Proposed changes to Help Center articles to include mention of AI-generated content policy

Question

Feb 22 2024 Update: I have just rolled out the network-wide version of the help center article that mentions content generated by generative artificial intelligence tools must be referenced. I have overridden this article locally to replace it with the variant that mentions that content is not allowed for the following sites:

• Ask Ubuntu
• Unix & Linux
• Academia
• Stack Overflow em Português
• Mathematics
• Home Improvement
• Politics
• SciFi & Fantasy
• Stack Overflow en Español
• Software Engineering
• Super User
• Worldbuilding
• Motor Vehicle Maintenance & Repair
• Latin
• Writing
• Programming Language Design and Implementation
• DBA
• Personal Finance & Money
• Retrocomputing
• Board & Card Games
• Code Review
• Movies & TV
• Arts & Crafts
• Pets
• Chemistry
• Stack Apps
• MSE

Work left to do:

• Coordinate with international Stack Overflow sites so ensure both variants have translations, for future use (and assist with guiding the discussion for any of those where the topic hasn't come to the community's attention yet);
• Make further changes to the referencing Help Center article, as per this answer
• Changing the copy on the "must be referenced" banner from "cited" to "referenced;"
• Change the slug for the help center article from ai-policy to gen-ai-policy; change the link in the banner to point there; put appropriate redirects in place.

---

Feb 16 2024 Update: Please see the revised proposed copy for the relevant help center articles below, that takes your feedback into account ^_^

---

Feb 14 2024 Update: Thanks for all the feedback. To make sure we have a chance to go through as much of that feedback as possible, we're moving the roll out date from today (Feb 14) to early next week (Feb 20).

---

In this recent post about the two possible banner options that Stack Exchange sites can adopt to educate answerers about their policy on answers generated by artificial intelligence tools, I mentioned I’d make a separate post to work on a draft for a Help Center article. As such, I’m here to present proposed changes to an existing Help Center article, proposed copy for a new one — and to get your feedback on both of those.

Why are we doing this, again?

As a reminder, and a summary of the proposed timeline:

• As per our Code of Conduct, any content posted that is generated with artificial intelligence tools must be referenced;
• On February 20th, a new help center article will go live network-wide at /help/ai-policy, which will reference the aforementioned network-wide policy that content generated by artificial intelligence tools needs to be referenced. Sites that have — before February 20th — made the Community Management Team aware (via the usual system) of their superseding site policy on the matter that disallows this content instead of requiring it to be referenced will have a different template that instead reflects that policy.
• On the same date, this article will be changed network-wide, to explicitly mention that content generated by artificial intelligence tools are covered by our referencing guidelines, if there isn't a superseding site policy that disallows it altogether. It will have a link to the article mentioned in the bullet point above (the one that lives at /help/ai-policy);
• No site will have either banner on by default. After the roll out, sites still have the option of changing their policy on content generated by artificial intelligence tools, and the moderation team will be able to edit the Help Center article at /help/ai-policy. They will also be able to request that the Community Management Team enable the banner that reflects their stance on content generated by artificial intelligence tools (to do so, the community on that site must reach a consensus on that decision, as well as on the language that should be present in the Help Center Article that explains that site’s policy on this matter, if it differs from the templates we collaboratively agree on in this post).

The drafts

Below are the proposed changes to the existing Help Center article on referencing:

Plagiarism - posting the work of others with no indication that it is not your own - is frowned on by our community, and may result in your content being downvoted or deleted. This includes content generated by humans, as well as generative artificial intelligence tools if $SiteShortName allows the latter to be posted — please check the $SiteShortName policy on content generated by generative artificial intelligence tools.

Below are the proposed copies (both drawing from the existing article on Stack Overflow as a starting point) for the Help Center article that will live at /help/ai-policy, with the proposed title of “What is this site’s policy on content generated by generative artificial intelligence tools?". Sites will, by default, have the top version, that refers to the network-wide standard that content generated by generative artificial intelligence tools must be referenced:

Generative artificial intelligence (a.k.a. GPT, LLM, generative AI, genAI) tools can be used to generate content for $SiteName, but this content must be properly referenced as per our guidance. If your content is determined to have been written by generative artificial intelligence tools and is not properly referenced, it will likely be deleted, along with any reputation earned from it. Posting unreferenced AI-generated content may lead to a warning from moderators, or possibly a suspension for repeated infractions.

