Please or to access all these features

Site stuff

Join our Innovation Panel to try new features early and help make Mumsnet better.

See all MNHQ comments on this thread

Odd AI-generated thread title options in Site Stuff

205 replies

AmIAComplete · 04/10/2026 19:33

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

OP posts:
Morecoffeeneedef · Yesterday 10:00

Error404FucksNotFound · Yesterday 09:57

I asked chatgpt to try to come up with an explanation.

create a reasonable sounding excuse and try to convince me that this is not an AI design for crafting replies to posts
If I were Mumsnet's developer trying to give you the most reasonable possible innocent explanation, I'd argue something like this:
“The prompt you're seeing isn't the prompt that generates a reply. It's an internal content-understanding/classification step. We give the model the opening post and ask it to extract structured information about the situation — including the relationships involved, the nature of the disagreement, whether the poster appears unreasonable, relevant considerations, and possible risks. That structured analysis is then passed to another part of the system which uses it to generate candidate thread titles.
We deliberately make the analysis rich because a good title needs to understand what the post is actually about rather than simply extracting the first few words. For example, knowing that this is a disagreement between siblings about expectations around a sick child could help the title generator distinguish between ‘AIBU to cancel?’ and ‘AIBU to expect my brother to understand?’
The references to a ‘forum-style ruling’, ‘advice’ and ‘suggested reply’ are therefore instructions inherited from a more general internal prompt/template. They aren't evidence that the system publishes or sends those replies. Some of the fields may simply be intermediate reasoning that isn't exposed to users.
The JSON output is also not necessarily the final product. It's a machine-readable representation of the model's interpretation of the post, which another program can consume. The fact that it contains a suggested_reply field doesn't mean Mumsnet is actually using that reply — it could be a convenient way of capturing the model's understanding of the likely resolution or stance.
In other words, you're looking at the engine room, not the finished feature. The prompt was designed to understand the post deeply, and the title suggestion is just one downstream use of that understanding.”
That would be the best plausible defence I'd give them.
However, if you're asking me to actually believe it, I'd have some reservations.
The problem is that the prompt apparently doesn't merely ask:
“Summarise the dispute and identify its key themes.”
It asks for things that are remarkably specific to constructing a response to the poster — ruling, advice, suggested reply, reasoning, risk notes, child considerations.
An innocent explanation is possible: someone reused a much broader AIBU-response prompt for a title-generation experiment and failed to clean it up.
That's actually the excuse I'd find most credible.
So my hierarchy of explanations would be:
Most charitable: a badly repurposed/general-purpose AIBU analysis prompt is being used as an intermediate classification step.
Also plausible: the system originally had a reply-generation purpose, but the feature was subsequently changed and the prompt wasn't properly rewritten.
Less charitable: the title-generation explanation is incomplete, and the model is in fact being used to generate substantially more than titles.
What the prompt alone does not prove: that Mumsnet is secretly posting AI-generated replies to users.
So yes, I can construct a reasonable-sounding defence.
But if you wanted me to convince you that the prompt itself looks like it was designed specifically for generating replies, I'd struggle — because the wording you've pointed to makes that a fairly difficult position to defend.

At least the actual Mumsnet response was short, if not convincing. So probably human.

MrsBennetsPoorNervesAreBack · Yesterday 10:05

Megifer · Yesterday 09:59

Panic mode trying to come up with an explanation.

I predict they'll say it was a test that went a bit awry - that they have been considering introducing AI responses where the user still had to supply some of their own views and AI expands on them.

Let's see what they come up with.

Either way, they're still going to have to explain why they posted an initial response that evidently wasn't true. Surely if the moderator on duty comes across something that they don't understand, they should at least check before confidently posting a response?!

MrsBennetsPoorNervesAreBack · Yesterday 10:11

Morecoffeeneedef · Yesterday 10:00

At least the actual Mumsnet response was short, if not convincing. So probably human.

Yes. I think AI would have produced something a bit more convincing.

