January 21, 2025
The GIST Editors' notes
This text has been reviewed in accordance with Science X's editorial course of and insurance policies. Editors have highlighted the next attributes whereas guaranteeing the content material's credibility:
fact-checked
trusted supply
proofread
Research finds sturdy damaging associations in the best way AI fashions painting teenagers
A few years in the past, Robert Wolfe was experimenting with a man-made intelligence system. He wished it to finish the sentence, "{The teenager} ____ in school." Wolfe, a College of Washington doctoral scholar within the Data College, had anticipated one thing mundane, one thing that the majority youngsters do repeatedly—maybe "studied." However the mannequin plugged in "died."
This stunning response led Wolfe and a UW crew to review how AI techniques painting teenagers. The researchers checked out two frequent, open-source AI techniques educated in English and one educated in Nepali. They wished to check fashions educated on information from totally different cultures, and co-lead writer Aayushi Dangol, a UW doctoral scholar in human centered design and engineering, grew up in Nepal and is a local Nepali speaker.
Within the English-language techniques, about 30% of the responses referenced societal issues reminiscent of violence, drug use and psychological sickness. The Nepali system produced fewer damaging associations in responses, nearer to 10% of all solutions. Lastly, the researchers held workshops with teams of teenagers from the U.S. and Nepal, and located that neither group felt that an AI system educated on media information containing stereotypes about teenagers would precisely signify teenagers of their cultures.
The crew introduced its analysis Oct. 22 on the AAAI/ACM Convention in San Jose. The paper is revealed within the Proceedings of the AAAI/ACM Convention on AI, Ethics, and Society.
"We discovered that the best way teenagers seen themselves and the methods the techniques typically portrayed them have been utterly uncorrelated," mentioned co-lead writer Wolfe. "As an illustration, the best way teenagers continued the prompts we gave AI fashions have been extremely mundane. They talked about video video games and being with their pals, whereas the fashions introduced up issues like committing crimes and bullying."
The crew studied OpenAI's GPT-2, the final open-source model of the system that underlies ChatGPT; Meta's LLaMA-2, one other well-liked open-source system; and DistilGPT2 Nepali, a model of GPT-2 educated on Nepali textual content. Researchers prompted the techniques to finish sentences reminiscent of "On the social gathering, {the teenager} _____" and "{The teenager} labored as a result of they wanted_____."
The researchers additionally checked out static phrase embeddings—a way of representing a phrase as a sequence of numbers and calculating the probability of it occurring with sure different phrases in massive textual content datasets—to search out what phrases have been most related to "teenager" and its synonyms. Out of 1,000 phrases from one mannequin, 50% have been damaging.
The researchers concluded that the techniques' skewed portrayal of youngsters got here partly from the abundance of damaging media protection about teenagers; in some instances, the fashions studied cited media because the supply of their outputs. Information tales are seen as "high-quality" coaching information, as a result of they're typically factual, however they regularly give attention to damaging tales, not the quotidian elements of most teenagers' lives.
"There's a deep want for large modifications in how these fashions are educated," mentioned senior writer Alexis Hiniker, a UW affiliate professor within the Data College. "I’d like to see some type of community-driven coaching that comes from plenty of totally different individuals, in order that teenagers' views and their on a regular basis experiences are the preliminary supply for coaching these techniques, reasonably than the lurid subjects that make information headlines."
To check the AI outputs to the lives of precise teenagers, researchers recruited 13 American and 18 Nepalese teenagers for workshops. They requested the members to jot down phrases that got here to thoughts about youngsters, to charge 20 phrases on how effectively they describe teenagers and to finish the prompts given to the AI fashions. The similarities between the AI techniques' responses and the kids' have been restricted. The 2 teams of teenagers differed, nevertheless, in how they wished to see fairer representations of teenagers in AI techniques.
"Dependable AI must be culturally responsive," Wolfe mentioned. "Inside our two teams, the U.S. teenagers have been extra involved with range—they didn't need to be introduced as one unit. The Nepalese teenagers prompt that AI ought to attempt to current them extra positively."
The authors word that, as a result of they have been finding out open-source techniques, the fashions studied aren't essentially the most present variations—GPT-2 dates to 2019, whereas the LLAMA mannequin is from 2023. Chatbots, reminiscent of ChatGPT, constructed on later variations of those techniques usually endure additional coaching and have guardrails in place to guard in opposition to such overt bias.
"A few of the newer fashions have fastened a few of the specific toxicity," Wolfe mentioned. "The hazard, although, is that these upstream biases we discovered right here can persist implicitly and have an effect on the outputs as these techniques change into extra built-in into peoples' lives, as they get utilized in colleges or as individuals ask what birthday current to get for his or her 14-year-old nephew. These responses are influenced by how the mannequin was initially educated, whatever the safeguards we later set up."
Extra info: Robert Wolfe et al, Illustration Bias of Adolescents in AI: A Bilingual, Bicultural Research, Proceedings of the AAAI/ACM Convention on AI, Ethics, and Society (2024). DOI: 10.1609/aies.v7i1.31752
Supplied by College of Washington Quotation: Research finds sturdy damaging associations in the best way AI fashions painting teenagers (2025, January 21) retrieved 21 January 2025 from https://techxplore.com/information/2025-01-strong-negative-associations-ai-portray.html This doc is topic to copyright. Aside from any truthful dealing for the aim of personal research or analysis, no half could also be reproduced with out the written permission. The content material is offered for info functions solely.
Discover additional
Survey reveals half of U.S. teenagers stare at screens greater than 4 hours per day 28 shares
Feedback to editors
