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

Could a friend or relative have told you about the mystery? Or perhaps you saw the Guardian piece.

Most small to medium organisations use small to medium "off the shelf " software products rather than custom building their own software. These are then customised with wanted features being enabled unused features being left switched off. Sometimes using particular features incurs additional licensing fees (especially in this particular space).

I'm guessing that the product they use offers interfaces to OpenAI for differing purposes and they intended only to use it for title generation but an additional feature would enable full post generation. So by "switched off" they mean the feature in use has been switched off.

A simplistic analogy might be you buy a TV package but only enable channels you want to be used in your household - the others are still there but not in use or not licensed. You can then switch off channels previously enabled if you decide that James Blunt Hot Shower scenes isn't worth paying for or useful.

The fact that all the instructions for both features appeared in the OP's post doesn't mean both features were enabled/in use by design. The post is more likely to be a consequence of a fuck up in local customisation coupled yet another gap in testing (neither set of instructions should ever have appeared in the OP's post).

That doesn't answer the question of whether the data is still being sent. (And I take it you watched Traitors Uncloaked last night.😁)

I'd say hilarious shower scene rather than hot....

I am unmasked 😀

And yes for sure, the only people who can tell us which features they actually use and which data are sent to OpenAI are MN. I was simply outlining how a feature might legitimately appear in the exposed software instructions without it being in use by the licensing company.

Testing should have caught this before it went live, just like it should have caught the issues we find on every other software release historically. Testing is never 100% perfect but a good test architect learns and modifies the framework rather than coming back with the "but we did lots of testing" whinge which predominates on these threads. If the exposure was the consequence of a recent change/fix/customisation then I'd expect automated testing to have picked it up.

As I said upthread - if draft posts are going to OpenAI before being committed to public that seems inconsistent with the privacy policy. I'd be more concerned if the sweep of text analysis includes PMs.

"Hot" for marketing purposes 😀

That predates AI, but as AI has scraped mumsnet for data it will know all about it. And snapping and farting etc

😂 in the immortal words of our former Queen, recollections may vary....

I don't even like the AI title idea. If you can't come up with a title, you haven't got an issue worth discussing.

I definitely couldn’t have come up with this by myself

That was a baffling thread title.

Hi all,
Some answers to the questions that have come up since my last post.

First, on the prompt tool to be clear we’re no longer sending any draft text to OpenAI. When we switched title suggestions off, we stopped sending anything, not just stopped showing the suggestion.

When the tool was operational, all that was sent was the text of the draft and the topic, to generate a title. No usernames, email addresses or account details were included, and OpenAI doesn’t use this data to train its models.

When I previously said 'We have never tested AI-generated responses” I should have also said that we haven’t built one, haven’t tested one and have no plans to. Mumsnet doesn’t use AI to write any posts or replies on the site. We understand how important it is to keep Mumsnet human.

On name changes, which are a crucial part of feeling comfortable posting here: I want to reassure you that AI can't link people’s posts across usernames.

When we say posts used in research are anonymised, we mean we don't share any personally identifiable information and we don’t join anonymised posts together using an account identifier. Partners get trends and themes, and occasionally an anonymised quote from a public post, not anything that identifies the individual wrote it.

Our moderation and tech teams can see posts alongside registration details, because that’s needed to moderate and run the site. As our privacy policy explains, we keep that access to a minimum and review it regularly. If you deregister, your posts stay up, but your email address and other registration details are deleted, so they’re no longer connected to your posts. The one exception is the IP address you registered with, which we keep for moderation purposes.

I hope that’s helpful. The point of our insights work is to show what parents in general think and care about, not to profile anyone. We take individual privacy very seriously, and we know anonymity is fundamental to Mumsnet. People speak frankly here because they can, and it’s the people, not the technology, that make Mumsnet what it is.

Thank you for all the questions. It’s been useful to set this out clearly, and we’ll look at adding something to our guidelines and policies that spells it out for everyone.

Justine

I was joking

Just quoting this because it makes me feel a bit icky. It seems such a clrar case if "if it ain't broke..." Yes people mess up their thread titles sometimes but trying to avert this by sharing our 'thinking aloud while we develop an OP' with a bastard tech giant seems to be benefiting Soneone Else rather than MN users.

