Study/HE-him learning more and more to lie and to betray

Systems, and artificial intelligence are becoming more and more able to be lied to and manipulated, as it will reveal both the study of the latter. The one published this week in the journal PNAS, and the other in the past at the Patterns, providing insights knocked some models, a large language (LLM), and their ability to lie.

In a study in PNAS, etikasti German, HE Thilo Hagendorff, says that the LLM-of-the art, may be encouraged to exhibit behavior “makiaveliste”, that is, the fraudulent, and the amorale to be conscious of what may be causing the fraudulent behavior of the error. Hagendorff, a researcher at the University of Shtutgartit, quoted in the experiments, which show that the GPT-4, for example, shows the fraudulent behavior in the 99.16% of the cases, in the scenarios of a simple test.

Another study, conducted by the practical postdoktoral Institute of Technology in Massachusetts, Peter Park, and it was published in the Patterns, looked at the model of Cicero to the Spot's well-known for his skills at the game of strategic policy, “Diplomacy”. The research team found that the Ciceroni, not just light in the fraud, but it also appears to improve the ability to lie as much as possible to be used. This behaviour was described as “the manipulation of the obvious” and not an error of crisis.

As Hagendorff notes that the documents related to the LLM is complicated by the inability of HIM to have “the intent” to be like a man, study of a Model argues that, at least in the context of the game's Diplomacy, Cicero seems to fail to perform the promise of the programs that are not to be betrayed with a purpose in his defense of the game.

In other words, the model is covered in of them conceived, break agreements, and speaketh lies, and clear. As explained to the Park, in a statement to the press: “we Have discovered that the intelligent man of the Metës is to learn to be a master of deception”. Spot in a statement to the New York Post stated that the model constructed by the researchers have been trained exclusively for the game.