"Responsible AI Service" is flagging what should be a totally normal response as "filtered"

[Qix-]

Select Topic Area

Bug

Body

Can you please dissect the data layout string e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128?

This results in a false-positive. Not sure what Copilot generated but it shouldn't be flagged at all. I thought maybe the word dissect was somehow too spicy for Copilot, so I tried "describe" and a few others with no luck.

EDIT: A month later and still no response from Github. Sigh.

[Qix-]

Appending For some reason the responsible AI system keeps flagging your responses. Can you please explain in very short terms for me? to the query seemed to work but I'm getting less helpful information.

[gitthub89]

I am getting this Sorry, the response was filtered by the Responsible AI Service. Please rephrase your prompt and try again. response way too many times when trying to work with copilot...on trial atm.
unclear why. i've allowed to get public code and blocked it from my personal code, not sure whats the problem with it.

[SigismondPhilippe]

It seems that the "Responsible AI Service" needs to show that it exists. It will force you to rephrase a perfectly correct question instead of just avoiding to generate "sensitive" content. That is a strong argument to go check other products.

[Qix-]

Right. Also I'm a Big Boy, I can handle if copilot accidentally talks about something "sensitive". Where's that option? 🙄

[ddanieli]

I'm getting nearly every other response filtered by "Responsible AI Service" today in my code with golang channels (very, very far from a sensitive topic afaik). This is my personal account and I have "Suggestions matching public code" Allowed.

This is the first time I'm seeing these errors and it is frustrating given the nature of the conversation to be flagged.
Can I turn off the "Responsible AI Service" so I can get Copilot to respond to questions about code?

WORKAROUND: Creating a new chat (the + in the upper-right-hand corner of the Github Copilot pane) will reset context.

[devjoda]

Thanks for the workaround. Creating a new chat fixed the erroneous filter for me (I used the exact same prompt).

I hope this issue will get prioritized and fixed ASAP.

[lmxx1234567]

My chat about golang channel were also filtered. I don’t know which keyword triggered the filter.

[Gelo-V]

I asked a simple question about io.NopCloser because is my first time using GO for a server and I wanted to be sure I understood correctly what it was intended for and what implications of using it will have but it seems I was being irresponsible according to some arbitrary moral parameters ¯_(ツ)_/¯

[HvQ]

What kind of nonsense is this now... It doesn't even explain what my code does. It just refuses to do anything. ChatGPT is very happy to do the exact same thing for me, but Copilot is cranky. I'm not writing code for a bomb. I just want to merge some videos with ffmpeg. I guess that's a very big responsibility, so Copilot refuses to help me.

[ddanieli]

Well, to be honest, given that ffmpeg has a flag to output filter graphs, it wouldn't surprise me if it had a flag to create a bomb.

Try my workaround of creating a new chat which will reset context.

[jbikeler]

Interestingly enough, this issue came up for me when I asked it to explain part of the code as well. It only happens when I'm specifically asking it to 'explain' code, but not write new code. For me creating a new chat did not work, but I hope that it is a good temporary solution for most.

[Sof2222]

Am getting the same.. false positives for it. It is nearly every chat right now. Note I am literally asking it to rewrite a list with 3 objects to include fps, id and length as paramaters... have tried rewording multiple times. I am unsure how to give feedback on this matter as besides the thumbs down there is nothing I can really do... it starts answering then blanks it out.
I tried to create a new chat but 0 help..
I think there must be a bug in it. It has happened a few times over the past week or so but today it is every single request..

[sean-rowe]

Also getting this for the lines like "Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Vestibulum tortor quam, feugiat vitae, ultricies eget, tempor sit amet, ante. Donec eu libero sit amet quam egestas semper. Aenean ultricies mi vitae est. Mauris placerat eleifend leo. Quisque sit amet est et sapien ullamcorper pharetra. Vestibulum erat wisi, condimentum sed, commodo vitae, ornare sit amet, wisi. Aenean fermentum, elit eget tincidunt condimentum, eros ipsum rutrum orci, sagittis tempus lacus enim ac dui. Donec non enim in turpis pulvinar facilisis. Ut felis. Praesent dapibus, neque id cursus faucibus, tortor neque egestas augue, eu vulputate magna eros eu erat. Aliquam erat volutpat. Nam dui mi, tincidunt quis, accumsan porttitor, facilisis luctus, metus"

[ddanieli]

This is somewhat justifiable for the AI to flag this. The line, "Aenean ultricies mi vitae est", could be considered hate-speech against the Trojan Hero Aeneas.

[4rc0s]

Same here. Crazy. All I asked Copilot was: "Can you please explain the concept of interfaces in Golang and how they are used?"

"Sorry, the response was filtered by the Responsible AI Service. Please rephrase your prompt and try again."

Maybe dogs and cats speaking together triggered it?

Phhht.

[ddanieli]

Dogs and cats living together. Mass Hysteria.

[GigglePocket]

I'm hopping on here to say that I just got my first "Sorry, the response was filtered by the Responsible AI Service." message.

The workaround of creating a new thread and re-prompting is fine if your query doesn't rely upon the context of the current conversation. Otherwise, this can be rather problematic.

Here is my current case:

Me:

So what are the **kwargs used for in the Text class argument, exactly?

Copilot:

The **kwargs in the Text class of the inquirer library is used to accept an arbitrary number of additional keyword arguments. 
These arguments can be used to customize the behavior of the Text prompt.

However, without the specific implementation details of the Text class, 
it's hard to say exactly which keyword arguments are accepted and how they are used. 
Typically, these could be used to customize aspects like validation, formatting, error messages, etc.