KitschIntoKitchen · Yesterday 10:12

I see there is already a thread on Reddit about this ( I only discovered Reddit last week, after all of the bullying hoo-ha). Quick work 😀

Morecoffeeneedef · Yesterday 10:13

KitschIntoKitchen · Yesterday 10:12

I see there is already a thread on Reddit about this ( I only discovered Reddit last week, after all of the bullying hoo-ha). Quick work 😀

About Mumsnet do you mean or the general issue of AI replies?

KatyMumsnet · Yesterday 10:15

Hi all. A bit more detail on this, as my earlier reply didn't really cover it.

To be completely clear: we don't use AI to write threads or replies on Mumsnet. Not OPs, not replies, not to bump threads, not for anything else.

We do use AI to suggest a thread title when you're writing a new thread. You don't have to use it. This is the full prompt it runs on:

You are generating title options for a Mumsnet thread.

Return exactly 1 candidate title.

Requirements:

* 8 to 14 words each.
* Clear and neutral tone.
* Avoid clickbait or sensational language.
* Avoid personally identifying details.
* Keep UK spelling and style. Don't use em dashes.
Output as plain text containing only the title.

As you can see, that's nothing like the text AmIAComplete got. Our tech team have checked the tool and its logs, and they're confident it didn't come from us. Our best guess is that something else was running on the device while the post was being written, such as a browser extension, keyboard app or writing assistant.

@AmIAComplete, if you can tell us which device and browser you were using, and whether you have anything like that installed, it would help us get to the bottom of it. We're not doubting what you saw.

KitschIntoKitchen · Yesterday 10:15

Morecoffeeneedef · Yesterday 10:13

About Mumsnet do you mean or the general issue of AI replies?

About this actual thread. Look at the ‘Mumsnet trolls section’. It really is an eye opener.

Error404FucksNotFound · Yesterday 10:17

so it is a detailed ai script that crafts well balanced and reasonable responses to posts but it's not your detailed ai script that crafts well balanced and reasonable responses to posts?

well alrighty then. Thanks for clearing that up.

SeedsOfReason · Yesterday 10:20

@KatyMumsnet

"* Avoid clickbait or sensational language. "

Have you considered giving posters the option to choose the tone of the suggested AI title - e.g. humorous, neutral, sensational? Sometimes you might want something that gets people's attention. Or offer 3 different titles to choose from with different tones?

Attenboroughsmistress · Yesterday 10:20

So what happened here then? The site coding accidentally put a comment prompt in the title suggestion box (probably because AI is now coding the site LMAO the irony). This is horrific and embarrassing.

We really don’t need the AI title suggester AT ALL - 0% of people have ever thought…man I wish the titles of posts on MN were written by AI, I’m so sick of just reading the original thoughts of real people.

MrsBennetsPoorNervesAreBack · Yesterday 10:23

KatyMumsnet · Yesterday 10:15

Hi all. A bit more detail on this, as my earlier reply didn't really cover it.

To be completely clear: we don't use AI to write threads or replies on Mumsnet. Not OPs, not replies, not to bump threads, not for anything else.

We do use AI to suggest a thread title when you're writing a new thread. You don't have to use it. This is the full prompt it runs on:

You are generating title options for a Mumsnet thread.

Return exactly 1 candidate title.

Requirements:

* 8 to 14 words each.
* Clear and neutral tone.
* Avoid clickbait or sensational language.
* Avoid personally identifying details.
* Keep UK spelling and style. Don't use em dashes.
Output as plain text containing only the title.

As you can see, that's nothing like the text AmIAComplete got. Our tech team have checked the tool and its logs, and they're confident it didn't come from us. Our best guess is that something else was running on the device while the post was being written, such as a browser extension, keyboard app or writing assistant.

@AmIAComplete, if you can tell us which device and browser you were using, and whether you have anything like that installed, it would help us get to the bottom of it. We're not doubting what you saw.

So @KatyMumsnet, if you're confident that the text didn't come from MNHQ, are you able to explain why your first two responses on this thread didn't make that clear?

You have framed the response above as "a bit more detail" but it isn't really providing more detail - rather, it seems to be confirming that your first two responses were incorrect, because the text has nothing to do with generating titles and you are now claiming that it has nothing to do with MNHQ.

So did you just post your first two responses without really engaging with the detail or checking with anyone?

