What does AI bring to conversations among humans?

Continuing the discussion from Discourse Version 3.2:

Since the focus for the upcoming Discourse release is announced to be AI, I’d like to take the opportunity to question this choice. After all, Discourse is made to support “thoughtful discussion and meaningful connections”.

Since the AI field is mostly dominated by huge corporations with deep pockets, and given that AI training is consuming large amounts of water, I’m genuinely asking why this feature is important to Discourse, the Discourse community, and the world at large. I can dig the “building forum software for the next decade” phrase but I’m not sure I want to live and converse with AI instead of humans — especially as humans seems to be requiring much less water over time for much better conversations. :wink:

The AI field is full of lies to attract investors. It seems that Discourse AI was introduced per customer request – probably some organizations running huge public instances and having to deal with tons of garbage. Sure, this is not for everyone, and probably appropriate if you have millions of users and must deal with that throwing more power at it…

I appreciate this disclaimer. But what happens to consent? Many people would prefer not to feed their posts to AI training, especially if they are on confidential communities or in restricted groups. There is a growing history of big corporations appropriating content for AI training without regard for Copyright laws nor consent. Is this really a good idea to feed them more high-quality content — and paying for it?

If there was a possibility to not consent to such personal data acquisition — and this can be a genuine demand under European GDPR Law —, wouldn’t this defeat the whole purpose of having AI policing a forum?

I really hope this will remain optional forever, and that proper consent will be displayed for this. Will it?

Well, I made this a separate topic from the plugin discussion because I think it deserves a wider take and did not want to derail the original discussion…

Here is a list of alternate titles “suggested by AI”:

  • The Ethics of AI Enrichment in Discourse :thinking: no, this is not an ‘enrichment’ I’m talking about, rather a concern for degraded human conversation…
  • Discourse and AI: A Controversial Proposition :disapproval: well, noooo… it’s not about controverse, but about human conversations!
  • Can AI Become the Future of Discourse Discussions? :face_exhaling: really… as in… bots talking to each other?
  • AI Integration in Discourse: Threat or Innovation? :triumph: bot, you’re not ready for the subtlety of language.

The AI title suggestions can be a touch excitable. :slight_smile: I think it’s clearly been reading too many glossy magazines in its spare time. Though I think if I could dial it down a notch it could be quite useful for topic splitting (which I appreciate is more useful to mods than everyone).

I started off quite sceptical about the use of AI for similar reasons. I’m more interested in what people have to say than how well an AI can polish it up. At the moment (and these things move quite fast) there’s also a certain tone that it generates that I think obscures/homogenises the ‘voice’ of the writer.

However, there have been a couple of Discourse AI tools that I think have turned out pretty swish. I love the Related Topics. :raised_hands: That has really had an impact on how I find and engage with different topics. As an avid forum reader, this has been a treasure trove. I’ve used it across a couple of forums now, and the rabbit holes of connected content I’ve lost myself down have been amazing. Also really great for finding answers and linking up topics as well.

Bert, the AI Bot, has also been brilliant. Especially now he has SQL Helper and Site Setting personas. I think these are more geared towards the admin side (and maybe my job :slight_smile:), but he’s been a real boon :trophy: (and I think we may actually be friends :slight_smile:)

I think what has won me over is how useful using the AI-based tools to help out with the everyday stuff has been. I still don’t think I’d want an AI filter on the contributions themselves, but having ‘clever’ helpers has certainly made my life easier, and allowed me to focus more on the things I like doing.


While we did use the word focus there, in my opinion more because of historical reasons, those are the internal teams we created to work in different areas of the Discourse ecosystem.

We do have a team focused in AI since March, where I’m the tech lead. We did a heavily experimental cycle during Discourse 3.1 so we are coming back again on the 3.2 cycle to add polish to the features we got good feedback from our users.

The idea is to still keep everything under the DiscourseAI, which keeps anyone not using the plugin shielded from any bugs introduced by it, and benefits everyone using Discourse as we create and exercise the software extensibility.

(citation needed)


I think you will be happy to hear that for this cycle we expend a large amount of time making every module compatible with open source models, so even the most powerful features can be ran in a GPU that you can find in a gamer PC.

That’s been a huge point for me, that we will not make features that are completely dependent on third parties, and I’m super happy with the progress we made here.

That’s true, why would you want AI moderation for a instance among friends? But it may be a deal maker/breaker for a instance where a lonely CM tries to handle a oversized community with a small team.

Discourse at scale is something we care about too, otherwise we’re leaving huge swats of the online communication for the walled gardens, and you don’t want that.

Nothing changes here: what you post in a Discourse instance is subject to that specific instance rules. If said instance prints your posts and sends it as postcards to unsuspecting grandma’s, they may need to make that explicit on their terms. Same for AI.

For the AI plugin, most of the “share with AI” flow is user initiated. You can ask AI to proofread, translate, etc. You can ask AI to summarize. Meanwhile bulk actions, like embeddings computation that power semantic related topics and semantic search are, by default, computed using OpenSource models that can be hosted in the same machines where you are running Discourse and safe from third parties.

And that’s has always been my plan on this project. It’s about building tools to make people conversation easier and reducing chores. Not about talking to machines.