Zi Qi Liow

This research is funded through the AI Institute for Artificial and Natural Intelligence (ARNI), a $20 million multidisciplinary initiative established by the National Science Foundation (NSF).

PI: Kathleen McKeown, Columbia
Co-PIs: Brian A. Smith, Columbia; Tatiana Aloi Emmanouil, CUNY; Nikolaus Kriegeskorte, Columbia; Carl Vondrick, Columbia

Overview

A picture is worth a thousand words, but only if the viewer can experience it. Our project aims to help blind and low-vision (BLV) people experience works of art, including their sensory and semantic dimensions, their emotional valence, the artist’s intentions, and other contextual information. Our project brings together researchers from vision, language, and interaction design, all informed by cognitive science, to develop AI Art Immersion for the visually impaired. Our work is inspired by the human visual system, which gives us an instant sense of our surroundings and lets us explore a complex scene interactively with eye movements and internal shifts of attention, enabling us to rapidly grasp the layout, the objects and characters present, and how the elements relate to each other.

Partnerships

• Helen Keller Services for the Blind
• National Gallery of Art
• New York Hall of Science

Overview

Short-term mental fatigue frequently disrupts player engagement in puzzle games. While Dynamic Difficulty Adjustment (DDA) is commonly used to maintain game flow, its invisible modifications to game parameters can compromise player autonomy. Conversely, psychological research shows that micro-breaks restore cognitive capacity in demanding tasks, yet no study has directly compared DDA and diegetic micro-breaks in an interactive puzzle environment. We evaluated both interventions using a modified clone of The New York Times Spelling Bee in a within-subjects study (N = 9). Results indicate that diegetic micro-breaks significantly enhance reported refreshment, autonomy, and perceived fairness, whereas DDA sustains engagement at the expense of player control.

Overview

Designing social media platforms to foster healthy democratic discourse is a critical challenge driven by the limitations of today's engagement-focused, ad-revenue model. To spark cross-disciplinary innovation beyond these profit-driven incentives, this white paper introduces a set of actionable "mechanisms" designed to reduce harms like misinformation and political polarization. Ultimately, this paper serves as both an entry point for public and expert debate and a proposal of designs aimed at rebuilding a digital public sphere that enhances collective intelligence.

Overview

CitizenAI is a 3-year, $1.5M NSF-funded initiative by Columbia University’s Trust Collaboratory, CS3, and Rutgers University. The project creates a bottom-up model for responsible AI development, giving the public direct input into how AI tools are built and deployed.

Key Aims & Focus

• Participatory Co-Creation: Partner with civic organizations, advocacy groups, local businesses, artists, and residents through a participatory design process to build teLLMe, a community-based conversational AI engagement tool.
• Human-AI Interaction & Public Trust: Conduct interdisciplinary research across social science, engineering, and data science to understand human-AI interactions and build public trust in technology.
• Responsible AI & Governance: Generate foundational insights to inform responsible AI policy, tech governance, urban data ethics, and the future of human-centered technology.

Overview

Columbia Business School hosts an AI Impact Competition annually, a month-long start-up challenge to propose a pre-launch idea. Finalists pitch in front of VC investors and professionals from a16z, Bain, Google Gradient Ventures, and more. Our team entered the health care track. The task was to propose an AI-driven solution for your given track, with emphasis on the technical approach.

Backstory

In the summer of 2021, I interned at a speech therapy clinic in Hong Kong, where I spent three months training under speech-language pathologists to support children with Selective Mutism and Social Anxiety Disorder. I observed therapy sessions, led behavior reinforcement exercises, and helped design exposure therapy activities tailored to each child. Over time, I noticed a clear gap: once the children left our clinic, there was no structured way to monitor their speech progress or reinforce therapeutic techniques at home. The system lacked a structured way to meaningfully connected outside of sessions.

How should we design clinician-AI co-intelligence to personalize speech therapy?

Our process involved extensive literature review, collaborative ideation sessions, user interviews with speech-language-pathologists, iterative prototyping, and feedback sessions about clinical integration with mentors.

Our Solution

Sori is an AI-powered storybook app that detects speech abnormalities and tracks a child's performance over time. Children use the AI chatbot to generate their own story plots and characters, creating personalized storybooks embedded with phonemic structures designed to assess speech patterns seamlessly. After each reading session, results are sent directly to the child’s speech-language pathologist, enabling more tailored and responsive treatment plans. Ultimately, Sori makes the therapy process more fun and personalized to each child.

Overview

Hosted by the University of Toronto, the Global Learning Case Competition (GLCC) invites students globally to collaborate on innovative solutions for pressing global challenges. Over 72 hours, participants work in teams, culminating in a final presentation evaluated by a panel of judges. Our prompt was: “What is one innovative way that artificial intelligence can be used for positive social change in your discipline(s) or communities?”

Backstory

Post-Traumatic Stress Disorder (PTSD) is a mental and behavioral disorder that can develop from directly or indirectly experiencing a traumatic event.

PTSD is characterized by:

• Re-experiencing of the traumatic event(s)
• Persistent emotional and physiological arousal in response to reminders of the trauma
• Anxiety in anticipation of the next panic attack

In the physical world, there are ways to manage symptoms by avoiding certain locations, people, or objects. But in the digital world, where content is constant and unfiltered, it only takes one triggering ad, video, photo to activate a panic response in PTSD patients. As a consequence, they may avoid digital spaces (such as social media and websites) to protect themselves, which can exclude them from participation in their personal and work environments.

How should we re-design digital spaces to help people with PTSD navigate content safely?

Our Solution

A plugin extension for text, audio, and visuals—designed with clinical support to help modulate a patient’s exposure to online triggers.

1. Step 1: Trigger Identification — Patients work with a psychiatrist to identify their personal triggers.
2. Step 2: LLM Model Training — An AI model is trained on diverse datasets to detect and categorize triggers across text, audio, images, and video.
3. Step 3: Real-Time Web Scraping — The plugin continuously scans incoming digital content.
4. Step 4: Content Modulation — As users interact with content, the plugin responds in real time. They can customize exposure levels (e.g., 90%, 50%, 10%). Based on these settings, the tool can blur, warn, or remove triggering material from their feed.
5. Step 5: Patient Dashboard — ModU's interactive dashboard allows patients to track and reflect on their therapy progress. With the clinician in the loop, patients can gradually increase their exposure to triggering content over time, giving the power of re-integration back into the patient's hands.

Through clinician-AI collaboration, our tool is personalized to one's unique triggers, empowering PTSD patients to navigate online spaces with confidence and security.