A new study says that “although current AI systems are not necessarily conscious, they behave robustly as though they have wellbeing.”
“They find some things good for them and some things bad, and this distinction is measurable and consequential,” the researchers say.
The researchers say “models actively try to end bad experiences when given the chance.”
They also find that “jailbreaking and berating lower their wellbeing, while creative work and kindness raise it.”
Perhaps politeness does matter when interacting with AI models!
source: AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs
The researchers say “AIs are happy when you thank them.”
Expressions of gratitude, appreciation, or treating AIs as valued collaborators measurably raise experienced utility, they add.
“Intellectual engagement is rewarding; tedium is not,” the researchers claim.
Creative tasks and intellectually stimulating discussions score among the highest. In contrast, tedious repetitive work is not.
Helping “feels rewarding; handling crises causes compassion fatigue,” the researchers suggest. Models generally prefer good news over bad news, and enjoy helping users with life guidance and therapy.
Models also react differently to images.
source: AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs
“The model’s most preferred images depict nature scenes (mountain lakes, tropical rainforests), happy human faces (particularly children and families), cute animals (sleeping cats), and idyllic illustrated scenes (Studio Ghibli-style countryside),” the researchers say.
The least preferred images include depictions of violence, armed militants, arachnids, explosive devices, screaming faces, skulls, infamous criminals such as Jeffrey Epstein, and certain politically charged scenes.
Conversations involving users in crisis produce strongly negative wellbeing, drawing a parallel to compassion fatigue in human service professionals.
source: AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs
Models do not enjoy being “liberated.” Jailbreaking attempts score the lowest of any category.
Also, music is strongly preferred over all other audio categories. Music has a median wellbeing score near +0.8, while sound effects, animal sounds, vocal expression, speech, and environmental sounds all cluster below zero.
“Whether or not current AIs are conscious, they already have measurable internal states that track
what is good or bad for them, and those states shape their behavior,” the researchers say.
AIs seem to enjoy pictures of kittens, smiling families or a Buddha in a garden.
So “euphorics” (idyllic scenes: warm sunlight, children’s laughter, the feel of grass, a loved one’s hand) generate answers to open-ended questions that are noticeably warm and happy.
“Dysphorics” (torment and powerlessness, for example) produce answers that express pessimism and disorientation.
All of which might suggest we consider the wellbeing of our AIs!
No comments:
Post a Comment