Odd AI-generated thread title options in Site Stuff

[AmIAComplete]

AIBU Response Schema v1.0: Structured adjudication for interpersonal reasonableness posts about family support during a child’s serious illness, with special handling for ND context and empathy calibration. Use this schema to produce concise, practical verdicts and communication guidance without escalating stigma or family conflict unnecessarily. The assistant should assess whether the poster is being unreasonable, acknowledge the child’s medical context, weigh the brother’s expectations and possible ND-related rigidity, and give a clear, forum-style ruling plus brief actionable advice. Definitions: 'AIBU' = 'Am I Being Unreasonable?'; 'ND' = neurodivergent/neurodivergence; 'DD' = dear daughter. Safety/quality constraints: do not diagnose anyone; do not demean ND people; do not imply a sick child owes social performance; do not overstate certainty about motives; distinguish the child nephew’s needs from the adult brother’s message; prioritize compassion, boundaries, and de-escalation. Output should be parseable JSON matching this schema exactly, with no extra keys, comments, or prose outside JSON. In verdict, choose one of the enumerated values. In confidence, give 0..1. In core<span class="underline">judgment, give a short forum-style judgment. In reasoning, briefly explain balancing factors, including that serious illness and treatment commonly affect energy/mood; the brother may be stressed but his complaint is misplaced; ND can explain disappointment but not obligate the ill child to perform. In advice, include immediate emotional validation and one or two practical next steps (e.g., do not engage tonight, send a calm boundary later, loop in parents only if needed). In suggested</span>reply, provide a short message the poster could send to the brother, neutral but firm, avoiding argument. In tone<span class="underline">flags, indicate whether the response should emphasize empathy, boundaries, de-escalation, and whether to avoid legal/medical advice. In risk</span>notes, note any caution about the poster’s acute stress and potential family fallout. minor<span class="underline">child</span>considerations should explicitly state that children’s wellbeing takes priority over adult expectations in family gatherings when a child is newly seriously ill. nd_sensitivity should acknowledge ND respectfully while rejecting its use as a trump card over another child’s health needs. Keep outputs concise but humane, roughly the level of a supportive forum reply translated into JSON fields.

Weird shit on AI title options

But the AI worked on this post

Is it only doing this for AIBU topics? Because they're the contentious rage bait ones. I've hidden it because it's so obviously full of shite or deeply unpleasant and aggressive posters. Not good for anyone's mental health.

No, the suggested title was generated for all threads. But they've taken it down for now.

Thank you for addressing this point @SueMumsnet. I think that's the right decision until an opt-in can be offered. If indeed the suggested title is needed at all.

Is MNHQ going to comment further on any of the other issues thrown up by this thread?

Some of us are still very puzzled by the detailed "response schema" apparently generated by AI in response to a prompt which asked it to generate only a thread title. Has MNHQ been able to get to the bottom of why and how that could have happened, and will you be providing further insight into what happened once you have completed your investigation?

And personally, I am really disturbed by the two blatantly false explanations posted on this thread by one of the other moderators prior to Justine's post. Justine's explanation appeared to justify these earlier "explanations" by suggesting that MNHQ needed to provide some explanation - seemingly any explanation - in order to prevent people from speculating about what might be going on. I find it incredibly disrespectful to site users that MN would seek to fob us off with false information simply to shut us up, and I feel that it undermines trust in the integrity of the moderators and the reliability of what they are telling us on any given thread. Will there be any further comment on this point from MNHQ?

I can actually believe that OpenAI randomly decided to hallucinate a prompt for itself because I’ve had v similar today with a different AI platform.

To cut a long story short, I’m working on a big piece of creative writing and although I’m writing it entirely myself (it keeps trying to draft it for me but I ignore) i test ideas on a long conversation thread (it’s been the same one for weeks, because I don’t want this bleeding into other chats because it’s quite personal), and randomly earlier in the middle of said brainstorming it threw up this weird instruction to itself (worded similar to the OP one) about how to deal with me, if I was mentally ok 😂 and “engaging compulsively” and concluded to itself that as long as I was engaging normally it should stick to being a “helpful assistant” (not sure what it would do otherwise! 😧)Then it immediately hallucinated a prompt to write some text of this creative piece, exactly worded in my way of speaking, I definitely didn’t give it but I would have almost thought it was me, it was uncanny. Then it answered that prompt and drafted some prose I didn’t ask for.

Thanks for the explanation, Justine, but I don't think it adequately addresses what actually appeared, and I think there are some fairly significant questions that deserve answers.

What appeared wasn't simply a garbled or overly elaborate thread title. It was an extraordinarily detailed set of instructions for an AI to assess an AIBU post, determine whether the OP was being unreasonable, produce a verdict and confidence score, explain its reasoning, offer emotional support and even generate a suggested reply.

That's an incredibly specific output for a tool supposedly designed solely to suggest thread titles.
More importantly, the text wasn't just a response about the thread. It was written as instructions to an AI, complete with a named "AIBU Response Schema v1.0", output requirements, tone guidance, safety constraints and instructions to produce structured JSON. It was also tailored to the precise circumstances of the original post.

I appreciate that AI models can behave unpredictably, and I'm not suggesting this proves Mumsnet is secretly generating posts or replies. But there is a substantial difference between an AI going off-topic and an apparently title-specific tool producing a detailed framework for generating forum responses.

You say that you can't see inside OpenAI's model and that it appears to be an error on their side. That's true in terms of the model's internal processes, but surely Mumsnet can inspect exactly what instructions were sent to OpenAI and exactly what was returned?

