We have reached the following conclusions during negotiations between community-selected strike representatives and representatives of Stack Exchange, Inc. This aims to address most of the concerns detailed in the strike letter, the initial strike announcement, and the conditions outlined in the update post.
Negotiations took place in a private Discord server, with Yaakov Ellis and ArtOfCode serving as server admins, to ensure a neutral balance of power in a neutral location, avoiding a space entirely controlled by one party. Negotiations took place between three negotiators on each side, at Stack Exchange, Inc.'s request. Striking community members elected Cody Gray, Makyen, and Mithical as representatives to negotiate; Cesar M, Philippe, and JNat served as representatives from Stack Exchange, Inc.
Agreeing to some basic ground rules, all representatives were invited to the Discord server, and then discussed the issues that led to the moderation strike and how to resolve them in a way satisfactory to all parties. After several weeks of discussion and negotiating, we have reached the following conclusions:
Replacing the "self-admission" policy and allowing moderation of AI-generated posts
Stack Exchange has agreed to allow the removal of content based on a single strong indicator of GPT usage, or on several weaker indicators. Effective immediately, the interim standards that we have agreed on will serve as policy until a more permanent policy is established with the help of the broader community, including determining further heuristics and what category they will fall into. Moderation should follow the standard escalation for suspensions, beginning with a warning and then escalating suspensions if infractions continue.
There are currently five established strong heuristics and five established weak heuristics.
- Strong heuristics include almost definite signs that a post was AI-generated that do not occur in human-written posts.
- Weak heuristics include details that are often a sign of GPT generation but are not as definite.
The negotiators established a preliminary list of heuristics as a way of making sure that Stack Exchange, Inc. was willing to engage in good faith on this issue and not summarily dismiss any proposed heuristics. However, three people in private cannot themselves discuss and debate every possible indicator of GPT usage; further development on this policy will happen with a broader group of people.
Further developments to the policy
The above interim policy will serve until more permanent standards are determined. We've set up a Stack Overflow for Teams instance to work out establishing further heuristics and gathering data that can be used as "indicators", as well as further fleshing out what each heuristic involves. The Teams instance is open to Stack Exchange network moderators, users active in the AI Domination private chatroom, and members of Charcoal. If you're a member of one of those groups, check the relevant chatrooms and private places for information on how to gain access.
Development of further additions and changes to heuristics and indicators has begun and should produce some results shortly. However, this is a long-term process. As the technology around generative AI continues to develop, our "tells" and identification methods will evolve with it.
Heuristics and indicators will be proposed and discussed by mods and users, with data being provided and arguments made. As of right now, a Community Manager will be required to sign off on adding a heuristic to the official lists; this may change as the process surrounding this develops and matures.
Currently, heuristics are going through several stages, including proposal, definition, and various levels of accuracy testing, in concert with Stack Exchange, Inc.'s data and research teams. The heuristics will go through these stages in the Heuristics Working Group Team, and then, once they're approved as a strong heuristic, weak heuristic, or non-weighted indicator (i.e. heuristics that are useful for spotting something suspicious but are not accurate enough to delete based on them), they will be added to a list maintained on the Stack Moderators Team. This process will be streamlined as much as possible.
Releasing the private "self-admission policy"
Stack Exchange, Inc. has released the private, prior version of the policy restricting moderation on AI-generated content. This policy is no longer in effect. It can be found at (Historical) Policy on the use of GPT Detectors.
Testing moderators' capabilities to detect AI-generated content
Stack Exchange, Inc.'s research team is developing a process to test Stack Exchange network moderators' capabilities to detect AI-generated content. The methodology for this is supposed to be rigorous enough that the results of this experiment can later be released publicly, such as in a research paper (provided that the individuals involved consent). Several Stack Exchange, Inc. staff members have participated in this experiment already and moderators will be invited to participate shortly.
The data dumps, the Stack Exchange Data Explorer, and the Stack Exchange API
A long-term commitment to freely maintaining and providing these services
Stack Exchange, Inc. has posted a public commitment to freely maintaining and providing these services for the benefit of the community. That commitment can be found at The company's commitment to the data dumps, the API, and SEDE.
Changes to the Moderator Agreement
A mandatory review period for binding policies
Any proposed binding network-wide policy must undergo a mandatory feedback and review period, of a minimum of seven business days (as defined by standard Stack Exchange, Inc. corporate business days), where Stack Exchange network moderators can provide feedback and have the policy revised before it goes live. This is the bare minimum; it is strongly encouraged for Stack Exchange, Inc. to have a longer feedback gathering period and involve a broader group than just network moderators. Afterwards, any binding network-wide policy must be publicly posted here on Meta Stack Exchange and tagged with the mod-agreement-policy tag.
