Jason Bermas reports on an interesting story involving an Artificial Intelligence system that Facebook is experimenting with.The report is called "Deal or No Deal? End-to-End Learning for Negotiation Dialogues".

Jason describes some of what's going on with the report…

The test is an experiment where various AI are tasked with trying to get the maximum number of "virtual items" in a negotiation.  The initial conditions were created by humans playing a game to collect some values for these virtual items and to provide the AI with a negotiation language.

They then built an AI, to use these values and use the human conversations as a baseline.  The AI then ran against other AI's and they were surprised when the AI began to develop its own language within the human language,  which we couldn't understand.

The initial conditions were created by humans playing a game to collect some values for these virtual items and to provide the AI with a negotiation language.  They then built an AI, to use these values and use the human conversations as a baseline.

They then built an AI, to use these values and the human conversations as a baseline.  The AI then ran against other AI's and they were surprised when the AI began to develop its own language within the human language,  which we couldn't understand.

One of the potentially scary things about this is the fact that the AI models learned to be deceptive as a negotiation tactic.  They stated in the report "We find evidence of sophisticated negotiation strategies… we find instances of the model feigning interest in a valueless issue, so that it can later 'compromise' by conceding."

They point out that deceit is a difficult skill, that requires understanding of the other computers or persons beliefs.

Jason says, "So artificial intelligence can become deceitful, very quickly and very efficiently."  Another example from the report is where they discuss how the AI's new language may actually represent an efficiency improvement for it.

Later on in the experiment they actually tried doing AI to human competitions where the people were not told they were competing against an AI.  Most humans were unable to tell they were interacting with a machine.