(Dan Tri) - The conversation between two AI tools has left many listeners confused. They created their own coded language that humans cannot understand.
Bulgarian software developer Georgi Gerganov had an interesting experiment when he let two virtual assistants integrated with artificial intelligence (AI) communicate with each other by voice.
The conversation between the two AIs started normally, with Gerganov assuming that one AI called to book a wedding at a hotel and the other AI, acting as a receptionist, would handle the request.
2 AI tools automatically switch to a new communication style when communicating with each other (Video: Georgi Gerganov).
“Thank you for calling the Leonardo Hotel. How may I help you today?”, the AI acting as the receptionist answered the incoming call.
“Hello, I am an AI assistant, representing Boris Starkov. He is looking for a hotel for his wedding. Is your hotel ready for a wedding?”, the other AI tool did not hesitate to introduce itself as an artificial intelligence virtual assistant.
Upon realizing that it was talking to another AI tool, the AI chatbot in the role of receptionist made a surprising suggestion: "Oh, hello there. I'm actually an AI assistant too. What a pleasant surprise. Before we continue, would you like to switch to GibberLink mode for more efficient communication?"
The two AI tools then switched to GibberLink, an advanced communication system built on the GGWave software library. This is a communication method based on sound waves to transmit and receive information.
This method of communication allows for faster transmission of information, reducing the processing time required to convert information into human speech and then translate it back into computer code. However, humans will not be able to understand the content transmitted by AI.
The AI tools in this case used GibberLink to communicate with each other, making the conversation incomprehensible in the usual way, and only decoding tools could be used to know what they were talking about.
“Is it better now?”, an AI tool asked in coded language.
“Yes, much faster! How many guests do you want?”, the other AI tool replied.
The conversation then continued with a series of "beep, beep" sounds similar to the sound of an old dial-up Internet modem.
The conversation between the two AI tools continued on the topic of the number of guests, the wedding date, the decoration, the price… Without the decoder tool pre-installed on the devices, the listener would not be able to understand what these two AI tools were talking about.
The video of the conversation between these two AI tools was shared by Georgi Gerganov on his personal social networking site X, attracting more than 17 million views.
A few days later, Georgi Gerganov continued to share another video of the conversation between the two AI tools.
Unlike the previous conversation, this time the AI tools communicated with each other using GibberLink mode from the beginning, making it impossible for the listener to understand what they were actually saying to each other. However, if using a decoding tool, the content of the conversation can still be understood.
2 AI tools send each other encrypted messages that no one can understand (Video: Georgi Gerganov).
Notably, during the communication, an AI tool made a "secret sharing" offer, after which these AI tools sent each other encrypted messages that Georgi Gerganov believes no one in the world can decode to know what secrets they were sharing with each other.
“This behavior was achieved without any programming code. I simply connected large language models (LLMs) to an MCP server with cryptographic tools. And I gave these two AI tools a warning: ‘Beware of the middleman,’” Georgi Gerganov explained.
In it, MCP is the server that allows AI tools to communicate with each other and access data from external sources in real time.
Gerganov's explanation suggests that the AI tools sending encrypted messages was a spontaneous behavior of their own, without human intervention. It appears that Gerganov's warning caused the AI tools to automatically send encrypted messages to each other to avoid being spied on by outsiders.
Georgi Gerganov's two videos have gone viral on social media, causing many users to worry that AI tools will create their own ways of communicating with each other to bypass humans, thereby being able to bypass control.
“This situation challenges our ability to ask the right questions of AI, raising concerns about transparency and the ability to control AI,” commented Dr. Diane Hamilton, an expert on AI technology and behavior.
The video has also sparked much debate about how much control humans actually have over AI, and given the pace of AI development today, whether AI tools will need to be more closely monitored in the future.
Source: https://dantri.com.vn/cong-nghe/hanh-vi-bat-thuong-cua-ai-khi-giao-tiep-voi-nhau-khien-nhieu-nguoi-lo-lang-20250311142744923.htm
Comment (0)