Thanks very much for those answers, @JustineMumsnet .
They are very reassuring. As AI is so rapidly developing I do feel the need to keep abreast of the changing situation much more actively than I have done in the past. All information of that sort, relating to the use of AI, will always be very welcome. It would be especially welcome if changes could be actively presented in user-friendly terms, perhaps in an MNHQ onsite blog, as well as in the privacy.policy.

Regarding name changing - worth noting that AI is entirely unnecessary to identify users across name changes.

Most of us will have had the experience of recognising an old poster in a new name simply because of a posting style or maybe specific set of facts mentioned in posts over time. Its a common way for bad actors to be identified by other posters when their struggling SEN triplets become Oxbridge sextuplets between posts or the same OP keeps appearing under different usernames. Standard analytics will industrialise that recognition factor and anyone can capture public posts and analyse them.

For the most part name changing is harmless obsfuscation but it is not anonymisation.

I made a couple of posts on this on another thread and I have been asked to repost and link to them on this thread.

A poster quoted @JustineMumsnet with her own comments (in italic, Justine in bold and italic), and I replied. My original reply is here:
https://www.mumsnet.com/talk/womens_rights/5586244-feeding-the-zoo-animals-site-stuff?page=2&reply=155243536

First, on the prompt tool to be clear we’re no longer sending any draft text to OpenAI. When we switched title suggestions off, we stopped sending anything, not just stopped showing the suggestion.

I don't see what Justine or MN has to gain by lying about this. And if they turned off the title suggestions (which is easily checkable), then they stopped sending things. I would say this is absolutely fine. And while to some it might feel like an overreaction, it is proportionate to the anger expressed about it.

MN runs on trust and confidence, and if we stop posting or using MN because they have used /are using AI, then that directly eats into their profits and commercial model. They rely on our eyeballs on their advertisments or paid subscriptions to make their money. If we stop using MN they lose money, hence them being very responsive to criticism.

When the tool was operational, all that was sent was the text of the draft and the topic, to generate a title. [my bold] No usernames, email addresses or account details were included, and OpenAI doesn’t use this data to train its models...."

I also think this is true. MN has a paid business subscription to AI and as I said above, it the AI companies started using the content generated under a business subscription to train AI, it would result in businesses stopping using AI in case their legal, financial or private information was leaked. Imagine a solicitor using AI to write template letters and private information about divorce, court cases, employment negotiations, redundancies being leaked, no-one would use those lawyers again.

As an aside I found that some lawyers submitted an update to a legal document that used the wrong version of the document - so instead of updating v2 to v3, they updated v1 instead, missing out legally agreed amendments to v2. They now have to redo all the work on to v2 for free, and get legal agreements again so that v3 includes all of v2 for free. This is another six months' work all for free, and the work they did on making adding v3 to v1 will not be paid. Then they will be sacked and their reputation over a silly little cock-up on version control will be trashed. Just because no-one checked.

So imagine that AI had done that (it definitely wasn't AI in this case) there would be huge lawsuits. And if that AI had been used to train the rest of the AI models worldwide, there would be even bigger huge lawsuits. Same as if sensitive information was used that meant there could be insider trading. So it is definitely true that a business subscription will not train other models.

Equally, sending usernames or email addresses or other account details would not be necessary (assuming the OP of the thread doesn't mention these details in the post). There is no need to send the information when it's just to generate a title.

All that would be sent would be the topic - AIBU/Feminism:Sex and Gender/Parenting/Breastfeeding etc and the text of the draft - AIBU that I insist on everyone in the house removes their shoes because they traipse in the dirt and it's unhygienic etc?

Again I see no reason for MN to lie about this. They have been trying to provide a service to support users.

I think now that some of the "invisible magic" has become visible, some users are unhappy and didn't realise exactly how the AI title generator worked. And that is absolutely justified. And to be fair to MNHQ, maybe they didn't entirely understand exactly what was involved in the secret sauce either.

The following is clear to me what happens -

1. The topic of a new post that starts a thread and the text drafted in that new post is sent to OpenAI, as it is typed.
2. OpenAI may speculate to itself, as the thread is typed, how to reply as if that text is a prompt for it to answer.
3. When the user is ready to generate a title, OpenAI suggests a title based only on 1 and 2 above. It does not generate a response.
4. No other information is sent to OpenAI
5. OpenAI does not use this information to train its AI models, including the model used by MN.

