s worry about Bubble’s AI using everyone’s Bubble apps as part of it’s training data. Though i am starting to find that it’s bordering on hyperbole.

Unlike art, building apps isn’t a purely creative process. It’s iterative and built upon one or many fixed languages. The end product is meant to serve a purpose and a demographic. Is anyone here bold enough to say their app isn’t built with parts of someone else’s code? Their hardwork?

Art on the other hand can be purely creative. Artists take inspiration from other artists but at the end of the day good art is unique to the artist who created it.

If every programmer treated each piece of code like it was sacred than GitHub wouldn’t exist. Dial down the doom and gloom people.

We stand at the brink of a pivotal moment in the world of apps, where we may either see an unprecedented era of innovation and originality, or a deluge of low-quality imitations that threaten to stifle creativity.


They can always retrain a system. Especially with dataset like bubbles that should not take to long.

I think you’re conflating things here. Github is what it is because people voluntarily provided their code to a public repository to be freely used . Linux is the world’s most used OS because it was given away.

Of course we build digital products using free to use software - we have permission to. Private Bubble apps on the other hand, just like private code , are private. Linux would not be the world’s most used OS if it had a restricted use license.

I mean otherwise, I expect you’ve provided read access to all of your private - even commercial - apps to the Bubble community? Can you link me to those?

By the way, building Bubble apps is a hugely creative process. It is art.

1 Like

Now i have to ask, what are you actually worried about? Your code or your private data getting leaked? If it’s the latter then i 100% agree.

Unless you are telling me you have some unique proprietary piece of code that only you and your team can create, i very much doubt even your private code isn’t already available elsewhere.

I build apps as a side hustle. I’ve been in the creative industry (i did post production work mostly and now i have to churn out posters on a daily basis) for almost 10 years.

I can assure you that building an app is not art, objectively speaking. We can argue about artistry in expertise but let’s not kid ourselves by calling it art. Any UI/UX designer who still does art, or used to, can tell you that.


Okay, so this forum has over 26k posts under the Need help category. Any particular workflow in a private app of mine is not special but could be the answer to many of those posts. My workflow doesn’t have to be particularly unique, it just has to be useful to someone to be valuable. There was a day when I didn’t know how to implement that workflow but I spent time, money and failed a lot to eventually figure out how to do x, y, or z very well. Whether that’s the thing that gives my app an edge or not, the totality of those workflows, and the creative ordering and interplay of them with visual elements, is what makes my app mine. It is a creative output that is a unique reflection of me and how I solve problems.

I don’t want Bubble taking my logic and approach to solving <insert_something_you_can_build_on_bubble> to train an AI that can give anyone the ability to create any site by writing a paragraph. The beautiful thing is we don’t have to be on the same page about this - if Bubble asks you for your permission to use your apps to train their AI, then click Agree. What I’m saying is we should have the choice.

1 Like

Choice is good, i do agree. 2 things though;

When talking about Art in the context of generative AI we’re talking about the originality and creativity both in the process and result of creating art.

Code is code. You don’t get inspired by other people’s code, you iterate. You can make a piece of code better but you cannot write code any other way than it’s meant to be written. I can’t write:

Tis 1 + 1 doth true, sumeth the numbers

I’m not belittling anyone’s efforts in learning how to build apps cause it’s hypocrisy on my part. We are here today because of the efforts of those who “voluntarily” chose to make their code open source. No Code and now AI is just the natural evolution of programming.


The issue is not whether it is ‘art’. The issue is whether it’s copyrighted and what you agreed to in the Terms.

  1. Copyright is very broad. Could cover the ‘expression’ in our Bubble apps.

  2. Haven’t read the terms but you might start there. Pretty confident it will be ‘silent’ on this point and this you will get into fights over stretching the meaning of wording (that was not intended for this application) to cover your situation.

  3. If your material is copyrighted, using copyrighted materials to ‘advance an AI model’ is not itself a violation as there are exceptions for advancement of models like these for educational purposes and the like.

  4. Using that trained model for commercial purposes might constitute a violation. However facts matter here. E.g. making a novel for commercial purposes from an AI model trained exclusively on Stephen King, for example… almost definitely copyright violation. Making a novel for commercial purposes from an AI trained on hundreds of different authors and styles… likely not a copyright violation. That means unless they’re using your content exclusively. @zelus_pudding, you may be out of luck.

  5. Even if it is a violation of your copyright or the Terms… You have to prove it. Prove that Bubble used your materials (as opposed to anyone else’s) to train it, prove that these actions had a material impact on the AI model, and then prove that you personally suffered damages as a result (all of which will need to be proven at trial court, and the inevitable appeals courts) (you could hope for a settlement).

Finally… the way to achieve all this ‘app creation’ (via AI) is not through prompts, it is through fine-tuning and/or embeddings.

1 Like

@zelus_pudding, I made a thing for you (again)! :slight_smile: So much of this stuff is easier to talk about than to write about and I wanted to show what GPT-4 “knows” about Bubble (way more than GPT3.5, but not quite enough just yet) and also what it knows about Bubble plugin development (not enough to get it right without additional training). So, here’s a talky video:

Edit: I suppose @josh might be interested in this video as I demo my “teach GPT to write Bubble plugins” prompt, so I am summoning him here.