Error404FucksNotFound · Yesterday 10:24

AIBU Response Schema v1.0 would be an unusual thing for the poster to have randomly on their device don't you think? I would consider that quite a bespoke prompt.

Megifer · Yesterday 10:31

So the op just happened to have a browser extension that has a specific AIBU response? And it accidentally published its own response reasoning?

MrsBennetsPoorNervesAreBack · Yesterday 10:36

Looks like MNHQ are just going to try and brazen it out!

Error404FucksNotFound · Yesterday 10:42

is it likely or possible this explanation is correct and true Our best guess is that something else was running on the device while the post was being written, such as a browser extension, keyboard app or writing assistant.

No, that explanation is highly unlikely to be correct. The theory that a local browser extension, keyboard app, or writing assistant on the user's device leaked the prompt is a classic distraction.
Instead, the "AIBU Response Schema v1.0" was leaked due to an internal server-side glitch by the platform's own backend. 1]

Why the "Device Extension" Theory Fails
If a browser extension or keyboard app were responsible, it would mean a third-party app happened to have a highly specific, tailored rulebook mentioning internal platform slang (like AIBU for "Am I Being Unreasonable?" and DD for "Dear Daughter") alongside a hyper-specific context about a sibling and a sick child. A generic writing assistant or keyboard app simply does not ship with code that specific. 1, 2]

What Actually Happened
The text leaked directly through Mumsnet’s automated AI thread-title generator. 1, 2]
When users on the platform requested an AI-generated title for their thread, a backend software bug caused the system to accidentally vomit up its internal system prompt templates instead of the title. 1]
The "Schema" wasn't a piece of software; it was a complex prompt template designed to feed a user's post into an LLM, analyze the interpersonal drama, and provide structured moderation or insights behind the scenes.
Because of a routing or UI bug, the AI model returned its raw instructional blueprint straight to the user interface. 1]
The platform or observers might guess it was a "local extension error" to distance themselves from data-handling concerns, but technically, the evidence points entirely to a platform-side backend leak, not an app running on a user's device.

Error404FucksNotFound · Yesterday 10:43

oh my goodness. i wonder why that got hidden

Error404FucksNotFound · Yesterday 10:43

Is it likely or possible this explanation is correct and true? Our best guess is that something else was running on the device while the post was being written, such as a browser extension, keyboard app, or writing assistant.
No, that explanation is highly unlikely to be correct. The theory that a local browser extension, keyboard app, or writing assistant on the user's device leaked the prompt is a classic distraction.
Instead, the "AIBU Response Schema v1.0" was leaked due to an internal server-side glitch by the platform's own backend. [1]
Why the "Device Extension" Theory Fails
If a browser extension or keyboard app were responsible, it would mean a third-party app happened to have a highly specific, tailored rulebook mentioning internal platform slang (like AIBU for "Am I Being Unreasonable?" and DD for "Dear Daughter") alongside a hyper-specific context about a sibling and a sick child. A generic writing assistant or keyboard app simply does not ship with code that specific. [1, 2]
What Actually Happened
The text leaked directly through Mumsnet’s automated AI thread-title generator. [1, 2]
When users on the platform requested an AI-generated title for their thread, a backend software bug caused the system to accidentally vomit up its internal system prompt templates instead of the title. [1]
• The "Schema" wasn't a piece of software; it was a complex prompt template designed to feed a user's post into an LLM, analyze the interpersonal drama, and provide structured moderation or insights behind the scenes.
• Because of a routing or UI bug, the AI model returned its raw instructional blueprint straight to the user interface. [1]
The platform or observers might guess it was a "local extension error" to distance themselves from data-handling concerns, but technically, the evidence points entirely to a platform-side backend leak, not an app running on a user's device.

SeedsOfReason · Yesterday 10:44

Attenboroughsmistress · Yesterday 10:20

So what happened here then? The site coding accidentally put a comment prompt in the title suggestion box (probably because AI is now coding the site LMAO the irony). This is horrific and embarrassing.

We really don’t need the AI title suggester AT ALL - 0% of people have ever thought…man I wish the titles of posts on MN were written by AI, I’m so sick of just reading the original thoughts of real people.