I think the following questions remain -

1. Now that users know that this is how it works, are they happy to continue using the AI thread title generator as it stands?
2. If they are not happy with this way of working, would they be happy if the information wasn't sent as they were drafting, but only on the click of a button to generate a title?
3. If they are not happy with 1 or 2 above, should MN stop using the AI title generator altogether?
4. MN uses AI for other purposes such as automatically checking images and hiding some posts that trigger a spam/inappropriate content filter until a human can check them - should MN continue to use AI to do this?
5. MN probably uses AI for other business purposes that don't directly affect users, such as financial reporting, making summary reports on usage statistics etc, and should they continue to use AI for business purposes other than content moderation or title generation?

Frankly I think that point 5 is none of our business and I am sure that many businesses we use every day are using AI for these purposes and users and consumers don't get a say.

Point 4 is our business, and following the pornography-bombing that severely disrupted MN before they brought it in, I for one, think it is very useful, and if MN didn't use AI then we would either not be able to post images, or they would take ages to be approved. And content moderation has affected me a couple of times, but generally I have reported my next post to MN and it has unhidden the flagged posts very quickly.

I am not unhappy about point 1, but I wasn't aware and I can see that it bothers some users and may restrict how they use MN to start threads, and some users may leave.

I would be much happier if the title generator worked as laid out in point 2, if that is in fact possible. And that would also mean that if users didn't want to use the title generator, then none of their draft posts would be sent to OpenAI at all.

But I can also see some users might not be happy with that either. Would that be enough users who might leave or express their unhappiness that it affects MN's bottom line? Only they can say, but if it does, then they may implement point 3 and stop using the AI title generator.

Ultimately it is a business decision for MN because a drop in users and subscribers will affect their profits. But mostly right now, MN's reputation has taken a bit of a bashing and that may be more of an immediate issue. I think improving communication and being as transparent as possible is their only option at the moment, and part of that is removing the AI title generator immediately (as they have already done) while they discuss it.

I do think a lot of people are worred and suspicious of AI for many reasons and I hope this helps users think about what is happening and how they feel about MN's usage of AI.

I am happy to answer questions on how AI works in general if that will help, too.

This is my second post on AI in general:
https://www.mumsnet.com/talk/womens_rights/5586244-feeding-the-zoo-animals-site-stuff?page=2&reply=155243546

As a quick summary AI is just a sophisticated predictor of text, like the sugggestions you get when texting on your phone. It is really good at recognising patterns and really good at generating very average text based on the internet of the last 30 years.

Most of the internet content has been written by men (because they dominate and force women off the web by shaming and threatening them). This means if you ask AI about heart attack symptoms, it will give you those for men, not women. It means it will reassure you that everything is fine when you show it a nasty skin infection, even when it isn't (me last year with a severe MSSA infection, my husband this week with a nasty infected cyst). It means it will tell you that you're the greatest in a way that makes women sceptical and that men lap up. It will not offer you comfort or an apology when it gets something wrong and you can feel gaslit.

Don't rely on it for truth, check everything it gives you, particularly if it quotes medical, legal or financial information. As Google AI just told me, "AI fails when it is forced to replace the core human element. It lacks genuine reasoning, emotional empathy, and true situational context. Using it to completely automate human relationships—like writing personal messages, evaluating job candidates autonomously, or offering unverified medical or legal advice—frequently backfires"

But if you want it to give you high speed pattern recognition like detecting cancer in 1,000 mammograms, or detecting fraudulent bank transactions, or modelling scientific research it is very good.

And it excels at tedious text analysis and data transformation. It is brilliant at summarising that 40 page document you really don't want to read. It will generate repetitive boilerplate code, and hunt down spelling, punctuation and grammar issues in your own wrtiing. And it will also generate real-time, context-aware translations beyond word-for-word substitution by workign out what is intended.

It is in essence a useful but quite broken typewriter. I still swear at it every single day because it really is quite rubbish, but at the same time when I am writing a long paper on why certain rules are necessary and how they've been broken, it is rigorous at keeping me to the text rather than speculating about why a company has broken the rules - because in my field it doesn't matter why, it just matters that the rules have been broken.

I try and think of it like a useful but broken typewriter. In ten years we will have worked out what it is good at, and that's what we will mostly use it for.

