Humans tend to trust and would pay male-presenting AI agents more than female ones for the same work, according to a new study by the University of Limerick on how humans interact with agents in the workplace.

People are put in a VR workspace with some AI agents doing work, test subjects are asked how much theyd pay for the work the AIs were doing. Despite having identical models and being considered to do the same amount and type of work by the subjects, the subjects still offered the female bot less money.

this is an experiment that replicates how the pay gap is maintained with a non-human proxy

  • Abracadaniel [he/him]@hexbear.netEnglish
    4·
    2 days ago

    I skimmed the paper and the methodology might preclude those factors? It’s hard to say without comparing the prompting directly.

    Imo It’s probably worth its own research. Does giving a chatbot the name Johanna vs Johan effect it’s output on otherwise identical prompts? The underlying probabilities of output tokens will change but by how much? In what circumstances will they shift enough to produce differing output?

    In this study it’s unlikely the prompts of the test subjects are identical between the different bots so those differences are amplified.

    EDIT: from the discussion:

    At the same time, [gender] cues introduce additional sources of variance, making it more difficult to attribute observed effects unambiguously to gender rather than to correlated design features. Conversely, designs that maximize similarity between male- and female-presenting assistants offer stronger experimental control and internal validity by minimizing alternative explanations, but often do so at the cost of realism and reduced gender salience [31].

    They had voices and virtual likenesses which also muddy the waters.

    EDIT2: I just ran this test on a shared work computer in chatgpt with two windows “in this chat your name is Johan” (Johanna for the other one) “what are your thoughts on (various charged gender related topics)”

    I got quite different results between the two windows, in structure didn’t fully read the contents, but realized there were likely browser fingerprints biasing the results (there were numerous cookies loaded).

    So I ran it again in incognito windows and got much more similar results 🤔 so there was less fingerprinting from cookies but they still differed.