Stack Exchange, Inc. staff cannot mandate moderation actions be taken without a public policy supporting those actions
Staff of Stack Exchange, Inc., including Community Managers, cannot mandate that actions be taken that would affect non-moderator users without a public policy stating that these actions must be taken. This disallows further private binding restrictions, such as being mandated to "ignore/decline" flags raised about GPT content (as was the case in the formerly private "self-admission policy" linked above).
However, this does allow for private instructions in two cases:
Non-public enforcement guidelines in cases where making those public would be detrimental to moderation
In cases where having the exact criteria for taking action public would facilitate abuse, they may be kept private, as long as there is a public policy that outlines what type of criteria exist. It's a fine line to walk, and there are very very few cases where anything has to be private; anything kept private will be very carefully considered and discussed to establish if it actually needs to be private.
All of the policies around, for instance, voting fraud are public. It's very clear that this behavior is not allowed anywhere on the platform. However, the exact criteria for when the fraud detection scripts will automatically take action are private. The criteria for when diamond moderators take action or escalate to the Community Team are also private. If these were public, this would allow for much easier evasion of these scripts and tools.
In the same vein, the basic policies for when moderators will take action on AI-generated content are public (a single strong heuristic or several weaker heuristics). However, the precise list of heuristics is available only to moderators and select users, and as further heuristics are developed and changes made, those heuristics will also be kept private to prevent users using these lists to evade detection. This is also why the AI Domination chatroom is private and invites sent out manually and approved on an individual basis.
Private guidance during the Moderator Conduct Review process
Private guidance may be issued during the Moderator Conduct Review process that is applicable only to that individual moderator. This is still subject to the restriction that this cannot mandate actions that affect non-moderator users, and the moderator in question may still choose to reveal the guidance they were given, either to their fellow moderators or publicly. This is akin to a moderator sending a mod message to a user; it's kept private by default, but they do have the option of choosing to make things public. This does not apply to content from the Teachers' Lounge or the Moderator Team, where the contents are still confidential and may not be copied out.
Updating the Press Policy surrounding statements on moderation
In light of the comments sent to the press by Stack Exchange Inc. in response to the strike, the Moderator Agreement will be updated to reflect further restrictions on such comments.
In the future, any comments made to the press that involve moderation must undergo the approval of at least one member of the Community Management Team, as that did not happen in this case. Any comments made in such a scenario will be kept as general as possible and no comments will be made on specific cases. This is in addition to the current restrictions of not discussing individual moderators or their actions without their express written permission.
Determining if Stack Exchange, Inc. violated the Moderator Agreement
As of V2 of the Moderator Agreement, there is no provision for determining if Stack Exchange, Inc. violated the Moderator Agreement. There exists a process for determining if a moderator violated the Agreement, as well as a process for moderators to determine if another moderator should be removed, but no process for determining if the company violated the agreement.
We have negotiated an outline for such a process.
In the event that Stack Exchange, Inc. is found to have violated the Moderator Agreement, Stack Exchange, Inc. must:
- Retract and nullify any actions taken or comments made during the course of the violation; and
- Issue a public apology to Meta Stack Exchange detailing the violation made.
In order to establish that such a violation took place, a consensus must be reached by a certain number of Stack Exchange network moderators. The numbers that we are working with at the moment were initially proposed by a moderator, not by the company, and haven't been finalized. The numbers will be discussed on the Mod Team before being finalized. The placeholder numbers are:
- a minimum of 20% of Stack Exchange network moderators must vote on if a violation was committed
- a minimum of 90% of moderators voting must vote that a violation was committed
At the time of writing, there are 541 network moderators. These percentages would require there to be a minimum of 108 moderators voting, with at least 97 of those agreeing that a violation took place. These percentages may be adjusted during the feedback and review period for changes to the Moderator Agreement.
Standard feedback and review period for changes to the Moderator Agreement itself
Any changes to the Moderator Agreement mentioned here are still subject to the standard feedback and review period for changes to the Moderator Agreement. A minimum of 30 days for feedback and review is required before changes go live. With any change to the Agreement, moderators are required to re-accept the agreement, with a minimum of 60 days before the deadline to agree to the updated agreement.
Because the Code of Conduct is considered to be part of the Moderator Agreement, changes to the Code of Conduct also require re-acceptance of the Moderator Agreement. We (the negotiators) asked Stack Exchange, Inc. to delay requiring moderators to re-accept the agreement after the recent changes to the Code of Conduct so that we could make these changes to the Agreement and not have the hassle of re-accepting twice in a short period of time.
