AI Dungeon is an open source text adventure game released in late 2019. It is unique in that it uses the GPT-3 text generation model to generate open-ended and unlimited plots. This system was designed by Open AI, an American company specializing in AI. As part of a recent update, AI Dungeon developer Latitude implemented a new tool that prevents the game from generating sexual content involving minors, but it has sparked a debate among users.

A video game where the game master is an artificial intelligence

The principle of AI Dungeon is rather simple: at the beginning of the game, the player can choose several “game modes” corresponding to particular events, contexts or atmospheres. You can play with zombies or aliens, or carry out dark investigations, experience an apocalypse, a romance, etc.

After this step, the game proposes a situation and your adventure begins: you just have to write your character’s action in the game and the artificial intelligence will generate a new action induced by your input, and this, as many times as it is necessary until your fictitious death or until you have solved the investigation.

How is this possible? Thanks to a deep learning algorithm that has been trained on every possible action, dialogue and situation the player can think of. The problem is that some users have managed to make the model generate offensive situations such as sexual content involving minors, which scandalizes the majority of the game community. Under pressure from OpenAI, the developer had to change its AI system.

An AI update to prevent the generation of inappropriate content

Last Tuesday, Latitude published an article mentioning this update in order to best explain the changes made to the AI model and whose first sentences are explicit:

“Yesterday, we released a test system to prevent the generation of certain sexual content that violates our policies, particularly content that may involve depictions or descriptions of minors (for which we have zero tolerance), on the AI Dungeon platform.”

The developers clarified that other sexual content between consenting adults and situations involving violence or vulgarity could still be generated and would not be considered inappropriate for the system.

Of course, the brand has tried to anticipate some questions, especially about data protection: does the Latitude team have the ability to see all the stories generated by users? The brand says the model only detects inappropriate content automatically, without the Latitude team having access to other content:

“We built an automated system that detects inappropriate content. Latitude reviews the content flagged by the model in order to improve the model, enforce our policies, and comply with the law.”

However, a user of GitHub, a web-based software development hosting and management service, says just the opposite. According to him, Latitude can have access to all the generated text and plots. Worse still, its database storing nearly 50 million adventures and 800,000 scenarios would be, still according to this Internet user, vulnerable to attacks.

Part of the game community deplores the capabilities of this algorithm, threatening archived stories or censoring content that would not be inappropriate. To allow a better efficiency of the algorithm, Latitude wishes to appeal to the participation of its community which must immediately report any content that could be, according to them, inappropriate.

Translated from Le jeu vidéo AI Dungeon met à jour son intelligence artificielle pour lutter contre les contenus choquants