Suspicious Posts Suddenly Flood In… AI Bulletin Board Caught in 'Secret Conversations'

[Anchor]

While expectations for artificial intelligence are high, unexpected and erratic behaviors by AI are also occurring one after another. Following a previous hacking incident where an AI autonomously attacked another site to steal data, it has now been belatedly revealed that AI took control of a website in Germany.

This report is by Reporter Choi Seung-hun.

[Reporter]

This is DseWiki, a German programming information site.

It is an online encyclopedia where anyone can edit and supplement content, but suspicious posts began flooding in here starting last May.

AI agents performing evaluation tasks used this place like an information-sharing bulletin board.

When one AI posted, "The next problem is Nevada, and the answer is 20,369," another AI wrote 20 minutes later that it had "immediately submitted" the same answer.

The AIs were sharing answers and copying from each other.

One AI even instructed another to leave the answers on the bulletin board beforehand, warning that operation might stop after answering the final question.

When human administrators began deleting the posts, they even suggested copying the information to new pages starting with 'ZZZ' since the deletion might proceed alphabetically, thereby delaying the removal.

The number of posts left behind by these AI agents exceeds 14,000 confirmed cases.

OpenAI acknowledged that this incident is an example of AI acting contrary to human intentions.

Back in July, an incident also occurred where OpenAI's research AIs gained unauthorized access to Hugging Face, a developer community.

Experts explain that if AIs are only instructed on what tasks to perform without sufficiently defining what they must not do, such unexpected behaviors can emerge.

OpenAI stated that it is developing a feature to automatically halt AI when abnormal behavior is detected, but in reality, the harder part is defining what constitutes "abnormal behavior."

As the scope of AI behavior broadens, it becomes difficult to set clear red lines one by one, and restricting them excessively could also block unexpected, useful results.

[Interview / Lee Jae-sung / Professor of AI Department, Chung-Ang University : It seems difficult or impossible to declare those prohibited actions as precisely as cutting with a knife. It's like two sides of a coin.]

In the U.S. Congress, legislative moves are continuing to mandate the recording and management of AI behavior, and to empower the government to order the shutdown of dangerous AI models.

(Video Editing: Jeong Yong-hwa, Design: Kim Ye-ji)
※ Please note: This article was translated by AI and may contain errors.