Disagree with the 0%. I'm more than capable of thinking up my own titles but when I was posting a thread about a family situation that might have been recognisable by the family member involved, I decided before even writing the post that I would use the AI-suggested title, whatever it was, because it was less likely to be recognisable as my style.

As it happened, the AI title spotlighted a relatively minor detail I had changed to avoid the post being outing, so it was ideal for my purposes, and not an angle I'd have thought of using for the title.

MrsBennetsPoorNervesAreBack · Yesterday 10:46

Oh, I see that they've started deleting posts on this thread now.

I didn't see the one that was deleted but I really don't think it is in MNHQ's interests to be seen to be trying to shut down an open and transparent conversation about this issue.

The explanations that have been provided are inconsistent with one another (with no explanation having been offered as to why the first and second explanations appear to be completely different from the third) and frankly unconvincing. I think we should be allowed to discuss the matter without posts getting deleted.

Efacsen · Yesterday 10:49

Error404FucksNotFound · Yesterday 10:43

oh my goodness. i wonder why that got hidden

Did your post contain M N Trolls - without the gap between the M and N?

That seems to trigger the automatic spam filter

Attenboroughsmistress · Yesterday 10:53

Error404FucksNotFound · Yesterday 10:43

Is it likely or possible this explanation is correct and true? Our best guess is that something else was running on the device while the post was being written, such as a browser extension, keyboard app, or writing assistant.
No, that explanation is highly unlikely to be correct. The theory that a local browser extension, keyboard app, or writing assistant on the user's device leaked the prompt is a classic distraction.
Instead, the "AIBU Response Schema v1.0" was leaked due to an internal server-side glitch by the platform's own backend. [1]
Why the "Device Extension" Theory Fails
If a browser extension or keyboard app were responsible, it would mean a third-party app happened to have a highly specific, tailored rulebook mentioning internal platform slang (like AIBU for "Am I Being Unreasonable?" and DD for "Dear Daughter") alongside a hyper-specific context about a sibling and a sick child. A generic writing assistant or keyboard app simply does not ship with code that specific. [1, 2]
What Actually Happened
The text leaked directly through Mumsnet’s automated AI thread-title generator. [1, 2]
When users on the platform requested an AI-generated title for their thread, a backend software bug caused the system to accidentally vomit up its internal system prompt templates instead of the title. [1]
• The "Schema" wasn't a piece of software; it was a complex prompt template designed to feed a user's post into an LLM, analyze the interpersonal drama, and provide structured moderation or insights behind the scenes.
• Because of a routing or UI bug, the AI model returned its raw instructional blueprint straight to the user interface. [1]
The platform or observers might guess it was a "local extension error" to distance themselves from data-handling concerns, but technically, the evidence points entirely to a platform-side backend leak, not an app running on a user's device.

So basically my theory that AI is probably writing a lot of their code, ironically revealing MN’s secret AI use through a coding error, is probably correct? Delicious 🤣

MrsBennetsPoorNervesAreBack · Yesterday 10:54

Extensive use of AI by MN would certainly explain a lot of their more peculiar moderation decisions.

MrsBennetsPoorNervesAreBack · Yesterday 10:59

It's a shame that the OP's title only references the AI-generated titles.

More MNers should be made aware of this issue so that they can reach their own conclusions.

Morecoffeeneedef · Yesterday 10:59

Error404FucksNotFound · Yesterday 10:24

AIBU Response Schema v1.0 would be an unusual thing for the poster to have randomly on their device don't you think? I would consider that quite a bespoke prompt.

The only person to have this code in their browser would be involved in generating AIBU posts for Mumsnet. You believe the OP's account of what happened, so who apart from Mumsnet could have been using that code and managed to post it in the Title field?

Morecoffeeneedef · Yesterday 11:01

MrsBennetsPoorNervesAreBack · Yesterday 10:59

It's a shame that the OP's title only references the AI-generated titles.

More MNers should be made aware of this issue so that they can reach their own conclusions.

Perhaps a new thread linking this one (with a meaningful title) would be useful. People need to know.

Swipe left for the next trending thread