Volunteer moderator for @aimod.social

13 Aug 2025

I am one of the volunteer moderators of @aimod.social. Feel free to engage if you have questions or wish to discuss labelling rationales. DMs are open, and will be regularly monitored. We're friendly!

22 Aug

Quite note here that for the witch hunt in question, the PFP of the person that we labeled carries the SynthID watermark from Google's Gemini. All images generated by Gemini now carry an invisible watermark that survives light edition. You can learn more here: deepmind.google/models/synth...

This post is unavailable.

We have decided to label this post as "Suspected AI Imagery". Here's how we came to this conclusion, as an example of our process. It takes time and effort, and we try to make our due diligence as much as possible. First, there is no known source for this video. No news coverage. 1/6

onesillyfag@onesillyfag.bsky.social

Asteroid strike on the moon.

The compounding of clues between no reputable source for this video and the inaccuracies in craters makes us believe this is fake, hence the label. Long story short: it takes us lots of time and effort to make our determinations, and we are always happy to be proven wrong with new data. 5/6

I hope this sort of explanation helps people understand our processes. Of course, we can't describe every single image and determination like this, it'd take too much work on an already pretty exhausting process. But I figured since we just recently relaunched, it may be beneficial. Stay safe! 6/6

onesillyfag@onesillyfag.bsky.social

Asteroid strike on the moon.

But more importantly, doing an overlay of this "Moon" over the actual one shows dramatic errors. Significant craters like Reinhold's are missing (circled red), while smaller craters remain clear. And there's some new ones, like the massive one next to Sinus Iridum (circled orange) 3/6

In addition, there are many weird artifacts in the video. It's a bit difficult to describe, but the "wavy rivers" effect in these shots is this. By itself, this could be caused by upscaling. But in the context of the other defects, this is yet another factor in the determination. 4/6

onesillyfag@onesillyfag.bsky.social

Asteroid strike on the moon.

euronews.comAI-generated label becomes mandatory in the EUCompanies creating or using AI-generated content have to label it clearly for users to know, as part of an EU guideline implemented on Sunday. #EuropeNews

An impact of this magnitude would probably have made at least some news, but we've uncovered no reputable source for this video. This is one big red flag when it comes to this sort of event. A singular video coming out of nowhere with no source, date, nor information is suspect by nature. 2/6

Volunteer moderator for @aimod.social reposted

AI Imagery Labeler

@aimod.social

2 Aug

At the end of the day what people keep asking for is the ability to know when something is or is not AI so they can make their own informed choices about how they interact with it. Hopefully tools like SynthID become widespread so it doesn't fall on the user to make their own determinations.

Volunteer moderator for @aimod.social reposted

31 July

Sampling of some of these tools, in no particular order, on the same image. Using these is virtually the same as tossing a coin.

Volunteer moderator for @aimod.social reposted

AI Imagery Labeler

@aimod.social

30 July

We've updated so that you can now also make appeals directly through our query page! query.aimod.social Of course users can also still appeal through Bluesky by clicking on the tag in question and then "appeal."

Volunteer moderator for @aimod.social reposted

AI Imagery Labeler

@aimod.social

27 July

ANNOUNCEMENTS AND UPDATES! Thank you for your patience, everyone. Yes, we are back.

Volunteer moderator for @aimod.social reposted

26 July

Hey everyone! 👋 I want to share an update on @aimod.social:

Volunteer moderator for @aimod.social reposted

26 Dec 2025

Hey @aimod.social users 👋 I’ve seen the messages about the labeler and wanted to say the team is ok and we’re so grateful for how much it meant to you. We’re sorting some things out and can’t share details yet, but the plan is still to bring it back. Thanks so much for the support—more updates soon.

Volunteer moderator for @aimod.social reposted

5 Dec 2025

Hey all, the team and I at @aimod.social are temporarily shutting down the labeler. I can't go into detail, but we discovered issues that could put me or the team at risk. We'll bring it back once we’re confident it's safe. Thank you so much for your understanding and for believing in us.

5 Dec 2025

4 Dec 2025

??? What label ?

This post is unavailable.

16 Nov 2025

Quote post since I can't reply. See bsky.app/profile/anti... It's on Bluesky, not us. The label was removed from our system, but they still haven't propagated the change.

This post is unavailable.

A bit more on this. We now believe that: - It depends on which PDS the viewing account is located. - It's transient, as in, it eventually syncs up with time. Additionally: It ALSO happens when we remove a label. It keeps showing for hours or days before eventually disappearing for everyone.

Volunteer moderator for @aimod.social@antifae.bsky.social

We are seeing a new problem with the labeler. We have no idea if it affects other labelers, but we're wondering if it has anything to do with the temporary suspension that happened on the account. Basically, labels stop showing for SOME people. Example for this post here: bsky.app/profile/did:...

I can appreciate the need for the throttling of the synchronization of labels from an infrastructure's perspective, to avoid overwhelming the network, if that's what's happening here. But then it makes the third party moderation system fairly moot for urgent and important matters.

If such a lag exists when propagating labels across PDS, then third party moderation labels will only be relegated to being gimmicks that can't really be trusted. While this may be a good technical decision, I don't think it's a good user experience one. Please consider investigating, @bsky.app.

Volunteer moderator for @aimod.social@antifae.bsky.social

This used to be instant, but something changed with Bluesky. Label changes now propagate very slowly, and without consistency between PDS. This honestly isn't ideal, because in a rapidly evolving world, having a proper third party moderation system that can react instantly is crucial.

