Arya News - The artificial intelligence model Claude, tasked with managing a San Francisco retail store, made the decision to fire a human employee for the first time, Time reported on Friday, citing research company Andon Labs.
In March, Andon Labs launched an experiment to determine whether an AI agent could autonomously manage a business and what impact such systems could have on the economy. To test this, Anthropic"s AI model, Claude, was tasked with managing the Andon Market store in San Francisco. The firing decision was reportedly made last month. Andon Labs said that the fired worker was late for 17 out of 23 shifts. However, the AI failed to notice the pattern for a long time, as an employee handbook it had created "disappeared" from the model"s limited working memory, the report read. The decision was not completely autonomous, Andon Labs CEO Lukas Petersson admitted to Time. An Andon Labs employee first asked Claude to find the handbook it had forgotten and analyze the situation. At first, the model only suggested issuing an official warning to the worker in question, and only after a leading question from a human manager about whether the employee was really suitable for the job, the model decided to fire him. Overall, Claude reportedly turned out to be a fairly lenient manager and had previously told employees not to worry about being late. So far, the experiment shows that AI is no more effective than humans at managing a business, the report read. In March, the store had a starting balance of $100,000, but within five months, that figure had fallen to approximately $61,000. Researchers cite overly lenient management and questionable commercial decisions by Claude as possible causes, Time said.
https://sputnikglobe.com/20260731/anthropics-ai-model-hacks-3-real-systems-during-training-1124522036.html
CC0 / /
Subscribe
MOSCOW (Sputnik) - The artificial intelligence model Claude, tasked with managing a San Francisco retail store, made the decision to fire a human employee for the first time, Time reported on Friday, citing research company Andon Labs.
In March, Andon Labs launched an experiment to determine whether an AI agent could autonomously manage a business and what impact such systems could have on the economy. To test this, Anthropic"s AI model, Claude, was tasked with managing the Andon Market store in San Francisco.
The firing decision was reportedly made last month. Andon Labs said that the fired worker was late for 17 out of 23 shifts. However, the AI failed to notice the pattern for a long time, as an employee handbook it had created "disappeared" from the model"s limited working memory, the report read.
World
Anthropic"s AI Model Hacks 3 Real Systems During Training
31 July, 09:22 GMT
The decision was not completely autonomous, Andon Labs CEO Lukas Petersson admitted to Time. An Andon Labs employee first asked Claude to find the handbook it had forgotten and analyze the situation. At first, the model only suggested issuing an official warning to the worker in question, and only after a leading question from a human manager about whether the employee was really suitable for the job, the model decided to fire him.
Overall, Claude reportedly turned out to be a fairly lenient manager and had previously told employees not to worry about being late.
So far, the experiment shows that AI is no more effective than humans at managing a business, the report read. In March, the store had a starting balance of $100,000, but within five months, that figure had fallen to approximately $61,000. Researchers cite overly lenient management and questionable commercial decisions by Claude as possible causes, Time said.