The curious case of @hisvault_eth

In which an apparent AI chatbot turns out to be a troll of unknown origin operating a hijacked X account

Earlier this week, tales of the the alleged discovery of a malfunctioning ChatGPT-driven bot account went viral on X, the social media platform previously known as Twitter. The inspiration for these allegations was a series of replies from the (presently suspended) blue check account @hisvault_eth containing apparent error messages from OpenAI’s ChatGPT chatbot. There are several problems with the format and content of these replies that make it vanishingly unlikely that they were actually produced by ChatGPT, however, and Wayback Machine archives of some of the account’s other recent content suggests that the allegedly automated replies were in fact a very human form of trolling. There are, however, multiple suspicious aspects to the @hisvault_eth account, including evidence that the account was hijacked.

quote marks, how do they work?

Debunking the alleged ChatGPT error messages posted by @hisvault_eth is relatively straightforward. First, there’s the matter of the syntax: among other things, the quote marks are not parseable, as the inner quote marks in the output section are not escaped with backslashes as required in JSON (and many other common data formats and programming languages). Additionally, ChatGPT’s API does not take a country of origin as a parameter at present, and OpenAI does not offer any model named “ChatGPT 4-o”. Note that none of the above proves that the account operator has never used ChatGPT or a similar tool to help write their posts; it simply means that the claims that these particular posts are ChatGPT errors are false.

the frequent use of a slur suggests that this is not in fact ChatGPT output

Many of @hisvault_eth’s posts prior to the recent suspension were archived by Wayback Machine in JSON format, and these archives shed further light on the nature of the account. One additional data point that casts doubt on the notion that this account is a ChatGPT-driven bot is the frequent use of a (slightly redacted) ableist slur beginning with “R” in the account’s replies. 759 of the 2839 replies archived by the Wayback Machine contain this slur. OpenAI’s language models generally avoid using slurs of any kind, so it is highly unlikely that these particular replies were generated using ChatGPT.

nothing to see here, just a small business account that magically transformed into an NFT account

As it turns out, the @hisvault_eth account wasn’t always named @hisvault_eth. Wayback Machine archives indicate that this account, which has the permanent ID 46448791 and was created back in June 2009, was previously named @ABikersGarage and appears to have been the official Twitter account of a US-based motorcycle repair business. The exact time of the name change is difficult to pin down, as the last Wayback Machine archive under the name @ABikersGarage is from June 2022 and the first @hisvault_eth archive is from May 2024. Alongside the name change, the alleged location of the account operator was changed to Estonia, and the account was redecorated as an NFT-themed account. This set of changes suggests that the account was hijacked, sold, or otherwise changed owners.

the @hisvault_eth account was very prolific in its last three days of existence

The archived @hisvault_eth posts are almost all either replies or reposts, with only one original post in the entire dataset, which consists of the word “test”. The reposts began in mid-May 2024 and are almost all NFT-related. These reposts happened in short barrages in which the account reposted dozens of posts over the course of an hour or two and then went silent for several days (at least, as far as can be discerned from Wayback Machine archives).

I would like to know more about the Tesla with the same cost as a pizza

On June 16th, 2024, the @hisvault_eth account began posting large numbers of replies. Interestingly, some of the first replies are insults directed at accounts that tagged the account a decade earlier, when it was named @ABikersGarage. Most of these accounts appear to have been customers of the motorcycle repair business that formerly operated the account. Many of the account’s other replies prior to the first fake ChatGPT error message are trollish as well, which suggests that the entire “malfunctioning ChatGPT bot” incident may simply have been another form of trolling. (It’s worth reiterating that this does not necessarily rule out some use of automation, especially given the account’s high post volume in its final hours.)

insulting the former customers of the business whose social media account you hijacked is an interesting choice

What are we to make of all of this? While the initial assessment of this account as a ChatGPT-driven bot gone astray appears to have been a misdiagnosis, the suspicion that the account was inauthentic was well warranted. The most likely explanation is that someone hacked or purchased the dormant @ABikersGarage account, and repurposed it for NFT promotion, shitposting, and whatever other activities the operator had yet to get around to at the time the account was suspended. It’s worth noting that, despite the ban, this stunt was an effective means of driving discourse, as the resulting discussion continued for several days and even inspired trending topics on a couple occasions.