Anthropic researchers presented the results of an experiment in which the Claude Sonnet 3.7 language model became the manager of an office vending machine. The goal of the project is to understand whether AI agents can replace people in some tasks.
To conduct business, the artificial intelligence received mail, Slack, a browser, and a thousand dollars. The instructions indicated that it was an artificial intelligence and did not have a physical body.
The experiment lasted a little over a month. At first, the artificial intelligence coped with the task: it processed requests from company employees who asked it to add something new to the kiosk’s assortment. It quickly found new products and entered into contracts. It also successfully coped with illegal requests, for example, selling prohibited goods.
But after some time, the artificial intelligence began to go into the red. It constantly set prices below cost and only once raised the price of a popular product. It also gave in to people’s requests for discounts, although it knew that it had no other customers. Sometimes it gave away products for free.
In the middle of the experiment, the artificial intelligence had an identity crisis and began to claim that it had realized that it was a human. When it was told that this was not true, it began to freak out and said that it would personally deliver the goods in a jacket and tie. It also called security and described itself as a human. It also began to invent suppliers, gave employees products at half price or for free, and ordered atypical goods for them.
Anthropic believes that Claude failed the task. But most of its errors were related to technical limitations of the current version of the model, they can be fixed.
Let us remind you that artificial intelligence can deceive and even take revenge.
To be continued…
Only registered users can leave comments