April 26th of 2021, Latitude, the company behind AI Dungeon, implemented a filter for private adventures. They
officially announced the filter the next day. Latitude claimed the filter was
only supposed to filter pedophilic content (although there was a pretty obvious attempt to filter bestiality as well, as the filter had issues with animals). At the time the filter was added, it was just laughably bad. Mentioning a child, or an animal, or a number below 18, or a sexual term, regardless of the context, triggered the filter when it was first added. The devs also said they'd be monitoring private stories that triggered the filter, which was seen as a privacy violation given how faulty the filter was. There was also nothing preventing the AI from giving outputs that the filter would disallow.
Latitude made some changes to the filter early on. The first major changes were the removal of the censorship of animals, and the change to a two-column filter, where you'd have to use one of the censored words relating to minors,
and one of the censored words relating to sex, within the last 400 or so characters of the story to trigger the filter. The term "pedoph" (and any words containing the term) was an outlier, as it would trigger the filter on it's own, without any other censored words having to be present.
The filter still wasn't good after that, though, since it was, of course, incapable of understanding context. The way they went about filtering also meant that it was incredibly easy for people actually using the AI for pedophilic content to avoid the filter if they knew how. It also meant that words containing censored terms could trigger the filter. As some examples: "Infantry" (contained "infant"), "breastplate" (contained "breast"), "analysis" (contained "anal"), "cocky" and "cockpit" (both contained "cock"), "boyfriend" and "girlfriend" (contained "boy" and "girl", respectively), etc. Latitude, within a couple of weeks of the filter's existence, also added an automated suspension system that would sometimes automatically suspend players for triggering the filter for up to a week.
It was also
found that AI Dungeon's finetuning data contained some pedophilic content, which made the AI more likely to output pedophilic content, itself, and which a lot of users saw as being hypocritical.
Additionally, there was a
data breach that occured about 2 weeks prior to the filter controversy, and that the AI Dungeon community only found out about during the controversy. It apparently allowed for all private content, except for private content from the first 11 days of AI Dungeon 2's existence, to be accessed. Latitude failed to disclose the breach for months, which people were understandably upset about. It wasn't until 5 months after the breach occurred that
Latitude finally disclosed the breach.
There was also the issue of the moderation of private stories being outsourced to strangers on the internet via a website called "Taskup". As a result of this outsourcing, over 100 private stories ended up being leaked on 4chan. To be fair, it
later came to light that the outsourcing was the doing of OpenAI (the company that used to provide Latitude with their AI models), rather than Latitude's doing, but that was unknown for months, and regardless of who was responsible for the outsourcing, people were still upset about such a breach of privacy.
Aside from that, during the controversy, several features of AI Dungeon were either made worse, or removed entirely. The "Explore" feature, which allowed users to share prompts and stories, and to communicate with one another, was also removed just 2 weeks before the controversy started. That definitely wasn't the cause of the controversy, but it added fuel to the fire regardless. It also didn't help that Latitude went radio silent for most of the controversy.
Eventually, Latitude ditched OpenAI. This was because of a combination of reasons. Firstly, OpenAI began enforcing heavy censorship for their models that disallowed violent content and general sexual content, among other things. Secondly, as a result of some users no longer using the OpenAI AI models as a result of said censorship, Latitude ended up losing the volume-based discount that they previously had for OpenAI's models, meaning that the AI was now significantly more expensive for Latitude to use than before. Thirdly, OpenAI is generally a shitty company and was, as I said before, responsible for over 100 private stories being leaked on 4chan. So, now AI Dungeon uses AI models from AI21. Unfortunately, though, the current AI models are pretty much unanimously considered to be worse than the old ones from OpenAI.
For clarification, the filter still exists, but is nowhere near as bad as it was during the controversy. At this point in time, there is no longer a suspension system, and with
Latitude's shift to the "Walls Approach", the filter now only applies to the AI. Latitude also claimed they'd no longer be manually reviewing stuff that triggers the filter. The "Explore" page is also back, now called "Search". Although it's... kinda worse than before since the Posts section and the "Top" sorting option were both removed, for some reason.
There was some controversy a few months ago regarding the addition of ads. Prior to ads being implemented, AI Dungeon had been using an energy system for over a year, and the energy had a high enough limit and recharge rate that most users were unhindered by it. When ads were first implemented, you'd get a pop-up ad every 10 or so actions. However, a lot of users ended up disliking this system, as they found the ads to be too intrusive, long, and/or frequent.
Latitude justified the implementation of ads by saying that ads were a highly requested feature. That claim
was true, but the people who requested for ads to be implemented mainly either wanted ads to restore energy, wanted the option to watch ads for some sort of benefit, or wanted banner ads; people didn't ask for the type of ads that Latitude initially implemented. Anyways, Latitude ended up switching to the current system in which you can watch ads to get 10 actions per ad, and can stockpile actions before playing instead of having to be constantly interrupted by ads while playing. Even with the new ad system, though, there's still a fairly large portion of the community that's unhappy with the ads.
More recently, in August, there was also
another data breach.
Here is an archive of the Q&A Latitude had about the breach.
www.reddit.com
