I just had a bizarre conversion with the help AI. It gave me bad advice twice, then completely denied having said the things it just said. Even when I pasted screenshots and text copies, it still denied it and accused me of lying to it. Either the AI is going insane or I am 
Anyway, its definitely malfunctioning in multiple ways, so I thought it was worth mentioning here.
If you happen to know which model was being used, along with any relevant settings such as the reasoning level, that information could be useful. Different models and reasoning settings can behave differently, so those details may help narrow down what happened.
Note:
I am a regular on the OpenAI Discourse forum, so I am fairly used to seeing reports involving contradictory AI responses, loss or misinterpretation of conversational context, and cases where a model incorrectly denies something it said earlier.
From what I have observed there, reports this extreme seem less common now than they were when reasoning models first started becoming widely available more than a year ago, although that is only my impression rather than something I can quantify.
Thanks - how do I check which model is being used? This is my first time as admin of a Discourse forum, so I am still getting used to it. I don’t recall ever making AI-related changes, so I assume the AI is the Discourse default.
This is a hosted forum btw.
That is exactly the question I would be trying to run down if I were an admin following up on a report like this.
I am an admin on another Discourse hosted forum, so will do a little digging.
Also asked ChatGPT to check my reply and it added the following which I am now going to verify but wanted to pass along so you are not waiting.
From the Discourse AI documentation, administrators can inspect the configured models under:
Admin → Plugins → Discourse AI → LLMs
Discourse AI features are also backed by personas, and a persona can be configured to use a particular LLM. The current AI feature configuration can show which persona and LLM are powering a particular feature.
Useful references:
What I have not yet found is a documented way for an ordinary user to inspect a completed Help AI conversation and see the exact model that handled that specific interaction.
If someone knows whether that information is exposed to users, admins, or in the site logs for an individual request, please add it here. That would be useful for reports like this.
In reviewing the AI-generated section above, which I am leaving unchanged for transparency, I would make a few corrections.
First, I would remove the Discourse AI – AI bot link. It does not appear necessary for answering the specific question about which model is being used.
I would also remove:
What I have not yet found is a documented way for an ordinary user to inspect a completed Help AI conversation and see the exact model that handled that specific interaction.
That limitation is not especially relevant here because you noted that you are an admin, so you have access to configuration information that an ordinary user would not.
Finally, now that I know where the relevant configuration details can be found, I would also remove:
If someone knows whether that information is exposed to users, admins, or in the site logs for an individual request, please add it here. That would be useful for reports like this.
At this point, the more useful path is to inspect the Discourse AI configuration and determine which LLM are assigned to the Help AI feature.
There is a very long list of AI settings, but these are the top few. Let me know if you need to look for anything specific, or I can post the whole list.
EDIT: I have a Claude account entirely separate to this and Claude is fairly helpful with Discourse queries, but I would expect the default Discourse helper agent to be better than a general AI.
As I am not an expert on every current Discourse AI setting, this is my best understanding.
On both my site and yours, the setting shows:
Default LLM model: CDCK Hosted LLM
From what I can find, this is Discourse’s hosted LLM service. More precisely, the current CDCK/MoM implementation is a Mixture of Models routing layer rather than one fixed model. It can route requests to different open-weight models depending on the task. Discourse has a useful explanation here:
Another useful place to check is:
/admin/plugins/discourse-ai/ai-llms
That is the Discourse AI LLM configuration page, where administrators can see the configured LLM connections and which models are available for Discourse AI features.
For example, on the site I am an admin I see:
which will indicate if that is backing the action you used.
That appears to answer the question I asked about model and settings.
At first glance, I was a little surprised by that list. Then I wondered whether you might be on the Free plan.
In checking your earlier Changing to Full Width topic, I found:
I am on the 14-day trial for the Pro version
So that satisfies my curiosity.
Another example so you know it is not just you.
I am on the full pro plan now, so it should be the default for that.
Also I’m much more comfortable with the Discourse width now and have made the necessary adjustments. It seems weird looking back at older full width forums
At this point, there is not much more I can add, but I did not want to leave you wondering whether I was still looking into it.
I am glad you posted this. I expect others may encounter something similar in the future, and this topic now provides some useful documentation and a clearer path for investigating it.