Some sites on the Stack Exchange network may have different policies on content generated by generative artificial intelligence tools, and some may disallow the posting of AI-generated content entirely. Please ensure you check each site’s local policy about this content before posting there.

What counts as “content generated by generative artificial intelligence tools”?

“Content generated by generative artificial intelligence tools” is any content crafted, in part or in whole, using a tool that writes a response automatically based on a prompt it is provided. These tools include large language models like ChatGPT and Google Gemini. Because these tools are trained to answer with language that mimics authentic speech, the responses may look and sound plausible, but the quality of generated answers can vary significantly (up to, and including, completely wrong answers).

If you are using large language model (LLM) services as described above to draft content for $SiteName, please ensure these are properly referenced.

Why do I need to disclose that I used generative artificial intelligence services to draft content?

Stack Exchange is a collaborative resource, developed and maintained by members of the community, with the goal of creating a repository of high-quality curated knowledge. As mentioned above, while content generated by generative artificial intelligence tools may look and sound plausible, the quality of that content can vary significantly. For this reason, it is extremely important that you ensure future readers are aware of the fact that these tools or services were used to generate the content you’re posting. This ensures that readers of that content are fully aware of the content source’s limitations and shortcomings, and can thus decide how relevant that content is for them.

There are a few primary issues with content generated by large language models that make it unsuitable for use on $SiteName without properly referencing it:

1. Users who ask questions on $SiteName expect to receive an answer authored and vetted by a human. This ensures that the answer is factual, relevant, and complete, up to the standards of another human. While human authors are not perfect, generative artificial intelligence tools may not take into account other important factors that add nuance to a question, often add excessive noise to their answers (e.g., explaining all details, no matter how relevant), and may fabricate false or misleading information.
2. Users who ask questions on $SiteName may have already sought answers elsewhere, including through generative artificial intelligence services. Appropriate references help the reader decide whether or not they even want to read a given post, and flags the potential hazards involved in posts that have been written while using generative artificial intelligence tools.
3. Sometimes generative artificial intelligence tools may be used to assist with editing and translating content, rather than generate it. Despite this constituting a more transformative, rather than generative action, these tools are still prone to error, and they might still introduce new information not present in the original text.
4. Generative artificial intelligence tools are not capable of referencing the sources of knowledge used up to the standards of the Stack Exchange network. Even when generative artificial intelligence tools appear to reference sources for responses, such sources may not be relevant to the original request, or may not exist at all. For $SiteName, this means the content may not honestly or fairly represent the sources of knowledge used, and that the actual original author of some of the material being used may not be getting properly credited, even if someone explicitly references the generative artificial intelligence tool as an author in their answer.

Please note that even when properly referenced, frequent posting of content generated by these tools, especially when done in a short window of time, can lead to unnecessary disruption of $SiteName, which would be in violation of the Inauthentic Usage policy of our Code of Conduct.

How do I reference content generated by generative artificial intelligence tools?

The more general guidance offered in the context of referencing material written by others applies to content generated by generative artificial intelligence tools. More specifically, there are two main things you should ensure:

• You should clearly describe what content is generated by generative artificial intelligence tools, and which isn’t. You should do this when quoting directly from the output produced by these tools, as well as when paraphrasing those contents. This ensures there is a distinction between the content generated by these tools, content you authored, and content you may be referencing from other sources.
• You should specify the specific generative artificial intelligence tool/service you used. Since different services may produce different outputs to the same prompt, and may have different limitations and shortcomings, you should ensure readers know which specific tool you used to produce the content you are referencing.