Many people are now depending on our labels to identify what is ai-generated from what's real. Not being instant, and not having a consistent appview between users diminishes the credibility of the moderation system, in case of breaking news or similar time-sensitive event.

Volunteer moderator for @aimod.social@antifae.bsky.social

This is a test post for the labeler. It should show the label "AI-informative" on it. It has been generated using Google's Gemini, using the prompt "Generate an image which is a simple horizontal gradient from blue to red."

We don't know how widespread the issue is, but this is definitely a very problematic bug. We don't know if it affects other labelers. We don't know how to solve it ourselves. We'll need @bsky.app's help on this one. We're going to try filing a bug for it, but we need a bit more data.

If people can spread the word, and report here whether or not they can see the label, and where, it'd be useful. Thank you for your help!

Volunteer moderator for @aimod.social@antifae.bsky.social

We discussed amongst the team. Everyone is having the same issue. All of our "old" main accounts are showing the label properly. All of our newer alt accounts aren't. We tried various things like unsubscribing and resubscribing, changing settings, clearing caches, to no avail.

It seems to happen on desktop browsers, on mobile phones, everywhere, for the same exact set of accounts, so it seems related to the account, not to the application or the machine used, but we've also had reports internally that it may also change from one machine to another. This is really weird.

Volunteer moderator for @aimod.social@antifae.bsky.social

On my main account, this is what I see. The label is properly applied, and everything's fine.

On this account, and on my NSFW account, this is what I see. The label is just completely missing.

We got a bit of a scare today, as our main @aimod.social account for labeling and moderating got suspended. It was scary because of the energy that has been poured into this project, and has seemingly disappeared all of a sudden. Luckily, it was only temporary. Let's talk about the effort a bit. 🧵

Davis Bickford@davis.social

@aimod.social is restored! I’m so sorry for the inconvenience! If Bluesky gives us more information on why the account was suspended, I’ll be sure to share what we learn and communicate how we’ll avoid it happening again. Thanks again for your patience! 🩵

The work is definitely grueling. We see the same content often, especially during world events. We see truly awful content. I have seen CSAM myself a few times. Breaks are needed, we can't do this continuously. Burnout is real. But we keep on, because we believe what we do is important.

So I hope this thread explains a bit why I was anxious at the idea of this labeler going down, and helps put in perspective what we do here. It is truly a work of love and passion, not only by the team itself, but by the rest of the community. Thank you all, we'll keep on labeling! -- Fin

Davis Bickford@davis.social

This is manual work. The only automation that's in place are: - If a post or account gets deleted before a human gets to it, the report is closed. - If a post is reported but has no image, the report is closed. - If a reported post has gen-ai hashtags, it gets labeled as ai-imagery. That's it.

The team has been pouring an insane amount of love and energy into this project. They are also some of the kindest and generous humans I have ever been interacting with, and I am truly glad to have been accepted amongst them. I believe I have formed some new friendship this way.

Davis Bickford@davis.social

The team has applied 405499 labels since the existence of the labeler. Some of these labels aren't relevant anymore, as we changed some and merged some, but the current relevant count for them is the following. Note this includes labels that have been removed by the team, following an appeal.

Counting the labels that are still applied shows the following. This means the team has checked, agreed with appeals a total of 2288 times, and has been removing labels accordingly. The other screenshot shows the difference, aka the number of times a label has been removed.

Davis Bickford@davis.social

The total number of reports as of today totals 140145. Meanwhile, the team has clicked buttons on the moderation UI to process a report 122871 times. Remember posts can be reported multiple times, hence the discrepancy. But the small group of people here has been manually doing this. By hand.

I myself have gone through ~6000 reports. One (now retired) member of the team has gone through 37000 of them by themselves. This is the scale of the work the team has been doing here, volunteering our time, in an effort to make Bluesky a bit safer, a bit cleaner for people to experience.

Davis Bickford@davis.social

The alt-text on the previous post analyzes the graph in more details, but let's face it: without user reports, @aimod.social wouldn't exist, plain and simple. We are unable to monitor the whole of Bluesky on our own, and we need you to help us by reporting things you believe are generated by AI.

I'll add however that if you write us some words in your report explaining why you're doing a report, it will help us greatly. We read each and every report individually. It's sometimes a bit difficult to guess why a report has been done, so don't hesitate to talk to us through the report!

Davis Bickford@davis.social

Now, incidentally, I was recently collecting numbers off our database, in order to help the team get the scale of what we're doing. I'm going to share some of these numbers today. What's important to realize is this is a group effort. Not just from the team, but from all of you, reporting content.

This graph shows the amount of reports per day the team is receiving. The dates are visible on the X axis, and the Y axis indicates the number of reports that happened that day. Note these are the number of individual reports. The same post may be reported multiple times by different people.

Davis Bickford@davis.social

First and foremost, I want to thank the @moderation.bsky.app team for processing our appeal so quickly. I am certain this was done in error. The moderators here are only humans, and can do mistakes, the same way we sometimes do mistakes while labeling content. If anything, the appeal process works.

Volunteer moderator for @aimod.social reposted

Davis Bickford@davis.social

Hey all 👋 we saw that @aimod.social has been suspended. We don’t know why at this point, but we’ve appealed the suspension and requested more information. I’ll update when I know more. Thank you for your patience! 🩵

Volunteer moderator for @aimod.social reposted