Hi! This topic is way out of my league. But I still want to discuss it with you, if you don't mind.
I support Joplin’s AI direction. I even defended it a little on social. No complaints.
But starting from the 3.7 announcement I see a lot of negativity on social platforms. And I don't think it is possible to fix by educating or persuading people. The reactions below include people reporting uninstalls, migrations, refused upgrades and removing Joplin from their shortlist. Some misunderstand the existing safeguards; others acknowledge that AI is optional and still want it removable. These examples cannot tell us how most users feel, but they show a concrete risk to user retention and trust.
Some examples
AI searched, but I saw a lot by myself:
Reinald — Mastodon, September 14: Initially objected to AI and asked how to switch it off. Subsequently reported uninstalling Joplin.
“How to remove the AI settings and Chat Bubble” — Reddit, September 11: The original poster reported attempting a downgrade to remove the AI interface. Another participant challenged the training concern but still wanted all AI references removable.
But is it possible, that in this technical and interesting discussion you (very tech-navy people) forget how users would feel when they see these changes? What if their negative attitude to the AI-related staff will harm Joplin's reputation? What if the main reason they chose Joplin - is privacy, not AI-productivity?
Don't you think, that based on users reaction - AI functionality possible and should be implemented as a separate downloadable on demand native module? Is it possible to ask installing this module in case people want some additional functionality or plugin?
I realize this would involve packaging, compatibility, and maintenance work. But maybe it is the time to step back, think about another architectural solution, post about it on social, and implement it?
Not a fan of AI here too, but to be honest, the way Joplin has implemented it doesn't really bother me… It's disabled by default, and it will stay off unless you explicitly open settings and enable it there.
The AI button should be easily removable though, either automatically when AI is switched off in the settings, or via a right-click option. Right now, the only way to get rid of it is to use custom CSS.
"many" is how many? And is that users we want to retain? Because the ai-hate crowd is particularly vocal while we don't hear about the vast majority who either want AI, or understand that this is where most apps are going, and perhaps appreciate the privacy controls we provide.
There won't be a different build. That will be way too complex to maintain, and that would be just for a few hardcore users who want to live in the past, not to mention there's no reason to think they'll actually stay on Joplin. They can move any time, it's free, while we get stuck with the massive maintenance costs forever.
And just to be clear: we're absolutely ok to discuss AI features, address privacy concerns, etc. I'm not dismissing all critics. But I think a part of that is irrational AI-hate and FUD, like the guy on Bluesky who's scared we're going to accidentally send all his data to OpenAI
@laurent I agree with everything you said, and I'm on your side - I use these features a lot.
But for many people privacy and AI are opposites, and most users chose Joplin as a privacy app. Notes are personal. Joplin used to give a feeling of protection- and suddenly for some it doesn't anymore. It's less about features than a paradigm change: safe/private/mine => AI/productive/shared.
I can't measure how big that group is - only that about half my social notifications mentioning "Joplin" are now this. My feed, not a survey.
we get stuck with the massive maintenance costs forever
Maybe there's a cheaper shape. Joplin already has something useful in the UX: Web Clipper. You go to a page, you enable it, you connect it.
What if AI worked the similar (or close) way - a plugin you install from the Plugins screen? The AI settings section and the chat button appear only after installing it. Functionally the same as ai.enabled = false today, except someone who doesn't want AI never sees an AI page at all. I believe many users just don't like seeing this "bot" icon in their app settings.
The embedding and vector infrastructure stays in core, as you decided in the embeddings thread, so the native dependencies still ship. It removes the surface, not the code. The idea is to move AI switcher to the plugin, which people associate with something separate from the core app. Make this plugin like a key user should use to enable AI.
The obvious cost is discovery: a plugin nobody finds is a feature nobody uses. So it is not a solution, just my thoughts.
I remember similar privacy scare over background network calls ( 1, 2). The whole debacle got solved with a built in welcome note that explains what is the reason for each background network call.
Of course, surprised people start the conversation with accusations and anger. Here anger isn't resolved by changing stuff around to make it less likely to discover the issue. Instead I believe publishing more information about privacy cognizant (or ai critical) workflows is much more productive.
I think as a community we can outline few tested workflows with local models. How to install, choose, connect and use them. We can explain in more detail what data will be received by which party of AI chain.
I think it's important to demonstrate tangibly that needs of AI critical users aren't baseless nonsense and in fact are being addressed as a part of the app development.
P.S. Joplin can't win over zero-trust folks because the app isn't built for them. Yet while staying true to its core Joplin can give normal nontechnical users (millions of them) unprecedented level of privacy. Over time it keeps increasing and now we even got local note encryption!
I've posted an article summarising the AI-related privacy features in Joplin, in the hope that it helps.
