GenPod: Constructive News Framing in AI-Generated Podcasts More Effectively Reduces Negative Emotions Than Non-Constructive Framing

TL;DR

GenPod presents a generative AI pipeline that creates constructive and non-constructive news podcasts from identical source material to study framing effects on listeners. In a mixed-methods study with N=65, constructive framing reduced negative emotions and, in some contexts, increased self-efficacy relative to non-constructive framing. The work demonstrates that presentation style alone can shape emotional and motivational responses to AI-generated audio, and discusses design, regulatory, and ethical implications for responsible deployment. The findings support constructive journalism principles and suggest pathways for AI-mediated personalization of news with safeguards against manipulation.

Abstract

AI-generated media products are increasingly prevalent in the news industry, yet their impacts on audience perception remain underexplored. Traditional media often employs negative framing to capture attention and capitalize on news consumption, and without oversight, AI-generated news could reinforce this trend. This study examines how different framing styles-constructive versus non-constructive-affect audience responses in AI-generated podcasts. We developed a pipeline using generative AI and text-to-speech (TTS) technology to create both constructive and non-constructive news podcasts from the same set of news resources. Through empirical research (N=65), we found that constructive podcasts significantly reduced audience's negative emotions compared to non-constructive podcasts. Additionally, in certain news contexts, constructive framing might further enhance audience self-efficacy. Our findings show that simply altering the framing of AI generated content can significantly impact audience responses, and we offer insights on leveraging these effects for positive outcomes while minimizing ethical risks.

Figures (3)

• Figure 1: Overview of GenPod’s pipeline to generate constructive and non-constructive podcasts. The initial input consists of news on the same topic, sourced from authoritative news platforms, with the final output being two types of podcasts: constructive and non-constructive. The process involves five agents in sequence, while the System Prompt serves as a foundation throughout the former four agents. Agent 1 dissects news sources into basic components: Agents 2 and 3 recompile these components using constructive or non-constructive frames. Agent 4: evaluate whether the recompiled content Agent 4 evaluates whether the recompiled content meets the requirements of news structure. Agent 5 transforms the plain text into podcast scripts. Finally, (Text-To-Sound)TTS technology and post-production sound effects are used to produce the audio.
• Figure 2: The detailed prompt structure and the function of Agent 1. The prompt for Agent 1 includes Task Instructions and several examples. explicitly guide the agent to dissect each original article into five components: headline, lead, body, background, and conclusion, without any alterations. This is based on a system prompt that defines and explains each of these five elements. Examples demonstrate some news articles alongside their corresponding dissection results. When multiple news articles on the same topic are input, Agent 1 dissects them and categorizes the results by element type, ultimately outputting five lists, serving as the input for the subsequent phase.
• Figure 3: Results of the post-survey showed that constructive podcasts significantly reduced negative emotions and enhanced self-efficacy on topic 1: food delivery workers. ** indicate significance of $p$ < 0.01.