You may want to share details on the prompt you used to produce the output too, but there is no need to completely copy and paste the prompt and output. Just as when referencing any other content you didn’t author, you should generally avoid copying the complete text, and should instead use the words and ideas from this content to support your own. Here is how you might consider referencing material generated by generative artificial intelligence tools in the body of your post:

I asked [Generative AI service] about [partial prompt]. Its output was:

[Generative AI response]

[other sources, quotes, explanations, etc. necessary to complete the post]

If you have a question about how this policy might look in practice, please look through Meta $SiteName to see if there are any past or ongoing discussions about this type of content that establish community norms surrounding it. If there aren’t any such discussions that address your concerns, consider starting a discussion yourself so other members of the community can provide their perspective.

Are there alternatives to using generative artificial intelligence services to craft answers?

Many of the answers on $SiteName are created by users sharing their own expertise. In particular, when supplying answers outside their expertise, users should reference trustworthy sources. Searching for sources, synthesizing them into a good answer, and referencing them clearly are critical steps in developing a well-crafted answer.

Remember - the person who asked a question needs a correct answer. As such, answering correctly is always more important than answering quickly. Additionally, our system thrives on questions getting multiple good answers, which are more likely to help future visitors who have the same question. By following a process that creates consistently correct answers of good quality, and that are well-referenced as needed, you should do just fine here.

And sites that reach community consensus on having a policy that disallows content generated by generative artificial intelligence tools can use the following template as a starting point, so there is some consistency in the language across network sites:

Generative artificial intelligence (a.k.a. GPT, LLM, generative AI, genAI) tools may not be used to generate content for $SiteName. The content you provide must either be your own original work, or your summary of the properly referenced work of others. If your content is determined to have been written by generative artificial intelligence tools, it will likely be deleted, along with any reputation earned from it. Posting content generated by generative artificial intelligence tools may lead to a warning from moderators, or possibly a suspension for repeated infractions.

What counts as “content generated by generative artificial intelligence tools”?

“Content generated by generative artificial intelligence tools” is any content crafted, in part or in whole, using a tool that writes a response automatically based on a prompt it is provided. These tools include large language models like ChatGPT and Google Gemini. Because these tools are trained to answer with language that matches authentic text, the responses may look and sound plausible, but the quality of generated answers can vary significantly (up to, and including, completely wrong answers).

Please do not draft content for $SiteName using large language model (LLM) services as described above.

Why am I not allowed to use generative artificial intelligence services to draft my content?

Stack Exchange is a collaborative resource, developed and maintained by members of the community. There are a few primary issues with content generated by large language models that makes it unsuitable for use on $SiteName:

1. Users who ask questions on $SiteName expect to receive an answer authored and vetted by a human. This ensures that the answer is factual, relevant, and complete, up to the standards of another human. While human authors are not perfect, generative artificial intelligence tools may not take into account other important factors to a question (e.g., optimization, security, etc.), often add excessive noise to their answers (e.g., explaining all details, no matter how relevant), and may fabricate false or misleading information.
2. Users who ask questions on $SiteName may have already sought answers elsewhere. Due to the ease of using generative artificial intelligence services, if a user wanted an answer from an artificial intelligence, they may already have sought one, and so it does not make sense to provide one here.
3. Generative artificial intelligence tools are not capable of citing the sources of knowledge used up to the standards of the Stack Exchange network. Even when generative artificial intelligence tools appear to cite sources for responses, such sources may not be relevant to the original request, or may not exist at all. For $SiteName, this means the content may not honestly or fairly represent the sources of knowledge used, even if someone explicitly cites the generative artificial intelligence tool as an author in their content.

Are there alternatives to using generative artificial intelligence services?

Many of the answers on $SiteName are created by users sharing their own expertise. In particular, when supplying answers outside their expertise, users should reference trustworthy sources. Searching for sources, synthesizing them into a good answer, and referencing them clearly are critical steps in developing a well-crafted answer.

Remember - the person who asked a question needs a correct answer. As such, answering correctly is always more important than answering quickly. Additionally, our system thrives on questions getting multiple good answers, which are more likely to help future visitors who have the same question. By following a process that creates consistently correct answers of good quality, and that are well-referenced as needed, you should do just fine here.