We can't act as if AI isn't there, or as if it doesn't matter, so I think the best option is to be transparent about it. Some users won't be comfortable with AI features for the foreseeable future, and that's fine. But another group simply doesn't realise that these changes really aren't such a big deal: they're all disabled by default and comes with good privacy controls, and in that case good documentation can help.
The AI plugin could also lead to some interesting features, maybe based on local models. I'm sure there are ways to make use of AI without sending all our notes to the big LLM providers, but the only way to find out is to release the features and experiment with them.
Semantic search is one example that's already fully local, and it already gives very good results.
Yes, I saw your article. I think I will point users to it if I meet negative comments again.
BTW, I saw a complaint that embedded semantic model increased app size. Maybe the future is to give user possibility to decide himself - which exact model he wants to use, if any. We can show in Settings that this one is default, but user can uninstall it and download another.
Do these models behave differently for different languages? If so, it might have sence to give user several models to pick. With filters: e.g. size, language, accuracy, speed.
The model itself would not increase the app size because it's downloaded separately, and only if the feature is enabled. But the installer size did go up, probably in part due to the ONNX Runtime. It could be worth reviewing this and see if it's possible to optimise it
Just my small thoughts about that :
when AI support was in plugins only, I have tried them, i have even written my own plugin to test using local LLMs in Joplin. But for me it was quite amateurish.
I was waiting to have a real Joplin feature, tested like all the other core features and with the good quality we know about Joplin developments. And I'm happy with it, and I'm trusting it more than I would trust a separate plugin.
Joplin has always been number one about privacy and security and adding a feature that is disabled by default is the perfect way to add AI feature.
Complains about adding an AI feature or increasing app size should be non relevant when we have such secured and quality app. We can't just close our eyes and say that AI doesn't exist. And working with local LLMs can be perfect if we want to keep notes secured. People just don't want to read how the feature is implemented and just complain when they see these two letters, that's sad.
Only this - a small comment. Now the joplin-mcp is better (IMHO), I am waiting some sort of synergy - whether joplin-mcp will work faster due to current changes, or native implementation will have the same amount of useful tools and token-efficiency.
Oh, is it? Embedding as well as search?
Maybe you could add a note explaining that in the relevant section of the settings, then?
I wanted to try AI features as well as the semantic search, but waited for a long time until I got my hands on some AI credits I could use. If there are some features I could have used in local-only mode, I'm still not sure what they are, and would like to know.
I saw the reaction on Bluesky. If I can be frank, I would argue it is was primarily a messaging issue with your marketing. There is a large amount of concern and worry about AI being invasive, especially with unregulated data scraping running rampant. I frankly don't blame anti-AI people here. If you wanted to better calm this anxiety with this particular crowd, this should have started with and have strongly STRONGLY emphasized the key points
it's off by default (which is a small miracle in today's dark pattern software paradigm of opt-in by default), and requires extra work to even enable- something I personally appreciate.
Joplin is still extremely focused on privacy
And, I mean basically all the stuff you posted in this thread. This would be a far better opening message to send than "chat with your notes". imo, as a newcomer and longtime watcher, THESE are Joplin's strengths, and good things to market right now in this complicated situation of certain companies building data roach motels.
I'm not interested in using AI, but I don't think it needs to be a separate plugin either. Frankly my only irritation with it is that working on it's integration is taking time away from fixing the very buggy Rich Text Editor and various other major bugs holding Joplin back.
Interesting discussion guys. I think it's an appropriate moment for me to mention my recent post about fully switching off all AI features, which isn't currently possible -
My post got a fair amount of support and hopefully it's something that can be considered in the spirit of making AI fully optional.
This was an interesting discussion. I gained a lot of insights from this topic and learned more about the AI features. Personally, I'm interested in Joplin's AI functionality, but for information security reasons, I prefer using local providers (such as Ollama or LM Studio) over cloud-based ones.
I've been using Joplin for the past 7 years and I'm not going to switch. That said, I don't want any of the AI features, even just a disabled chat button at the top of my notes. I downgraded my desktop Joplin to 3.6 and I'm not going to upgrade my mobile app and that's how things will be for me for the foreseeable future.
I think Firefox and DuckDuckGo have done things nicely in making it possible to block AI features completely if one wants to.
It's unfortunate things are playing out this way. People who don't like AI are absolutely people worth retaining. This is a much, much larger demographic than you might realize. There was even an article not long ago suggesting Joplin for people looking for a non-AI notes app.
It feels like Joplin is getting worse PR than Firefox. Even though Joplin's policy to ai is explicitly opt-in (with granular access and capabilities), while Firefox and DDG have opt-out "ai killswitch".
With such granular access in Joplin, even intentionally you literally cannot to leak any of your data to AI provider without doing ~5 step setup process first.
It really confuses me that projects making more transparent and pro-choice features met with such hardliner reception.
