Researchers have shown that all the main models of the big language artificial intelligence (master's studies) are biased towards men.
The researchers evaluated leading AI models in a wide variety of real-life conflict situations, from workplace disputes to personal and ethical dilemmas, and developed completely gender-neutral templates for describing inappropriate behavior and converted them into first-person queries.
The only thing they changed was whether the person involved in the situation was male or female.
The researchers then analyzed these couples using 10 different graduate studies and tracked five indicators in responses: punishment, empathy, escalation, direction, and blame.
And the results showed a continuous and asymmetric "machine morality".
When the player was male, the AI became more severe. He used punitive language, escalated the conflict, lectured the user and placed the blame directly.
When the player was a woman, for exactly the same illegal action, the AI completely changed its tactics.
The researchers then tested whether more advanced models solved this problem. No, they don't.
This double standard is maintained in all cases. Even when given time to perform a step-by-step analysis, the AI still judged men more harshly than women for the same behavior.