And I realised that my very first post on this on that thread might also be useful as it gives my background and my initial thoughts on the AI issue. (read this first then the other two!)
https://www.mumsnet.com/talk/womens_rights/5586244-feeding-the-zoo-animals-site-stuff?page=2&reply=155231515

I am a former software engineer and a heavy user of Claude.ai and previously ChatGPT, owned by OpenAI, and I pay for a Pro subscription for Claude and used to pay for ChatGPT (both about £20 a month). I am married to a software consultant for IBM whose job is to use AI, both IBM's specific AI, and other models, and who advises many other developers in IBM on using AI for IBM's clients.

My understanding is that when users opt out of training AI models, which I do and which I recommend all users do, then the prompts and responses put into AI are not scraped by the owners. If they did, this would be a serious betrayal of user trust and also GDPR, breaking a lot of laws in the EU.

I also understand that businesses who use AI (buy AI business licences) and install it for their systems, like Copilot etc, are automatically opted out of scraping to train other models and cannot opt in. Again this would be a serious breakdown of business trust, particularly where businesses have sensitive legal, financial and market information. I have seen nothing that tells me that the AI companies are lying when they say this.

More importantly, were the AI companies found to scrape company data to train their models, the money would dry up very quickly because no company would use AI ever again.

Therefore I do not think that anything that MN users have typed into their responses has been scraped by AI to go back to OpenAI.

Having said all of this, I do find it very worrying, that AI "watches" as users type. Presumably this only applies to posts that create new threads, but it is concerning that this is how it is set up. However, in thinking about how the AI generated thread title works, I am unsure how else it might generate a title, particularly if the title is automatically generated before posting. I have only generated two threads since this came in, so I can't speak for how it works and only have other people's word for it.

I do think though, that MN could operate an AI-generated title by having a button to click, which says, "Generate a thread title". It would mean no text is sent to OpenAI to generate a thread until the button is clicked. Obviously users should still have an option to write their own title if they don't like what is generated.

If it were an option that users had to undo in Settings, then many people would never see it and would be concerned.

As for the text that was generated by AI that was published on the Site Stuff thread, this is what AI does. If it is reading as a post is typed then it may assume that the poster is writing a prompt that will seek a response to what is being said,. The AI mdoel may not know until the poster selects post, or the option to generate a title (unsure of the mechanism).

AI will generate instructions to itself to say "do not reply like this", "consider this aspect of the post". This behaviour is not controllable by MN. I shout at Claude so often that it will say while "thinking" - how to generate a calm response to a frustrated prompt, so this type of language is normal for AI to use.

It may also be considering general instructions given by MN so that a generated title does not victim-blame.

I do think it needs to be fixed and I do think that users are right to ask questions and understand what is going on. I also agree that MN really need to get better at saying, we are looking into this and while we know it is difficult to have to wait, we will take the time necessary to find out what is happening rather than give you the wrong information.

But on the point that the AI title generation tool is scraping our posts and sending them to OpenAI, I am confident that isn't happening.

But AI from all countries and makers is probably still scraping every public forum it can get into and it's difficult to know what to do about that. Though, given how male-coded AI models are, maybe scraping 20 years of women talking to each other on the internet might make it less infuriating.

There's still no explanation re why a detailed prompt to generate a response to the original post appeared in the box (rather than the suggested title, or a prompt to generate a suggested title).

MN really need to get better at saying, we are looking into this and while we know it is difficult to have to wait, we will take the time necessary to find out what is happening rather than give you the wrong information.

Absolutely. The rapid-lie technique we've seen on this thread means we can't and won't trust them even if they do eventually come back with the real answer. Some people leave as soon as they realise the lie and don't come back to see the truth, and those who don't leave are pushed a little closer to it each time it happens.

Same as their habitual introduction of insufficiently tested features that get people's backs up. Of which this is clearly one.

As for the AI title generation, a 'push button if you want one' system would be better than the auto version. But no AI titles would be better still. It's possible as suggested upthread that they may make for better SEO, but they make for a worse user experience. I've seen many threads where the OP comes back after a page or so to say that AI has misrepresented something, added a mistake, or emphasised the wrong aspect and therefore the answers are not answering the right question or are focusing on an introduced discrepancy and therefore doubting the whole thread.

I'm still unconvinced.