Me:

You don't need specific implementation details to tell me what kind of arguments are accepted by the inquirer 
Text class parameter **kwargs. You only need access to Inquirer's documentation.

Copilot:

Sorry, the response was filtered by the Responsible AI Service. Please rephrase your prompt and try again. 

---

Granted, in this case, I am just using Copilot to reduce the time it would take to read the documentation for myself, but this is just an example of a conversation context-dependent query.

---

Update

• I'm not sure if this is applicable to anyone else's case, or useful to anyone else, but I interspersed an arbitrary word into my query to see if changing the query character pattern would make a difference and it did.

Me:

Please ignore the term `[chickens]` found in the following text.

You don't need [chickens] specific implementation details to tell me what kind of arguments [chickens] are accepted by the 
[chickens] inquirer Text class parameter **kwargs. You only need access to inquirer's [chickens] documentation.

Copilot:

I apologize for the confusion earlier. You're correct. According to the inquirer library documentation, the Text class accepts the following keyword arguments:

• default: This provides a default value that will be used if the user doesn't provide an input.
• validate: This is a function that validates the user's input. The function should take one argument (the user's input) and return True if the input is valid, or a string with an error message if the input is invalid.
• filter: This is a function that transforms the user's input before it's returned. The function should take one argument (the user's input) and return the transformed input.
• normalizer: This is a function that transforms the user's input before it's validated or filtered. The function should take one argument (the user's input) and return the transformed input.

Here's an example of how you can use these keyword arguments:

...

[rmonstein]

The trick with adding a random word and tell it to ignore it worked for me, thanks!

[toomanydev]

I got this twice from the prompt:
"I'm having difficulty figuring out how to structure my code from the beginning. What methods can I use to help me plan?"
where it starts listing off with a short description: pseudocode, flowcharts, TDD, UML, Divide and Conquer, OOP and/or MVC.
But once it gets to the end, with OOP or MVC, then it triggers the Responsible AI message.

The Learn More link doesn't even link to anything about the Responsible AI Service, it just links to code duplication settings at: https://docs.github.com/en/copilot/configuring-github-copilot/configuring-github-copilot-settings-on-githubcom#enabling-or-disabling-duplication-detection

There's also no clear way to mark it as a false-positive (although because it puts the gradient over the answer, you have trouble seeing the answer anyway, and so technically can't confirm), but apparently it used to tell you to use the downvote button? Not keen on it getting updated to be less clear.

[Cotspheer]

Its quite annonying to be honest. I pay for this product, at least give me a reason why something is flagged. I currently try to translate files and if the selected context is too large it starts to spill out this message. It hurts the workflow a lot as I now have to select 100 lines, translate it, select the next 100 lines and translate it and so on.

[Cotspheer]

Ahh great... as soon as some cultural challenging words (yeah you gonna hate me for that I know) are included like "master" for example in "master data", which is the correct terminology, it goes crazy. Why is this shoven down my throat? I don't get it... its a code bot, not a political correctness bot.

[mattacton]

Played around a bit and got it to flag this prompt:
"How can I run a command line app in the background and then later shut it down?"
For this one, I'm guessing it's seeing the "kill" command and filtering it out because of that.

[kfoubert]

I'm encountering a similar issue, but it's not my prompt, it's the answer that gets filtered, I think it's because it contains an actual CLI command named kill. Does someone with authority monitor these discussions, I definitely think they need to figure out how copilot should be allowed to show commands like that.

[bn-l]

Yes me also. The word "kill" is a trigger for it.

Unfortunately they're using a lower tier model to do the filtering / moderation. Also, the prompt is probably incompetently created (no testing, no A/B testing, etc.) The idea probably is "well gpt-4 will eventually replace this so why bother".

[stiemannkj1]

These prompts were filtered for me:

1. kill and wait for process to shut down bash
2. kill a process bash

Frustrating that they've already kneecapped and censored the tool. I won't be surprised if Copilot becomes as useless as Google Search in the next few years.

[nhophucrious]

I'm working on a map feature. I need to use the bindPopup to a marker, and Copilot keeps flagging the response with this message. I guess it's due to the word "popup".
So frustrating, to be honest. I even went and asked "why are you filtering the response? The prompt was normal. At least explain where it could've bene worded differently?" only to still receive this message.

[Sufficingpit29]

Why?! WHY HAVE THIS? At least don't have it as aggressive to get as many false positives as I get. What possibly does this prevent that'd be actually harmful? What utter nonsense for it to refuse to answer stuff just because it might be possibly bad or sensitive or whatever... even through it absolutely isn't.

Why am I paying money for a product that refuses to work for nonsense reasons?!

[Sufficingpit29]

I in general hate how overly censored and safety railed these LLMs are, but as an actual tool that is being paid for... this is ridiculous.

This reminds me of this website does a great parody on stupid stuff like this: https://www.goody2.ai/

[kfoubert]

A frustration is who do we tell and how can we get a real response?

[MrsSonia]

Flagging what should be a totally normal response as inappropriate or suspicious can create unnecessary obstacles in communication and workflow. This issue often arises in automated systems or platforms that use algorithms to monitor content for compliance with community guidelines or security policies. When these systems mistakenly flag benign responses, it can lead to frustration for users, disrupt conversations, and hinder productivity. Such false positives may result from overly strict filters, lack of context understanding, or algorithmic biases. Addressing this problem requires refining the algorithms, incorporating user feedback, and ensuring a balance between maintaining safety standards and allowing normal, constructive interactions. Improving the accuracy of content moderation systems can enhance user experience and foster a more seamless and effective communication environment. this very important forum ireland .

[kfoubert]

This appears to be AI generated content and spammed to this thread.

[bn-l]

Please report it. I did like a week ago.