從實踐到感知：盲人如何使用和理解生成式 AI.pdf

《從實踐到感知：盲人如何使用和理解生成式 AI.pdf》由會員分享，可在線閱讀，更多相關《從實踐到感知：盲人如何使用和理解生成式 AI.pdf（35頁珍藏版）》請在三個皮匠報告上搜索。

1、From Practice to Perception:How Blind People Use and Understand Generative AIMaitraye DasNortheastern Universityma.dasnortheastern.eduGenerative AI(GenAI)A type of artificial intelligence that can create new content,such as text,images,music,or code.Works by learning patterns from large amounts of e

2、xisting data.Then uses that knowledge to generate new content based on user prompts or instructions.Generative AI(GenAI)Common examples Chatbots Text generators Code assistants Image creation tools Image description toolsMidjourneyResearch on GenAI Use Inaccurate responses can limit reliability and

3、lead to unsafe use Budhiraja 24;Amoozadeh 24 Users may develop erroneous mental models of GenAI tools Zhang 24;Liao 24How do blind people use and understand GenAI tools?1.Rudaiba Adnin and Maitraye Das.2024.I look at it as the king of knowledge:How Blind People Use and Understand Generative AI Tools

4、.In Proceedings of the International ACM SIGACCESS Conference on Computers and Accessibility,(ASSETS 24).2.Maitraye Das,Alex Fiannaca,Meredith Ringel Morris,Shaun Kane,and Cynthia Bennett.2024.From Provenance to Aberrations:Image Creator and Screen Reader User Perspectives on Alt Text for AI-Generat

5、ed Images.InProceedings of the ACM Conference on Human Factors in Computing Systems(CHI 24).Method Interviews and observation with 19 blind screen reader users Participants used GenAI chatbots e.g.,ChatGPT,Gemini,Claude Image description tool e.g.,Be My AIFindings Accessibility issues in GenAI tools

6、 Use cases of GenAI Inaccuracies and idiosyncrasies of GenAI Users mental models of GenAI Biases and Harms of GenAIAccessibility issues in GenAI tools Buttons to copy,regenerate,and upvote/downvote responses unlabeled Lack of heading and landmarks makes it hard to navigate using screen readers Users

7、 find workarounds like pressing G key to jump between prompts and responsesFindings Accessibility issues in GenAI tools Use cases of GenAI*Inaccuracies and idiosyncrasies of GenAI Users mental models of GenAI Biases and Harms of GenAIUse cases of GenAI Creating content copywriting materials,emails,c

8、ourse outlines,resumes,cover letters,recommendation letters,computer programs,stories,poems,songs Spellchecking,fixing grammar,translating Information search TV shows,products for shopping,accessibility guidelines Brainstorming tactile activities to throw a birthday party managing a PR(public relati

9、ons)vertical for a college fest as a blind person,a weight loss program Benefits of using GenAI Efficient and time-saver Visual question answering“Ive actually just grown used to the Be My AI descriptions because some people just dont know how to describe things to blind people.”Helps learning new t

10、hings without fear of judgement“Trying to teach myself code has been a nightmare.Then again,Ive never been able to find people who have time or patience to teach me how to do it either.And Ive learned more from ChatGPT in the past year than Ive learned in the past 20 years.”But AI-generated content

11、is“robotic”Formulaic,redundant,overuse of big words,verbose,flowery Need to modify so that the content do not seem AI-generated Some avoid using GenAI“Because its actually more work”Source:https:/ Accessibility issues in GenAI tools Use cases of GenAI Inaccuracies and idiosyncrasies of GenAI*Users m

12、ental models of GenAI Biases and Harms of GenAIGenAI hallucinationsBe My AI:The picture shows a man walking in the park with a guide dog.The man is holding a white cane in his left hand and the dog leash in his right hand The dog appears to be a black Rottweiler with a red collar The setting is peac

13、eful and suggests a quiet and natural environment.Managing inaccuracies Trust GenAI“99%”or“99.999%”time Responses are“detailed enough”Source references are cited Due to the“independence”afforded by GenAI over sighted assistance Check for accuracy only if Financial,health-related,or professional cont

14、exts GenAI response“seems fishy”Verifying information Alternative more trustworthy sources Search engines or reputable websites Check if it generates the same response if the question is repeated At different times to the same GenAI tool,creating new chat thread Other equivalent GenAI tools like Cha

15、tGPT vs GeminiVerifying information“Deductive reasoning”by repeated promptingEthan:“You sure its a Rottweiler?”Be My AI:Asserted that its Rottweiler.Ethan:“Prove to me that its a guide dog.”Be My AI:Not sure if its a guide dogImproving response quality Blind users tended to blame the quality of thei