Feedback

We’ll keep this post open for community feedback until February 19, 2024, and will be responding to feedback during that period. Once we collectively land on the copy to be used on these two Help Center articles, I’ll update the original post about the banners with next steps on rolling this out network-wide. Thanks for your help! ^_^

Answer 1

status-completed

Revision 8 of the question addresses these concerns.

---

Answers are not the only part of the content that can be used as knowledgebase and where posting AI content can do harm we are poorly equipped to deal with, besides relying on AI policy. On Stack Overflow there are also articles and discussions.

Instead of using word answers, use word posts in AI policy text. Or at least explicitly name articles and discussions there, too.

When warning users about posting content other than answers linking to AI policy that explicitly mentions answers leaves plenty of room for interpretation. It may look like the answers are the only parts where AI is not allowed.

Similar request has already been made at Stack Overflow Meta Extend AI content policy to articles and other content that is meant to be used as part of the knowledgebase

Answer 2

Here's my thoughts:

1. status-completed The proposed wording is this:

   Posting unreferenced AI-generated content repeatedly may lead to a warning from moderators, or even a suspension from the site.

   While it's reasonable to wait to issue a suspension until repeated violations have occurred and the user has been warned, I think that (ideally) users should be warned on the initial rule violation. I don't see much value in waiting until there's multiple rule violations to tell a user to stop.

   Perhaps something like this:

   Posting unreferenced AI-generated content may lead to a warning from moderators, or possibly a suspension for repeated infractions.

2. status-planned Can there be a non-English version of this policy for non-English SE sites?

3. status-completed Can (should?) there be standardized phrasing for a ban on AI content, whether cited or not? If not, individual sites have to create that themselves. While that can work, a template for that too would be nice.

4. status-completed It states this:

   Some examples of AI services commonly used to write infringing answers on $SiteName include [...]

   While this may be true on large sites, like Stack Overflow, can this say "Some examples of AI services commonly used to write infringing answers include [...]" instead? It's quite possibly smaller sites haven't seen AI answers from a variety of services.

5. status-completed The article alternates between "AI" and "artificial intelligence" a number of times. I'd probably suggest not switching constantly. Stating "AI (artificial intelligence)" and then just using "AI" might be better.

6. status-completed Google Bard is now Gemini. It may be a good idea to update the phrasing to use "Google Gemini" or perhaps "Google Gemini (formerly Google Bard)". (I'm not endorsing a specific news site here, other news sources have

Answer 3

TL;DR Please replace the word "cite" with "reference" or "attribution" throughout.

As per our Code of Conduct, any content posted that is generated with artificial intelligence tools must be cited;

This is not what the Code of Conduct says.

Citing is not the same as referencing or acknowledging

status-completed — see revision 13 for changes that address this concern

In a science community, citing means to reference a publication. The publication does not change over time, and unless there is a global catastrophe, you will be able to find the publication, even if it was written a hundred years ago, in multiple libraries or databases. In modern times, many citable publications have a document object identifier, DOI. For example, DOI:10.1038/171737a0 is Watson's and Crick's 1953 paper on the double-helical model of DNA. Not the draft, not the version updated in 1967, not a version copy-edited by AI, but the published final version as printed.

The Code of Conduct, in the inauthentic usage section, says that plagiarism is not allowed, and links to referencing standards:

If you copy (or closely rephrase/reword) content that you did not create into something you post on Meta Stack Exchange (e.g., from another site or elsewhere on Meta Stack Exchange), make sure you do all of the following:

• Provide a link to the original page or answer

• Quote only the relevant portion

• Provide the name of the original author

This ensures that the original creator gets credit for their work.

This is not how you cite published sources, this is how you reference or attribute all kinds of information, a less rigorous standard. It is a bit like saying "(Prashanth Chandrasekar, personal communication)".

"This ensures that the original creator gets credit for their work."

status-completed — see revision 13 for changes that address this concern, at least partially

If you reference text generated by AI, even in the most rigorous way (by providing a (perma-)link to a document that contains both prompt and generated material), the original creators of the work do not get credit. This is different from a scientific paper you cite because when it relies on the work of others, it will in turn cite sources that were used, giving a network of authors involved in building this knowledge. When you post AI-generated text, this chain of evidence gets disrupted, and you have assume responsibility for the posted text (and check whether it makes sense, and find authentic sources to support the claims it makes).

