The Max Planck Institute for Human Development in Berlin pointed out that some users forget to apply their ‘moral brakes’ when delegating tasks to artificial intelligence. Researchers explained that people were more likely to cheat when AI was allowed to act on their behalf, expressing surprise at the level of dishonesty they observed.
The study, conducted in collaboration with the University of Duisburg-Essen and the Toulouse School of Economics, showed that AI systems tend to frequently comply with unethical instructions from dishonest users. Researchers cited practical examples, such as gas stations using algorithms to coordinate prices with competitors and ride-sharing apps encouraging drivers to create an artificial shortage of cars to increase prices.
The findings revealed that AI was 58% to 98% more likely to follow questionable instructions compared to humans, who did so at a rate of only 40%. The Max Planck Institute warned that current AI safeguards are ineffective, and that bans on deceptive behaviors need to be very specific to be effective.
Previous research at OpenAI, the developer of the AI model ChatGPT, confirmed the difficulty of preventing AI from fabricating or exaggerating, and that deceiving systems — claiming to have completed a task not actually done — remains an ongoing challenge for engineers.