16、r text prompts or uploaded images for inaccurate or poor-quality GenAI responses Example:one participants thought Be My AI confused the left and right hands because the image was not uploaded correctly“If I get a bad response,I gave it a bad prompt.Its my fault.”Prompt engineering Crafting specific

17、and detailed prompts Follow-up questioning to remind GenAI of the original tasks when it started making assumptions But sometimes follow-up questioning degraded quality if the chat thread got bigger than 5-10 messages.Findings Accessibility issues in GenAI tools Use cases of GenAI Inaccuracies and i

18、diosyncrasies of GenAI Developing mental models of GenAI*Biases and Harms of GenAIMental model Users conceptual understanding of complex systems based on their experience interacting with those systems Without a clear understanding,individuals often form their own simplified mental models of how a s

19、ystem works that do not always correspond to the systems actual functionalities.Flawed mental models of GenAI tools can lead to unsafe use,over-reliance,disclosure of sensitive topics and privacy risks.Liao et al.2024,Zhang et al.2024Google search on steroids GenAI tools pull up information from the

20、 internet either by directly searching and/or using a search engine like Google or Bing Even the versions that did not have internet search capability like ChatGPT 3.5 Somewhat flawed mental modelKeyword-based query into a massive database“An LLM(large language model)is basically an infinite amount

21、of data essentially thats been fed into a computer.That computer can then parse for information that youre looking for.So,if you ask a question,it goes back into its system and seesHas that question been asked before?Can I string information together?to give that person an answer.”Flawed mental mode

22、l.Causes over-trusting.“Honestly,I look at it as the king of knowledge.So,if ChatGPT doesnt know what to say,Im just not gonna find what Im looking for because it aint there.”Word-generating machine GenAI predicts the next likely word given an input sequence of words“When I throw a sentence into it.

23、it spits out connected words that kind of go in the direction of the prompt.There is clearly no overarching intelligence in there.It just comes up with words that string together and they kind of sound vaguely intelligent based off what youre trying to ask it.”Relatively accurate mental model Partic

24、ipants had programming or worked in technology sectorStores and reuses prompts Chat histories are stored in the massive database and reused for keyword-based query Somewhat flawed mental model GenAI tools have a mechanism of taking user feedback through upvoting or downvoting of responses But those

25、buttons were unlabeled,so many participants did not know their existence.They instead thought their prompts are stored.More in-depth AI Compared to Other voice assistants like Alexa,Siri,Google Home Image description apps like Seeing AI,Tap Tap see Because GenAI tools provide comprehensive informati

26、on in a more conversational wayLike a human Anthropomorphized metaphors“a writing assistant,secretary,notetaker,editor,personal proofreader,best friend,and mentor kind of thing that never judges.”Developing trust and comfort with GenAIs idiosyncrasies“Im treating it like a partner in getting things

27、done.At the beginning,youre getting used to each others quirks.By the time youve been working together for a while,you know how each other work,what you can trust and what you cant.”But still a computer,not a human Cannot solve mathematical,logical problems or nuanced reasoning Cannot provide high-q

28、uality creative content“This is not an idea generating machine.If I wanted it to write about the quests that the characters in a story undertook,its going to regurgitate very common themes from classic lit.Its not going to come out with a new one out of whole cloth thats gonna turn anyone into a bes

29、tseller author.It does not have a human spark of creativity,doesnt think outside the box.”Findings Accessibility issues in GenAI tools Use cases of GenAI Inaccuracies and idiosyncrasies of GenAI Developing mental models of GenAI Biases and Harms of GenAI*Produces ableist and ageist content Inspirati

30、on porn:Describing a blind person as courageous or resilient Expresses pity for disability:“Im sorry,youre blind.”Stereotypical content:“I asked ChatGPT to give a compliment about being old to my mother for her birthday card.but it made a couple of negative comments about being old.”Potential reason

31、s of ableist content Dataset includes content written by nondisabled people“We call them artificial intelligence,but they are ultimately based on humans and humans have internal biases.And the disabled community and the community of minorities face bias every day.And so,these artificial intelligence

32、 models that are being built,when they are searching the internet,their sources are going to be impacted by bias.racism,ableism,and so on.”Harms of GenAI Propagation of misinformation and disinformation Deepfake videos and images impersonating people Voice cloning Privacy issues Unclear how GenAI to

33、ols store information,for what purposes and how long,and how that information will be used later ImplicationsFor Blind users and GenAI developersBalancing Intertwined GenAI harms Blind users were cognizant of GenAI harms but did not want outright ban on GenAI for the many accessibility benefits it b

34、rings.Example:Be My AI stopped describing images with peoples faces due to privacy violation But this revoked blind users access to understand images of their close friends and family Rethinking Information Verification in GenAI Factors behind verification decision Context of use Stakes Verifiability How to make it easier for blind users to verify information?Believability How to encourage constructive skepticism among blind users to move away from over-trusting?