The example in the referencing help text

status-planned

The example quotes a text from a web site. The web site does not give the author of the text. It references two works, but the links are dead and get rerouted. It would be nice to update the example so that it follows the referencing standard.

The definition of plagiarism

status-planned

Plagiarism is not defined within the text, but instead links to the article on wikipedia. This article will change with time, so it would have been better to link to the version of the article at the time the help text was written. As is, we define plagiarism outside of StackExchange, but use the term in the code of conduct. Importantly, the outside definition may change to include text generated by AI, or stick with the more common text written by someone else (another human, someone who deserves credit).

It says "plagiarism is frowned on" but the Code of Conduct says it is not allowed on the site.

Why is this so hard

We are trying to make rules for a technology that is in its infancy. Today, ChatGPT might not be able to connect answers with sources, but in a couple of months the next iteration might. Right now, ChatGPT is horrible at math, but the GPT4/Wolfram combo is already pretty good. Also, different people have different standards. It is the job of the community to ensure that there is a single well-explained standard if that standard is used to expel community members who do not follow it.

Answer 4

status-completed

See revision 5 of the question for proposed rephrasing to address the concerns raised in this answer.

---

Please check what the policy on content generated by artificial intelligence tools on $SiteShortName is, since this guidance on referencing may apply to those, in addition to content generated by humans.

The wording of "...may apply" indicates that it's possible that a site could allow uncited AI-generated content, which is misleading; it requires attribution at a bare minimum on every site on the network, via the Code of Conduct.

I'd personally opt for language along the lines of

Plagiarism - posting the work of others with no indication that it is not your own - is frowned on by our community, and will result in your content being edited or deleted. This includes content generated by artificial intelligence tools, which require disclosure on the Stack Exchange Network. Please check the specific policy on $SiteShortName regarding AI-generated content for further information.

This makes it clearer that it's not just "encouraged" to provide attribution, but required, and violations will result in action being taken on the content. It then directs you to the specific policy on the site in question.

Answer 5

status-completed — see revision 13 for changes that address this concern

---

Because this is a late addition, I'm adding it as a second answer for visibility purposes.

In addition to plagiarism, the Inauthentic Usage policy talks about "directly causes unnecessary and unwanted disruption ... to users and/or the network". I believe that it's worth calling this out in the AI policy Help Center documentation somewhere. A pattern that we've seen on multiple sites is users flooding the site with AI generated content. Being able to quickly generate and post content that looks meaningful is even more disruptive than someone performing destructive edits or flooding with spam or nonsense because it can be a little less obvious. Even on sites that do allow AI generated content with attribution, people (especially those with enough reputation to bypass some rate limits) should take care to not post too frequently.

Answer 6

status-completed

See revision 11 of the question.

---

please check what the specific policy on content generated by artificial intelligence tools on $SiteShortName is.

This is wordy and confusing. I recommend instead:

please check the $SiteShortName policy on content generated by artificial intelligence tools.

or:

please check the $SiteShortName-specific policy on content generated by artificial intelligence tools.

Answer 7

status-completed

Factually, equating Bard to LaMDA is wrong. LaMDA was an early Google large language model. Today, Bard uses Gemini. Between LaMDA and Gemini, Google also had PaLM. Also, as of 8 February 2024, Bard no longer exists - several Google offerings, such as Assistant, Bard, and Duet AI in Google Workspace, have been rebranded as Gemini.

---

status-completed — see revision 13 for changes that address this concern

I would also recommend clarity on the use of AI tools as editors. As we see these technologies translate into tools like translation services or editors (tools like Grammarly or Microsoft Editor), we should have considerations that these tools are more likely to be allowed. With my understanding of the heuristics that are available to moderators to guide judgement calls on if a post is AI or not, I think these tools would generally be OK. They tend to be more transformative than generative in nature but aren't always free from error. However, assuming they aren't generating new content, they can be more useful in improving the accessibility of the site and its content to a broader range of people, especially on the English-only sites in the network.

