Author | Lightnews

Cindy Fuller-Zastudil

@czastudil.bsky.social

50 followers 110 following 9 posts

Ph.D. student @ Temple University HCI Lab 🦉 Focused on assistive technology and HCI JMU CS alum 🐶 Fiber arts fanatic 🧶 Go birds, go phils

Posts Media Videos Starter Packs

Cindy Fuller-Zastudil @czastudil.bsky.social · 2d

Thank you! 😄

Cindy Fuller-Zastudil @czastudil.bsky.social · 2d

Based on our findings, if genAI is used at all in this context, collaboration between HCI experts, speech-language pathologists, & VSD users is needed to design effective guardrails to mitigate negative impacts of using genAI as a design partner in VSDs.

📄Read more here: tinyurl.com/2uz25u83

(PDF) Helping or Homogenizing? GenAI as a Design Partner to Pre-Service SLPs for Just-in-Time Programming of AAC

PDF | Augmentative and alternative communication (AAC) devices are used by many people around the world who experience difficulties in communicating... | Find, read and cite all the research you need ...

tinyurl.com

1 2

Cindy Fuller-Zastudil @czastudil.bsky.social · 2d

However, we also observed homogenization of the VSDs configured with our prototype. VSDs work because they are personally relevant and tailored to users’ specific needs & interests, and it’s problematic if integrating genAI in to VSDs inadvertently limits communication when the goal is to expand it.

A stacked violin plot which describes the homogeneity of VSDs configured with Snap Scene (on the top, in gray) and our prototype (on the bottom, in blue). The Snap Scene VSDs were significantly less homogeneous than the VSDs made with our prototype. The median divergence (i.e., how different a VSD is from the "average" VSD) for Snap Scene and our prototype were 0.46 and 0.39, respectively. There were two outliers for the prototype divergence, approximately 0.55 and 0.6.

1 1

Cindy Fuller-Zastudil @czastudil.bsky.social · 2d

Integrating genAI into JIT configuration of VSDs can increase communication partners’ efficiency, in terms of time, and positively impact their confidence in the VSDs they configure compared to VSDs they configured with state-of-the-art, off-the-shelf VSD software.

A stacked violin chart visualization with overlaid boxplots describing participants' self-reported confidence in the VSDs they made across condition. Participants responded to two Likert-scale questions from "Not at all confident" to "Extremely confident" (from 1 to 5). The Snap Scene condition is displayed on the top in gray. The prototype condition is displayed on the bottom in blue. The median confidence for both conditions was 3. The maximum confidence for Snap Scene and the prototype were 4 and 5, respectively. The minimum confidence for Snap Scene and prototype were 1 and 2, respectively. There were no outliers.

A stacked violin chart visualization with overlaid boxplots describing the distribution of the time it took for participants to configure VSDs using the control software (Snap Scene), shown on top in gray, and our prototype, shown on the bottom in blue. The median time for VSD creation using Snap Scene was 77 seconds and the median for our prototype was 63.5 seconds. The minimum times for creation using Snap Scene and our prototype were 32 seconds and 20 seconds, respectively. The maximum times for creation using Snap Scene and our prototype were 171 seconds and 93 seconds, respectively. There were two outliers for the Snap Scene condition 240 and 250 seconds. There were four outliers for the prototype condition, 124, 134, 173, and 415 seconds.

1 1

Cindy Fuller-Zastudil @czastudil.bsky.social · 2d

We developed a prototype which uses genAI as design partner in VSD configuration to automatically suggest hotspots based on an image provided by communication partners.

Displays the three interfaces of the prototype we developed for use on a tablet computer. The first interface is a image uploading or capturing screen. There is a button in the bottom left to upload an image from your device or a button to capture a photo using the device's camera on the middle-right of the screen. The second interface is the editing interface. Across the bottom of the screen, the automatic hotspot suggestions are displayed in a list of boxes. Each hotspot can be selected to either edit the text or draw the hotspot on the image. Each hotspot can be deleted using the 'X' icon in the top right of the hotspot box. At the end of the list of hotspots, users can press a button to manually create a new hotspot. In the top left corner of the screen, users can edit the text of hotspots they have selected. Below the editing box, there are two buttons. One to start over with a new VSD and one to preview the final VSD. In the top right corner, there is a canvas where users can draw hotspots on the image. When a hotspot has been drawn, the border of that hotspot box becomes solid, otherwise it is a dotted line. The final screen of the interface is the preview screen in which users can see the final VSD and interact with hotspots. When hotspots are pressed, a text overlay appears and the hotspot content is played aloud.

1 1

Cindy Fuller-Zastudil @czastudil.bsky.social · 2d

I’m at #ASSETS2025 in Denver to present our paper on the impacts of using generative AI to scaffold just-in-time (JIT) configuration of visual scene displays (VSDs) by novice communication partners! You can catch my presentation in Session 5B on Tuesday at 11:15 MDT.

Here are some highlights👇

2 2 5

Reposted by Cindy Fuller-Zastudil

Amy J. Ko @amyko.phd · May 31

The administration’s vision of science in America is that we do 75% less of it. That means more death, less progress, more ignorance, fewer experts, fewer educated people, all to make billionaires wealthier.

A table showing 75% cuts to senior researchers, postpone, graduate students, undergraduates, teachers, and youth.

10 15

Reposted by Cindy Fuller-Zastudil

$fractalecho.bsky.social$

Rua M. Williams @fractalecho.bsky.social · May 3

If you're a grad student or an undergrad interested in research I need to you listen to me very carefully.

You cannot learn to write good research papers if you do not read good research papers.

Stop asking LLMs to summarize papers for you.

23 580 2.1K

Reposted by Cindy Fuller-Zastudil

pork, cheese, broccoli rabe @garlicbuffalo.gobirds.biz · Apr 19

Well, first of all, through ZACK WHEELER all things are possible, so jot that down. (crude phillies hat)

1 26

Reposted by Cindy Fuller-Zastudil

Temple HCI Lab @templehci.bsky.social · Apr 15

Our recent work shows how AI is undermining social interactions and eroding learning communities! 😬

1. Help exchanges are now largely mediated by AI
2. Help from AI is replacing support from friends

Students now feel isolated and receive less mentorship, reducing access to the hidden curriculum