Ash Feng

AI Engineer · Former Journalist in China

Nagoya, Japan

Designing media environments for mutual understanding.

Social Computing Perceived Polarization Perception-Targeted Intervention Design

About

When a reporting series I helped produce received the China News Award — the profession's highest honor in China — I found myself unable to celebrate. I had entered journalism believing that better content could change how people understood one another. But I came to see a structural limit: content matters only within the systems that distribute and amplify it.

I entered journalism to change what people read; I left it to change how they encounter information. So I moved upstream. I taught myself to program and became an engineer, shifting from producing content to building the systems that mediate it. Today I build content-generation AI agents as an engineer in Japan.

Engineering taught me how to build interventions, but not how to know whether they work. That requires controlled experiments and causal inference — the training I now seek through a PhD beginning in Fall 2027.

Full CV →

Current Research

Turning mechanisms of misperception into system design

People systematically overestimate how extreme and hostile the other side is — a phenomenon known as perceived polarization. I focus on how media environments shape the cues people use to infer what others believe.

My work follows a mechanism-to-design approach. I start with documented mechanisms that distort social inference, translate each into a platform-agnostic intervention function, and then instantiate that function through the affordances of a particular system. This separates the principle of an intervention from any one interface or platform — and makes it possible to ask whether the same principle generalizes across interaction forms.

The goal is not to change what people believe, but how accurately they perceive one another.

Two test arenas

Algorithmic feeds · the first experimental setting, where these intervention functions can be instantiated as ranking and interface manipulations and causally tested.

Conversational AI · a second arena for testing whether the same underlying functions can be realized — and remain effective — in a different interaction architecture.

I built Lichtung to make these interventions experimentally testable. In my PhD, my next step is twofold: to test their causal effects in feed environments, and to translate the underlying intervention functions into conversational AI as a second test arena.

Experimental testbed

Lichtung 林间空地

A platform-independent social-feed simulator I built to turn these intervention designs into experimentally manipulable conditions, without relying on commercial-platform APIs. Three modules are currently implemented:

— Counter-stereotypical exemplars · surfacing out-party voices that cut against the stereotype, implemented through ranking
— Distribution overlay · revealing that moderate views form the majority within each camp, through interface cues
— Non-political identity surfacing · restoring multidimensional identity cues by displaying attributes people share outside politics

Live system

lichtung.ashfeng.com

No signup — the feed loads running.

Status

Built three intervention modules

Proposed causal testing, in my PhD

Open the live demo →

Walkthrough | GitHub

React · TypeScript

How I got here

Full research narrative →

My M.A. research examined how party-press social media organizes patriotic propaganda — a content and frame analysis of the People's Daily WeChat account; later, I built R-Selected, a news aggregator that placed coverage of the same event from different political perspectives side by side, based on the intuition that broader exposure might reduce polarization. A literature review made me abandon that premise: evidence for filter bubbles was weaker than I had assumed, and cross-cutting exposure could even backfire. That failure redirected my research from what people see to how they interpret what they see.