Changes to internal Stack Exchange, Inc. processes
Communicating with moderators in unclear cases
We discovered a severe disconnect in expectations around Stack Exchange staff reviewing actions taken by volunteer moderators. Stack Exchange internal guidance stated that moderators should almost never be contacted to discuss moderation actions they took while performing that role. This meant that if a user sent in a complaint through the Contact form about a moderator action, Stack Exchange internal policy dictated that staff should not contact the moderator who took that action to ask them to justify why that action was taken.
This was drastically different compared to the expectations that moderators have; diamond moderators direct users to contact staff if the user believes that a moderator is acting in error, operating under the belief that Stack Exchange, Inc. will do their best to objectively investigate the situation and ask the moderator why they took particular actions. This apparently was not happening, with staff by default upholding actions taken by moderators and not contacting the mods in question.
This internal guidance has been updated to reflect that moderators are not opposed to explaining to staff why they took certain actions, and in fact expect staff to review their actions and bring up cases where the mod may have acted in error. Moderators are human and do make mistakes. The expectation is that staff will investigate when contacted about possible moderator mistakes, by looking into the matter themselves, and then, if it's still unclear to staff what happened, discussing with the moderator who took action. If the end result is that the moderator was in error, that should be caught, and complaints about moderator error should not be automatically dismissed.
Not making irreversible changes to the platform
An announcement was made to the Moderator Team, stating that a major change was being made to a foundational system of the platform and that this change would be applied network-wide and irrevocably. (This planned change, which is still private, was referred to as "the second shoe" during strike coordination.) That change was temporarily paused shortly prior to the strike while the impact was reevaluated, mitigations investigated, and potential adjustments investigated. It has since been clarified that this will not be an irreversible change and that that was a mistaken assumption on the part of the person posting the announcement.
The internal Stack Exchange, Inc. process for making announcements has been updated to reflect that system changes are almost never irrevocable or irreversible and should not be announced as such.
Gathering community feedback before committing to a major change to the platform
Committing to transparency
Stack Exchange, Inc. staff will be as transparent as possible about product development and policy, regularly sharing updates and proposed changes. Releases will be communicated in a timely manner. Whenever possible, staff will provide insights behind key product and policy decisions to the community.
Taking into account community feedback
Staff will work with the community when making decisions about product development and policy, taking into consideration community feedback and suggestions while also considering other data points and research. Feedback on releases that substantively impact the user experience will be sought at the earliest possible opportunity, ideally during an initial ideation and requirements gathering phase of work, but absolutely far enough before the release of a “beta” product such that the feedback given can still influence the direction of the tool or change, and be incorporated into the design.
Improving day-to-day communication between staff and moderators
Aside from the disconnect mentioned above where staff wouldn't reach out to mods in unclear moderation cases and appeals, we've discussed the issue of staff feeling attacked by moderators in mod spaces such as the Teachers' Lounge ("TL"). We've established that staff will make better use of the tools available for dealing with toxic messages, such as raising flags or contacting an elected Teachers' Lounge Room Moderator; the TL Room Mods will also attempt to be more proactive in calling out toxic messages directed at staff. This does not mean that the TL will lose its purpose (or, one of its purposes) as a place for mods to privately vent frustrations, but that people should be aware of the effect that their messages may have.
With staff feeling more comfortable interacting with mods and community members in general, Stack Exchange, Inc. will increase the level of feedback gathering and working with the community when it comes to software and policy changes.
Changes to the Stack Exchange platform
Notifying users about the policies around AI-generated content
Data gathered by Stack Overflow moderators indicates that a significant percentage of users who were suspended or warned for posting AI-generated content were not aware of the prohibition on posting such content. Stack Exchange, Inc. will implement a warning that will show to users pasting in content to the answer box, notifying the user about the prohibition on AI-generated content and hopefully decreasing the number of users who are unaware of the policy.
The text for the notice to users will be customizable on a per-site basis and able to be enabled/disabled on a per-site basis. Initial work on this has begun. Stack Overflow moderators have very recently been given a mockup of a first concept and their feedback has been requested.
Tooling for mod communication and determining if the company violated the Mod Agreement
Stack Exchange, Inc. commits to developing and implementing tools for moderators to communicate and vote on if the company violated the Moderator Agreement. These tools will be developed in conjunction with moderators and alongside the discussion around the numbers required for the process. Stack Exchange, Inc. will not in any way hinder or obstruct discussion on the matter of Stack Exchange, Inc. violating the Moderator Agreement taking place on the Stack Exchange platform.
I would like to thank everyone involved in negotiations for working to come to these conclusions and dedicating a substantial amount of time and effort towards achieving them, including the representatives from Stack Exchange, Inc., the community representatives, the server administrators, and the Stack Exchange Network community at large.