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

This still looks bothersome to me, too, @BlackbirdSingingintheTrees & @MyrtleLion.

For one thing, I can't see why text for moderation would be sent during composition. It only requires evaluation after being posted, when it reaches the public domain. The poster's thought processes while writing are of no external value.

Most importantly, the (partially?) revealed instruction is CLEARLY a reply template being applied to the specific post. If not for creating automated replies, why the hell would this template exist? This isn't for moderation purposes, the purpose is obvious.

I saw some people upthread suggesting that OpenAI is writing these instructions for itself. Again, what led it to decide it should compose detailed replies to posts? The rules and admonitions contained in the instruction are nuanced and specific: how many of these rules exist and what are they based on? There must be at least one foundation document describing MN attitudes and approaches.

It still comes down to evidence that Mumsnet has the active capability to publish automated replies. Even if it's not currently being used, this is concerning in itself. If Mumsnet didn't know it had such capability, that's even more concerning!

As MyrtleLion pointed out, all this must be using a hell of a lot of tokens. Before title suggestion was suspended, MN servers were sending out continuous title prompts during composition. Now it transpires that it's also sending continuous prompts to evaluate the composition for meaning and compose a reply. What a fucking waste of resources.

I agree with your point about male-skewed content, Myrtle. I'm all in favour of AIs absorbing as much female-written material as possible, and this site's a valuable source. However, it doesn't need our words before we've posted them. And it doesn't bloody well need to write replies.

I'm with BlackbirdSingingintheTrees, GarlicEverything and the others who still aren't convinced.

MyrtleLion, I appreciate you're trying to make sense of it, but your explanations seem to be getting increasingly complicated, and less plausible. I do find that rather curious.

The latest theory about the moderation system and title generator getting their wires crossed doesn't add up. The leaked instructions weren't about moderation. They were about judging whether someone was being unreasonable, giving advice and writing a suggested reply. Why would a moderation system need any of that?

There's no evidence that two separate AI systems were involved or that their outputs somehow got mixed up. It's speculation, just like the earlier suggestion that the AI was writing instructions to itself.

The claim that AI can't link posts across usernames is not accurate. There's an established field called stylometry, or authorship attribution, which looks at writing patterns, favourite phrases, punctuation and other quirks to identify whether posts were written by the same person. Add in all the little details people share about their lives over the years and it becomes even easier to join the dots. External AI tools can do this with publicly available posts too.

I don't think those still questioning this are being difficult. Quite the opposite. I think they're asking the right questions.

As someone who knows very little about code and AI, I don't understand how 2 instructions for different operations (regardless of what those operations are) could be mixed up so that the results appear in the wrong field simply because they were sent 'close together'. Or how it could happen at all.

Is it possible to explain that part in a way that makes sense to someone who last programmed in BASIC?

Surely AI cannot have programmed that response on their own so someone needs to have put that together?

I’ve been chatting with my DP this morning who is a senior software engineer who also works with AI coding and he suggested that it might have come from, in his words, “a janky third party” system used to integrate the Open AI feature into Mumsnet’s site. Just adding to speculation about how it could have occurred. Not claiming he knows, just thought it was an interesting theory.

The biggest problem MN has here is the loss of trust.

For those of us who were doxxed a few years ago by the intern with a grudge, and who have been given ‘strikes’ and threatened with bans for speaking the truth, we know that MNHQ isn’t always on the side of regular and loyal users of the site.

Of course it’s a business and has to answer to people way more important than the ones producing its content, but we’ve come to be a little wary of being fobbed off. The early replies to this thread being a classic example. Maybe it’s just grown too big for its boots, but I’m beginning to feel like those posters who reminisce about when it was all fields round here. AI may be the death of many things, MN included.

BTW, my posts are still being saved as drafts. Saved to where?

The most plausible explanation is that MN have at some point been experimenting with a tool that could reply to AIBU posts, which used a schema called AIBU response schema v1.0. There was too much structure in the response (de-escalation instructions, to deliver a forum style response etc) for it to be anything other than a real template. An LLM designed to generate a title would have system instructions to do just that, and would not go rogue like this unless there was something else to get confused by.

As to why this schema got spat out rather than a title, who knows. Maybe there was a configuration bug and for a moment, the title generator was accidentally pointing to the AIBU reply test environment instead.

That to me is most likely.

