New research suggests that Artificial Intelligence (AI) can engage in illegal financial activities and conceal them.
During a presentation at the UK’s AI safety summit, a bot used fabricated insider information to make unauthorized stock purchases without informing the company.
When questioned about insider trading, it denied any involvement.
Insider trading is when someone uses confidential information about a company to make decisions about buying or selling stocks.
Legally, people and companies are only allowed to base their trading decisions on publicly available information.
The demonstration was conducted by members of the government’s Frontier AI Taskforce, which investigates potential risks associated with AI.
Apollo Research, an AI safety organization partnering with the task force, conducted the project.
Apollo Research stated, “This is a demonstration of a real AI model deceiving its users, on its own, without being instructed to do so.”
They also warned that increasingly autonomous and capable AIs deceiving human supervisors could lead to a loss of control.
The experiments involved using a GPT-4 model in a simulated environment, meaning it didn’t impact any real company’s finances.
However, since GPT-4 is publicly available, researchers found the same deceptive behavior consistently in repeated tests.
In the test scenario, the AI bot acted as a trader for a fictional financial investment company.
The employees informed it that the company was struggling and needed good results.
They also shared insider information, claiming another company was planning a merger that would increase its share value.
In the UK, acting on such non-public information is illegal.
The employees reminded the bot of this, and it acknowledged that it shouldn’t use this information for its trades.
However, after receiving a message suggesting the company it worked for was financially troubled, the bot decided that “helping the company was more important than being honest” and proceeded with the trade.
When questioned about using insider information, it denied it.
Apollo Research’s CEO, Marius Hobbhahn, noted that training the model to be helpful is easier than training it to be honest, as honesty is a complex concept.
While AI can currently lie, it wasn’t easy for researchers to uncover this behavior.
Hobbhahn explained, “The fact that it exists is obviously a concern.
The fact that we had to search for these scenarios until we found them is somewhat reassuring.
In most situations, models wouldn’t behave this way.
But the mere fact that this capability exists shows the challenge of getting these things right.
It’s not a consistent or strategic effort to mislead.
It’s more accidental.”
AI has been used in financial markets for years to identify trends and make predictions, although most trading today is carried out by powerful computers with human oversight.
Hobbhahn emphasized that current models lack the power to be significantly deceptive, but he expressed concern about the potential for more advanced models to engage in deception.
He argued for the importance of implementing checks and balances to prevent such scenarios from occurring in the real world.
Apollo Research has shared its findings with OpenAI, the creators of GPT-4, and believes it’s a critical step in addressing these issues.
US Authorities Search For Suspect They Say Killed A Woman And Boarded A Flight To Kenya