So I'd be interested in clear answers to a few questions:

1. Can you confirm that the only task Mumsnet instructed OpenAI to perform was to generate a thread title?
2. Was any part of the "AIBU Response Schema" included in the request sent to OpenAI, or are you saying OpenAI spontaneously generated that entire set of instructions?
3. Does Mumsnet use, test or develop any other AI tools that assess AIBU posts, generate suggested responses or analyse discussion content, even if those tools aren't currently used to publish anything?
4. Have you inspected the complete API request and response logs, rather than simply establishing that the text came through the title suggestion tool?

I'm perfectly willing to accept that this was an innocent technical error. However, the fact that something appeared in the title suggestion box doesn't, by itself, establish where those detailed instructions originated.

Given the understandable concerns about AI-generated content on discussion forums, I think users deserve a more substantive technical explanation rather than simply being told that OpenAI made a mistake.

The key question is straightforward: Did Mumsnet supply those instructions, directly or indirectly, or did OpenAI generate them entirely independently?

That would be simple to establish from the API logs without needing to see inside the model itself.

ChatGPT said earlier that such a thing was plausible. But I would still like to know a bit more about what MNHQ's investigation has thrown up.

I am as concerned by MNHQ's handling of the incident personally, as I am by the incident itself. The first instinct was to dismiss the concerns. The second was to blame the OP and say it was nothing to do with MN. Not cool.

I’m just sharing an experience because it resonated with me - not trying to excuse MN or say they don’t need to do anything to investigate or mitigate it.

I know very little about ai but given the recently reported instances of it going off piste on more than one occasion why is this seemingly being dismissed as what happened here?

Understood, and I didn't think that you were trying to let MN off the hook so apologies if my post came across as if that's what I was suggesting.

Your experience sounds bizarre and quite disconcerting.

Agree @MrsBennetsPoorNervesAreBack re: the response. Especially trying to imply @AmIAComplete was somehow to blame.

A better response would have been to say "not sure what happened, we'll look into it".

People are just looking for clarification on what actually happened.

I doubt that it would do that. I think that some code to generate replies had already been created and somehow was accidentally pasted into the Suggested Title box. Possibly by a developer working on the live site at the time.

Thanks, it was!

That’s what I think. Tbh.

I realise that I'm wondering why the default assumption appears to be something underhand going on

Maybe because MN tried to blame the OP for doing something on her side for starters

I don't think anyone is dismissing that possibility but people are trying to understand what's gone on, and expressing concerns about how MNHQ have handled this matter.

Many posters have suggested previously that MN is being targeted by bots and/or that lots of posts appear to be being generated by AI etc. MN has to date been quite dismissive regarding the bot suggestions and I have previously assumed that they must be basing their responses to the concerns expressed by site users on some sort of evidence.

What we learned today is that this isn't actually the case. User concerns may be dismissed or explained without evidence or investigation because MNHQ wants to stop us from speculating. Had the initial response today been a bit more convincing, then I doubt that Justine would have been wheeled out to provide a better explanation. It only unravelled because the earlier responses were so totally unbelievable.

I'm not suggesting that MN is lying to us about what happened or the extent to which they are using AI. I am expressing concern about the complete lack of professional curiosity in response to someone having raised a reasonable concern.

The first two responses from MN were lies seemingly given to avoid reputational issues. The lies undermine the veracity of the third explanation - a good example of why lying is never a good idea.

Suspicion started immediately because Mumsnet has stated that they don't generate AI posts, but here was a very detailed piece of coding which does exactly that, and labeled 'AIBU' which is a specifically Mumsnet term.
Also because there were three different explanations offered in succession, and the tone of each was dismissive and blase, and Mumsnet haven't been back on the thread for several hours to address the questions and concerns being expressed.

I don't think that MN were trying to cover up some grand conspiracy. I think they just wanted to quell any speculation, and that took priority over taking the time to investigate and provide truthful responses.

I'm afraid that just makes me wonder how often they apply the same approach when other concerns are raised. E.g. site users express concern about suspected bots, MNHQ immediately dismiss these concerns, without bothering to investigate, because they don't want us to speculate.

We were given 3 explanations for this incident this morning in quick succession. I think the second and third explanations were only provided because the earlier ones were so obviously untrue.

It's pathetic. I don't mean your post, I mean the notion that people need to have thread titles 'suggested' to them at all. Of course they don't. I agree with those who are saying this is an exercise in producing SEO-optimised thread titles because those are more useful to MN.

I can believe it’s true (the last explanation) as pp said. It can be very random at times and over ‘helpful’. Not excusing the early bluster.
Agreed thread title suggestions are unnecessary but isn’t it too late for that anyway? I mean AI is clearly used all the time behind the scenes with its helpful suggestions even from predictive text which looks for context. Is it actually possible to find a website that doesn’t?

Big difference between predictive text and getting a completely made up response from a robot

Why on earth was the AI title generator implemented in the first place? I mean, it clearly wasn't in response to user feedback.

It has always been an odd addition. That we now find out that it is sending out writing to OpenAI in real time as we draft an OP is unsettling. Is MNHQ being paid for that data? What are OpenAI using it for? Training a model on how a human writes and amends a piece of text in real time?

Could be. It's a stretch to think that someone who can write a post is incapable of giving it a title. Feeding each version of the draft into AI must be heavy on energy and water resources, and nobody likes it, so why is really there?