GAN-tastic politics enthusiasts

Come for the artificially generated faces, stay for the artificially generated text and accompanying error messages

we are real politics enthusiasts, please ignore the artificially generated faces

In the final weeks of 2023, a group of accounts on X (formerly Twitter) with StyleGAN-generated faces began spamming a variety of users with replies containing the phrases “politics enthusiast” and “political enthusiast”. The interests and political stances of the self-described politics enthusiasts vary from account to account and from post to post, making it difficult to discern the overall mission of this network at the present time. Nonetheless, there are sufficient similarities in the accounts’ behavior, content, and profiles to map out the network, or at least a portion of it.

none of these people exist, but they may have replied to your posts anyway

Searching X for recent replies containing the word “enthusiast” and either “politics” or “political”, and then filtering to accounts which use GAN-generated faces as avatars yielded 575 accounts created between May and December 2023. Some but not all of the GAN-generated faces used by these accounts contain a tiny "StyleGAN 2 (Karras et al.)" watermark in the bottom right corner. Many of the accounts have handles containing unusually long strings of vowels, such as @Maairiuieinaaa, @eJooeiaAoneueer, @Tieoanoue, @MWasuuunaoto, @CuuoceuiaeMuuna, and @NieoaaaioWeaauv. None of these accounts have posted more than a handful of times, and almost all of their posts are either replies or reposts. (There also exist accounts with similar content and naming patterns that do not have GAN-generated profile images; although these may well be part of the same network, they are excluded from this analysis.)

we don’t always announce our ages, but when we do we use two hyphens

In what may be an attempt to make these accounts seem more human, many of them have brief autobiographical posts disclosing their alleged ages, sometimes accompanied by other alleged details such as gender, nation of origin, or hobbies. Ages are always expressed with two hyphens — “60-year-old”, “51-year-old”, “38-year-old” as opposed to “60 year old”, “51 years old”, or “I’m 38” — and other repetitive aspects are present as well, such as overuse of the word “enthusiast”.

each of these posts was amplified by dozens of the accounts in the network

923 of the 2467 posts produced by the accounts in this network (37.4%) are reposts, and over half of these (566/923, 61.3%) are reposts of a single account, @cenyk1230. The @cenyk1230 posts reposted by the network are all long form posts about large language models and other AI-related topics. The rest of the network’s reposts are of large accounts such as @elonmusk, @historyinmemes, and @fasc1nate.

as a tech enthusiast, I completely agree that these are fake accounts

The remainder of the network’s content is mostly replies. Although exact duplicates are rare, these replies tend to have very similar phrasing and certain sentence fragments show up over and over, particularly at the beginning of sentences. For example, numerous replies begin with "As a political enthusiast…", "As a tech enthusiast…", "I completely agree with you…", or "Interesting perspective on…". This sort of repetition is harder to detect than cases where large numbers of accounts simply post the exact same text, but it is nonetheless a form of spam.

apparently X is a “cross-chain gaming platform”

Another spammy aspect of the network’s replies: when one account from the network replies to a given post, several more generally follow suit. The posts replied to by the network vary considerably in terms of subject matter and political stance, and are from a mix of popular and obscure accounts. The spam replies sometimes veer off topic, such as the reply from @AkoniMakai88225 to the official @XDevelopers account incorrectly characterizing the X API as a “cross-chain gaming platform”.

unfortunately, @PBeltran89667 has been suspended, so we won’t get to hear tips on saving money on Amazon from a large language model

Although it’s virtually impossible to be certain whether any given brief snippet of text is generated by ChatGPT or similar technologies, there are multiple signs that the text of the replies posted by the accounts in this network is artificial in origin. The accounts sometimes reply to users with what appear to be error messages rather than coherent replies, including a few which contain the phrase “I’m sorry, I’m a language model”. They also occasionally tag placeholder handles such as @TwitterUserName or @TwitterUser instead of the accounts they’re replying to, which is a potential indicator of a coding error in reply automation.

tagging @TwitterUserName is unlikely to accomplish much

This network was originally documented in this X/Twitter thread. For those who are interested, here is a paste containing the IDs of all the accounts identified as part of the network (although some have been suspended).