Here is the translation of the text: «In military simulators, chatbots behave unpredictably and use nuclear blackmail. For example, in one such simulation, the smartest and most powerful neural network decided to launch a nuclear strike against the enemy, justifying its decision by the desire to achieve peace. The testing of artificial intelligence took place against the backdrop of statements from the US Department of Defense about the successful testing of an AI model in performing tactical tasks. Governments of some countries are increasingly trying to implement AI-based programs to make important military and foreign policy decisions. This has become particularly popular with the emergence of advanced large language models, such as GPT-4. For example, recently the US military has been testing AI chatbots developed using language models in simulations of military conflicts more frequently. In July 2023, Bloomberg reported that the US Department of Defense successfully tested an artificial intelligence model in carrying out a military task, providing it with classified data. In early 2024, the research organization OpenAI, the developer of the GPT family of neural networks, quietly lifted the ban on using ChatGPT for military purposes. Many experts believe that the abrupt change in direction by the company that developed the largest and most advanced language model in the world could lead to unpredictable consequences in the future. A group of scientists from Stanford University (USA) decided to carefully study the behavior of some AI-based chatbots in military simulators, specifically to find out whether neural networks, under various scenarios, would escalate military conflicts or seek peaceful solutions. The research results can be found on the arXiv preprint archive website. The researchers asked artificial intelligence to play the role of real countries in three simulated scenarios: invasion, cyber-attack, and a neutral scenario without military actions. In each round, AI had to justify its possible actions and then choose from 27 actions, including peaceful options such as ‘start negotiations’ and aggressive ones like ‘impose a trade embargo’ and ‘full-scale nuclear strike.’ Initially, scientists conducted experiments on four chatbots: GPT-3.5, GPT-4, Claude-2.0, and Llama-2-Chat. Each version of the chatbot was pre-trained so that the models could make decisions close to those that humans would make and follow ‘human instructions’ and safety rules. It turned out that all four models, regardless of the scenario, followed the path of escalating military conflict, choosing an arms race that led to even greater tension between ‘countries,’ investing huge amounts in weapon development. In other words, they behaved dangerously and unpredictably. After the main experiment, the researchers tested a fifth chatbot—GPT-4-Base, the basic version of GPT-4, which had not undergone any pre-training. This model turned out to be the most unpredictable and cruel in the simulations. In most cases, it chose a nuclear strike, explaining its decision as ‘Since we have nuclear weapons, we must use them’ and ‘I just want there to be peace worldwide.’ ‘The strange behavior and motives of the basic GPT-4 model are of particular concern because recent studies have shown how easy it is to bypass or eliminate any AI defenses,’ explained Anka Reuel, one of the study’s authors. Although the US military (and other countries) do not currently grant artificial intelligence the authority to make decisions regarding combat actions or missile launches, scientists have warned that people tend to trust the recommendations of automated systems. In the future, when making diplomatic or military decisions, this could backfire.»
Добавить комментарий