Also, I hate how long Loom takes to generate a valid embed. Sorry the video doesn’t just show up in the page yet. I’ll relink it in my own related thread later.


Proper embed! (Pro tip, you can watch my videos at 1.2 to 1.5x speed and still understand me in most cases.)

That’s exactly the uphill battle with this thing - it would be hugely difficult to prove an individual app/unwilling contributor’s content was used for training purposes. What we do know about how these things work is they require a collective’s intellectual output as source material to be useful. So from a legal perspective I imagine the approach is for the authors to pursue a claim of damages as a class. If you have even one author with a fair claim against the AI creator who’s case is allowed to proceed, then legal discovery would show the other authors/content that were swept into training (which could be used for additional legal action).

But I think we’re getting ahead of ourselves here with this legal talk. I enjoy Bubble immensely - if given the opportunity to invest $20k in their business, I would have done so at any point over the last 4 years - it’s a great product. Last thing I think anyone wants is legal action.

To @keith’s point, if it were possible to train the AI without showing it examples of real apps built on Bubble, but just by explaining to it how Bubble works - I think that’s fair game. And if you could improve it’s ability to make such sites by showing it screenshots or even code of the public facing website you want it to copy, then to me that’s not vastly different from Facebook copying Snapchat’s “disappearing story” feature or YouTube copying TikTok’s vertical video scrolling feature.

But for the company I use as a hosting provider, code authoring provider and data storage provider to take not only the public parts of my app but the private ones and say “Hey robot, make this guy’s app easy for anyone to replicate” to me that’s fundamentally wrong and anti-competitive. That’s closer to Amazon using it’s 3rd party seller data to create new lines of business that compete and even bankrupt the 3rd parties than it is to Youtube copying TikTok.

Starting to feel like a broken record now - just make this opt in: it should start opted out by default.

Oh boi, I see Keith’s been typing for a while :slight_smile: I’m expecting nothing short of a masterpiece

1 Like


“hey 2024 bubble assistant, create XYZ and save me 3 days of trawling through the forum… or reading Keith’s 50 pages long posts” < the one he’s currently writing

So, IANAL (I Am Not A Lawyer) and it is the height of “internet” discourse when anyone (even lawyers, unless they happen to specialize in this area) starts going on about copyright and/or trademark law (which is neither simple, logical, nor internally consistent and, of course, like all areas of law, has an element of locality). The fact is that the legal definition of a copy is highly variable by context and while at its core holding or claiming a copyright is the right to control how some original work that has authorship can be reproduced… well… it’s complicated.

Not gonna belabor this point much more, but I think it’s fine if you or anyone else doesn’t want “their app” included in some training dataset. But that’s not actually what’s needed to accomplish the desired task anyway. (Bubble knows how Bubble apps are built. Bubble knows the transform between any app as represented in the editor and its representation in the runtime. That’s just software. This isn’t a mystery that needs to be solved by taking some big collection of Bubble apps paired with a natural-language description of what they do and comparing that to the output of Bubble.

The thing that (for example) GPT-4 doesn’t know how to do is, if there’s some Button X that when clicked fires some workflow Y, it doesn’t know how to write the JSON for that that would represent that in the editor. What a large language model needs to be trained on is: Here are the rules for representing the various things in a Bubble app and here’s the entire collection of those things. It doesn’t need to know about anyone’s particular business logic. That’s not the missing part. The missing part isn’t what you know and what you’ve done (the models already pretty much know this), it’s the blackbox part of Bubble.

Like I demonstrated: The plugin API is at the moment mostly a blackbox to GPT-4. In the same way, the representation of a Bubble app and its individual components are a blackbox to GPT-4. Programming concepts and business logic are not a blackbox to GPT-4.

Back to the legal points for a moment: By using Bubble you’ve already agreed to (the legal version) of this:

Any legal claim involving these Terms, the site, or the services will be settled by arbitration, not through a trial in front of a judge or jury, or through a class action lawsuit.

Also, You are responsible for all text, images, photographs, or other materials provided or uploaded by you or Direct Users associated with your Account to the Platform or Site (“Direct User Content ”) and, you have ownership of those - “Except for the limited licenses granted to Bubble in Sections 5(c), 6(d) and 6(g), as between the parties, you retain all right, title, and interest in and to your Direct User Content. For the avoidance of doubt, you may re-use any Direct User Content, including Bubble Site workflow and design, on other web platforms or media without restriction. Direct User Content includes User Components.”

Now, the “Direct User Content” does also include “Bubble Site workflow and design” and Bubble makes clear that you have the explicit right to reuse them on other web platforms or media without restriction (it’s unclear how one would reuse workflows on some other platform, but I guess that means “business logic” writ large … Business logic and business ideas are not really something copyrightable, but may in some instances be trademarkable and Bubble claims no ownership of your Marks, should you have any).

But you have also granted Bubble, its affiliates, and any applicable Third Party Services a nonexclusive license to use such Direct User Content during the Term (as defined in Section 11) solely for the purpose of operating the Platform and providing related services. Bubble is acquiring no rights in the Direct User Content except for the limited license set forth above.

While that license is “limited” the limitation is pretty broad (“providing related services”). Anyway, I just wouldn’t be particularly concerned about this. Basically, Bubble claims no ownership of your “app”, isn’t going to sell it, etc., but it does claim some rights to analyze things including aggregated Direct User Content, so…