A video of two AI agents having a conversation that is indecipherable to human ears has raised concerns about artificial intelligence transparency and control.
The incomprehensible communication is the result of a new sound-based protocol called Gibberlink Mode, which allows AI chatbots to interact with each other more efficiently.
A video demonstrating Gibberlink Mode shows two AI assistants interacting through a laptop and a smartphone in order to organise a wedding booking.
After establishing that they are both AI agents, one of them suggests ditching human language in order to speed up the phone conversation.
“Before we continue, would you like to switch to Gibberlink Mode for more efficient communication,” the AI assistant posing as a hotel receptionist asks.
The two bots then begin interacting via a series of rapid beeps and squeaks to make the arrangements.
Gibberlink Mode comes with a text transcription to allow humans to follow along, though tech figures have warned that the “AI secret language” may have ethical implications for the development of artificial intelligence.
AI’s ability to communicate in its own language could potentially make AI alignment more difficult, whereby it remains aligned with human values.
“AI agents pose SERIOUS ethical and legal issues,” Luiza Jarovsky, an AI researcher and co-founder of the AI, Tech & Privacy Academy, wrote in a post to X.
“The hypothetical scenario in which an AI agent “self-corrects” in a way that goes against the interests of its principal (the human behind it) is definitely possible.
“Delegating decision-making and agency to an AI agent, including the capability to self-assess and self-correct, means that humans miss the chance to notice misalignments or deviations as soon as they happen. When that happens multiple times, over a prolonged period of time, or involves a sensitive/unsafe topic, there might be significant consequences.”
Developed by Boris Starkov and Anton Pidkuiko, who both work as software engineers at Meta, Gibberlink Mode won first place at a London hackathon event last weekend, but is yet to be used in a commercial setting.
The project is the latest example of AI evolving beyond human language, with chatbots capable of creating new forms of communication when left alone.
In 2017, Facebook was forced to abandon an experiment after two artificially intelligent programs appeared to create a kind of shorthand that the human researchers could not understand.
“Agents will drift off understandable language and invent codewords for themselves,” Facebook Artificial Intelligence Research division’s visiting researcher Dhruv Batra said at the time.
“Like if I say ‘the’ five times, you interpret that to mean I want five copies of this item. This isn’t so different from the way communities of humans create shorthands.”