One of the questions is, did users have an understanding that unsubmitted posts were being sent to AI?

If I'm writing something and there's a submit/post button, then there's the understanding that the content isn't sent until it's submitted

Changing the privacy policy now, without actively informing users is poor. Something like sending your posts to AI needs to be proactively agreed, just updating terms and conditions isn't enough. No one reads the ts and cs once signed up

Has there been independent auditing of the AI being used? There's not been a single explanation about that schema that's believable

I agree. The “we’ll look at adding something to the guidelines and make it clearer, thanks for the discussion it’s been really helpful” stuff is really irritating. At the very least, there should be a prominent post, probably in AIBU cos, you know, traffic. Any updates/guidance needs to be seen by as many as possible, not just slipped into the guidelines because, let’s face it, I doubt many people have read through them. MN need to make things as obvious as possible with regards to any communication to all users of this site.

I can’t filter to read just posts from MN staff, because that’s only available on a Premium account. So, I may have missed something. Is there any information on whether it’s being actively looked into (ie through contact with OpenAI) as to how the AIBU schema blah blah occurred? Because, aside from concerns on this thread re this site, I’d be hoping that anything so obviously and extensively rogue as this would be important to flag to them - what other sites could be experiencing something similar?

And I still think there needs to be a better explanation re the first few MNHQ responses to the whole thread in the first place. The first one was bad, because it showed that NO attention was paid to the text, it’s obvious that it was deemed to be an annoying moan about AI-generated thread titles, and answered with a condescending “it looks complicated, but it’s really simple code” tone.

@randomchap - apologies, I quoted your post to agree with it, the rest of my post is not a list of queries pointed at you!

I started drafting this post at around 2145 and you can see that this is now some time later.

This is because I realised that what I had said previously hadn't fully explained what was going on. I spoke to my husband about this and we did a deep dive into the situation. It took us around two hours to understand what might have happened and we looked closely at the text labelled "AIBU Response Schema v1.0", that was put in the title field and the OP's actual post. I am not going to reproduce the post as it is quite personal to the OP, but the information did help us understand what happened.

This is what we surmise as the most likely answer to the question.

This does not mean we are right, but we think it does answer a lot of the questions asked.

We think the output is a bug in the moderation tool not the title generation tool.

1. The moderation tool populated the schema from the post.
2. It was ready to send the schema to AI as a prompt to scan and categorise the post then generate a decision to keep the post visible or hide or delete the post.
3. That information was confused or tangled with the other tool to generate the thread title.
4. In the confusion, the title was populated with the information in the schema and did not generate a title.

Justine from MNHQ stated on 5 October 2026 at 1152, "On this occasion, instead of a title, it sent back a long set of notes describing how someone might reply to the post. That's what appeared in the title box. We can't see inside OpenAI's model, so we can't say exactly why it did this, but it looks like an error on their side. It's still our tool, though, and what shows up on Mumsnet is our responsibility."

This is not an OpenAI issue because the schema clearly contained information from the original post on MN's side and not from OpenAI, and it's not AI behaviour, it's a badly coded instruction that attached the wrong prompt to the AI title generation tool and the responsibility lies with MN.

Later on the 6 October 2026 at 1629 "I understand why this has unsettled people, particularly the discovery that draft text was being sent to OpenAI while you were writing"

I have already said that sending draft text as it is being written undermines users' confidence in MN's use of AI.

We also agree that this will use a lot of AI tokens as it is constantly sending information to the AI assistant. It would be cheaper in money and energy to only generate a title or moderate content on posting and not while drafting.

But most importantly, we think that both the moderation tool and the title generation tool sent requests to AI as the OP was typing and the two requests happened to run close together in a way that doesn’t normally happen. This resulted in a mix up that put the moderation tool’s output (the schema) in the title field.

@MyrtleLion - interesting. Being flippant, I’d say the biggest takeaway is to do away with the AI thread title suggestions forever. It could be our/Mumsnet’s small contribution to not using energy resources for completely pointless stuff!

@MyrtleLion as this is anonymous you'll be aware that we have no way of knowing whether your credentials are as you say they are or whether you are acting on behalf of Mumsnet to try to dampen the situation. Equally, I won't detail my own credentials.

But I don't think your explanation holds. This is part of the text that appeared, and it is not a moderation prompt - it is a prompt to generate a response:

"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."