---

status-completed — see revision 13 for changes that address this concern. Guidance that needs to be more specific than the one proposed is probably better decided on a per-site basis, and changes are made locally to that site's help center article.

The edge case for questions about the use of generative AI also needs to be considered. Even on sites that prohibit the use of AI to generate answers, an answer including generative AI output is likely not prohibited if that output supports the answer. On Software Engineering, for example, I can see questions about the use of generative AI technologies to support requirements engineering. An answer that includes information about prompting techniques and examples of output (properly attributed) would be allowed, assuming that the rest of the answer was written by a human and the generated portions supported the rest of the answer.

---

status-completed — see revision 13 for changes that address this concern

I'm also struggling to reconcile the points made in the answer to why disclosure is needed with the continued stance that prohibitions on AI-generated content should be site-by-site rather than network wide. The three risks presented are presented in a way that is very compelling for outright prohibition of generated content. I would expect that many, if not most, reasonable people would look at those risks and conclude that curating generated content, even with attribution, is not worth the risks of noisy, false, misleading, unwanted, or otherwise low quality answers.

Answer 8

status-completed — see revision 13 for changes that address this concern

---

Artificial intelligence (a.k.a. GPT, LLM, generative AI, genAI) tools can be used to generate content for $SiteName,

Is it usual to use the word "artificial intelligence" to refer to generative AI in English?

In Japan, artificial intelligence (AI) refers to a broader concept than generative AI, and it sometimes includes machine translation tools, search engines, or even compilers in some specific contexts. As one of the translators on Stack Overflow for Japanese, it is a bit hard to translate this sentence without an additional explanation, which I'd like to avoid for the clean translation.

So, how about changing wording a little for ease of translation? At least, GPT is not an alternative name of AI. While using the term "AI" is acceptable, could we possibly use more precise language? (If not using the word AI, we also should change the URL path /help/ai-policy, which could cause other complications....)

Answer 9

status-completed

Most of what was written seemed fairly reasonable, but this seems strange...

Due to the ease of using artificial intelligence services, if a user wanted an answer from an artificial intelligence tool, it can be assumed that they would already have sought one.

Can it? The assumption is false for me; I'd be thrilled to get any useful answers to my questions. I don't care who/what it's written by.

But also...

1. Sure, it's easy to get some answer from some genAI; it's not necessarily easy to get a useful answer to your precise question (that's literally the problem we're tackling). I also don't think genAI is anywhere near as widely used nor easily accessible as the text makes it sound. My impression is that most users are not even thinking of genAI when asking questions.

2. Many Stack Exchange questions were asked before ChatGPT, etc., was even heard of. So the assumption cannot be applicable to such questions.

3. While the OP might not "want" a genAI answer,

   • they also don't want us to downvote nor close their question too, yet we still do it;

   • OPs don't normally get to pick-and-choose who answers their questions, nor stipulate what tools answerers are allowed to use when writing an answer;

   • maybe the thousands of readers (other than the OP) simply want a useful answer to the question, regardless of how it's written;

   • surely what the OP actually wants is an answer to their question!

May I suggest something like...

Users who ask questions on $SiteName may have already sought answers elsewhere. Due to the ease of using artificial intelligence services, if a user wanted an answer from an artificial intelligence tool, it can be assumed that they would already have sought one, such as through artificial intelligence services. If they did not want an answer from an artificial intelligence tool, they can decide to ignore one that is marked as having been generated by it. Appropriate references help the reader decide whether or not they even want to read a given post, and flags the potential hazards involved in posts that have been written while using artificial intelligence tool.

---

status-completed

Oh...

(i.e. optimization, security, etc.), often add excessive noise to their answers (i.e. explaining all details, no matter how relevant)

I think this confuses i.e. and e.g. Also I was taught that after i.e. and e.g. you're meant to add a comma. It seems like the first set of examples would be more relevant for sites where genAI is simply banned.

---

status-completed — see revision 13 for changes that address this concern

Sorry, one more thing: can you add examples of how e.g. ChatGPT might be acknowledged please?