There are definitely a few accounts that feel like they are responding to posts in the form of these V1 instructions for various topics. Check out https://www.mumsnet.com/talk/property/5586809-are-sellers-being-unrealistic?utm_campaign=thread&utm_medium=share for an example - the current final poster sounds so AI it’s disconcerting and I can’t figure out why a genuine poster would run their responses through AI like this. Hopefully this is vague enough to not be troll hunting 😅😅😅

I think that when posters use AI to write their posts, they need to expect a bit of cautious pushback in the responses.

@EmpressaurusKitty yes my tinfoil hat conspiracy theory is that they aren’t posters using AI - they are bots monitoring the topics who are responding via AI instructions like the one uncovered by OP. Whether they are made by MN or something else is uncertain (as is my theory itself lol)

And if they’re well trained enough they might realise they need to add a note saying ‘I’m using AI because….’

Definitely, but contrast on this one - very weird!

www.mumsnet.com/talk/property/5580790-would-you-stretch-for-a-ps775k-flat-or-stay-within-budget-clapham

Bloody hell 😮

@BlackbirdSingingintheTrees hmm true those seem human! Maybe just a poster using AI then. sheepishly puts tinfoil hat away

Oh I was getting more tinfoiley @Attenboroughsmistress and thinking about account hacking by AI!

I'm loving the (new to me) word "janky". It seems to merge several ways of being bad all into one word. 'Clanky/clunky' - the kind of metallic noises of an inefficiently configured machine. 'Junk' - crapola software infesting our lives. 'Manky' - all-round bad and distasteful. Jagged - sharp edges creating disruptive catchpoints. Great stuff.

@BlackbirdSingingintheTrees oooh the plot thickens! Or somehow sold/handed over to AI, like a real person seeds the account with a few posts and a persona and then plugs it into a bot and away it goes from there? Full disclosure: I do not really know what a “bot” is or how they work 🤣

I know they’ve switched off the AI but every post I type now I’m thinking of it being read while I’m drafting it. It’s unsettling.

On twitter there's an extremely prominent form of AI bot post that is some extremely outrageous scenario that plays out over several tweets, with the idea that each tweet ends in a way that makes you want to read the next one. This boosts engagement figures for that bot and presumably brings in more money.

They are written like this

All on separate lines

One sentence per line

With the scenario dragged out for as long as possible

Anyway, I saw an OP like this on MN that was so obviously AI but was dragging posters in that I reported it to MNHQ who couldn't see any issue with it.

What I don't understand is why bots would be posting like this on MN which isn't monetised and you don't get rated on your followers or engagement. Are they using it to train bots to get more engagement on twitter where it is monetised?

I think there are TRAs who occasionally troll FWR who do this. We’ve had bits of code posted before. Either that or the bad actors want us to think that’s what is happening to intimidate people.

One of the clear monetisation aspects is that Google SEO used to be the way to achieve hits on your website/advertisements for sales.

That has dramatically shifted with AI, where actually the top results if you search for something on AI are coming from forum discussions, so Reddit and MN comments.

You will see if you search for “best cot” or whatever, AI will summarise the brands discussed on forums. So, there is a motive to seed fake “real” accounts throughout forums that can then mention products and brands to get these coming up in AI search results.

I could see a clever strategy where you attempt to seed thousands of accounts for various topics, and then sell comments. Like influencer marketing but forum comment marketing.

So a property/DIY account might spend a few months setting up as a “real person” and then casually mention a particular real estate agent in the future, causing this company to get a boost in AI search hits.

Very cool interpretation, I’ll pass that on to him!

I think this is definitely very plausible.

Also, Mumsnet sells itself for market research etc based on number of users / visitors and so on...

https://advertising.mumsnet.com/workwithus

I can see why you think that and that was our first thought but we do think it’s a moderation schema. I will lay out the reasoning for you later but I’m about to go out. I’m really sorry as I never go out and I’ve delayed talking about this subject twice because I’m going out, so I can see it might look like an excuse. Honestly it is not an excuse but the schema will take a bit of time to explain. I was going to write it out for everyone last night but my husband said it would be boring to read.

Also I was up past 2am this morning to show my reasoning after looking at it for a long time late last night.

I agree that a timing clash is probably not the reason why it appeared in the title field but it is not clear why it happened. Essentially my husband said, “it’s a bug, they need to fix it,” and the responsibility for working it out and fixing it is on MN.

When I’m back I will set out the schema and what it is trying to do (from my perspective). I promise you I have nothing to do with MNHQ.