I just tried the AI function to see if it could make my post more gender inclusive.
Alas, it did not detect the 10 to 0 “he” ratio at all.
1 Like
Which function did you try? What was the prompt you used?
1 Like
All I know is this “AI” button appeared so I pressed it.
Then it said generating.
Then I accepted it.
Then it also asked me “try again?”
So then I clicked that too.
But then I got an error message.
1 Like
Yea I don’t think AI grammar revision will work. Best to just write it gender neutral the first time.
5 Likes
This is totally off topic and for all of you just a curiosity, but if I write something in finnish that doesn’t use genders and I ask answer in English, it is always such funny to see how much AI struggles.
Very weak point is the term AI is overpositive per se and those are actually very limited. But… I’ve never tried what happeds if I’ll prompt gender neutrality 
2 Likes
I had a thought, tried an experiment, and it worked. What I did was put my request in as the first line of the draft:
Please rewrite the following using gender-neutral language. Thanks.
The AI took the hint, and did exactly that.
5 Likes
For clarity, this is what I got back when I asked the AI to proofread @jidanni’s drafted text, with my extra one-line request:
We know that when a user replies by email to a thread that was in the meantime subsequently closed, they will receive a Sorry email response. All well and good.
But let’s examine the cases where the thread they are replying to via email has not been exactly closed, but the topic has been redirected or whatever various other combinations moderators might do.
OK, so there they are, replying blindly to an hours-old email in their Gmail etc., without knowing that indeed, if they would just have a look at the website, it was now part of a whole other thread or whatever.
OK, off goes their email. But when it hits the Discourse server, what happens?
Well, it doesn’t generate a Sorry email reply.
But it also doesn’t get posted anywhere.
Maybe a few weeks later the user, back on the website, swears they did mention something, but cannot find it now.
4 Likes
As an amusing further illustration that a one line instruction in the body of the draft can push the ‘proofread’ function into a chosen direction, I asked it to “rewrite the following as a rambling anecdote in the style of Moby Dick” - it obeyed and removed my request, leaving this:
Let us consider, if you will, the plight of a lone user, adrift in the vast digital sea of emails and threads. Picture this mariner, replying with earnest to a missive received in the bygone hours, not unlike a captain responding to a ghostly ship’s hail in the fog. This user, unbeknownst to them, is dispatching their thoughts into the ether of a thread that has, in the meantime, been closed or perhaps shifted in its course like a fickle wind.
Now, imagine this scenario: the thread to which our digital wanderer sends their reply has not been sealed shut with the finality of a closed door, but rather, it has been subtly redirected, its course altered by the unseen hands of moderators, much like a current redirected by the ocean’s hidden swells.
There they are, our steadfast correspondent, dispatching their reply from the familiar shores of their Gmail or suchlike, wholly unaware that the digital landscape has shifted, that the thread of their focus has been woven into another narrative entirely.
And what of the fate of this message? It ventures forth into the vastness, reaching the server of Discourse. Yet, it finds no port; it does not return with a sorrowful ‘Sorry’ nor does it anchor itself in the visible threads of discussion. It simply vanishes, like a ship swallowed by the horizon.
Weeks later, our mariner, navigating once more the expanses of the website, is confounded. They are certain they had charted this territory, left their mark upon the conversation. Yet, there is no trace to be found, their words lost to the digital depths, leaving behind naught but a whisper of frustration in the endless sea of discourse.
4 Likes
I’d like to point, yet again, that for custom prompts users must use the custom prompt input.
Using the content itself to inject instructions may work today, as the current crop of LLMs is very susceptible to those injection attacks, but it’s not a sustainable approach we recommend as it will invariably stop working in the future and lead to broken workflows.
7 Likes
It isn’t clear to me that this custom prompt is something which is available to the user, as opposed to the site admin. I can’t seem to find it, anyhow: here on Meta as a user I am offered just four choices, one of which is “Proofread”.
I think I might in the same boat as @jidanni - I am approaching this as a user, not expert in the featureset that’s offered. Hopefully that position is a useful one, for constructive feedback.
4 Likes
Site admins determine which groups have access to AI Helper presets, and which groups have access to custom prompts in AI Helper.
Here on Meta, for example, presets are TL2+ and custom prompts are TL3+. Those settings are gated to groups to help site admins control their costs.
6 Likes