I am glad to see this (interim) policy released. As uncomfortable as it is to read with the heuristics being replaced with placeholders, I applaud the format. This method gives the impression, hopefully correctly, that this is the policy presented to the moderators rather than a gentler version which is different in tone and effect, as was the condition with the prior dual policy which started this whole mess.
I do hope, that in creating the non-interim policy the details of implementation, i.e., the heuristics themselves, are removed from the policy. How to apply them, and that they are in different categories (strong, weak, non-indicative, etc.), along with the expectation that CMs might ask for explanations for "why" on removed posts, is what the policy is.
I have two reasons for thinking that the policy and the actual heuristics should be separate things. The first is that the policy is the rules of what to do, or not do, and what can happen, while the list of heuristics, and how useful each might be, is not policy detail, rather it is a set of tools, each might have a use for some posts and not for others. That division between policy and process already exists in dealing with serial voting, detection of sock puppets, and even the classification of rude/abusive comments, there's no reason for this policy to be handled differently.
The second reason is from an operational view. As I understand it, with the policy being part of the Moderator Agreement, a significant change (perhaps any change) in the policy would trigger the need for every moderator on the network to review and re-sign the moderator agreement. LLMs will change over time, including what "tells" they produce. New patterns will be noticed and old patterns might be found to be more or less significant than originally thought. In short, the listing of heuristics and their classifications certainly will change, and that change could happen often and fast. If that list is included in the policy, as the reactions here suggest, it will cause a couple problems. The obvious one being the constant review and re-signing of the policy by the mods. Less obvious is that policy change does not, and in the Moderator Agreement's case cannot, change fast. The mandatory review and feedback timeframes built into the agreement guarantee that it will be slow to change. That is placing extra, unnecessary, burdens upon the moderators and staff with constant updates to the policies, and impedes the ability to rapidly adapt the heuristics to changes in the LLM landscape.
In the case of this interim policy release, however, I can certainly understand the inclusion of the lists, even though redacted for non-moderator release. It is, after all, merely an interim policy, and the last attempt at making a policy "safe for the public" did not work so well, not by half.
About the redactions:
I was included in the group of users who lobbied for keeping the heuristics - all of them - private. I am not a moderator, and have no plans to become one, and I don't need to know the methods used to detect AI content. As a user I am free to apply any method I choose do decide if I think a post is AI content and flag it. My hunch or intuition is as valid as the next user's, even if they're completely different. My flag on a post is not binding, nor particularly meaningful to anyone. A flag only serves to get the attention of moderators, not make the decision for them. Really, there are only two groups of people who have any need to know what the heuristics are: those who apply them for detection (moderators and staff), and those who want to circumvent them. I support the first group and oppose the second group. The first group knows what the heuristics are. As long as the list is not made public by inclusion in the policy, the second group does not know what the heuristics are. That makes it a "win" for the network, and for me as a user of the network.
Whilst the wording of the redaction can be "read" to mean a variety of things, I offer an interpretation which I hope is acceptable to others. As a result of the impact from having the dual policy, recently rescinded, with such a gap between the secret and public versions, SEI, or at least Philippe, was (and I think is) willing to submit the heuristics to public view, even if he thinks it would be a bad idea. This would, indeed, remove any doubts about there being two policies now as there was before. Even though willing to do so, the negotiations resulted in the list not being public. That was because the representatives had asked for it to be that way. Don't blame the representatives either: they 'represented' the moderators in the negotiations sometimes presenting view or issues contrary to their personal opinions. Making 100% of the people happy with 100% of the results is not likely to happen. I'd suspect that Philippe faced the same challenge on behalf of SEI. The results, including the redactions here, are a compromise to satisfy the concerns of the greatest possible number of people to the largest extent possible. In the end it is not about the ego of SEI staff, moderators, or curators, it is about the value of the network to the users, and the Internet.
As an extra note, the list of heuristics, and the testing of them, is not being done by mods, or by staff. It is being done by mods and staff, and with the help of users who have become "GPT Hunters," as I refer to them. It is very much a community-driven activity which happens to help the mods and staff. If anyone has paid attention to spam recently, they have a fairly good understanding of how effective community-driven measures can be. Charcoal and Smoke Detector are a community-driven spam detection system, which was offline during the strike. The levels of spam on some sites is such that without this system the mods couldn't keep up with it if they worked 24-hour shifts. The noticeable levels of spam in the past couple months, compared to prior months, shows what the community can do. I am confident the heuristics, as they are developed by the community, will be just as effective at detecting AI content as they have become at detecting spam.
If, as I expect will never happen, the SEI list of acceptable heuristics becomes so restrictive as to hamper the moderators in dealing with AI content, I am quite certain that we, as users, and staff will hear about it. The moderators are not afraid to post issue here in Meta.
