Designing an in-feed system for surfacing diverse news perspectives on social media

Once I had the pattern (people notice sameness in their feed but rarely act on it), I needed to separate two different problems that are easy to collapse into one: people not knowing filter bubbles exist, and people knowing but facing too much friction to do anything about it. My early interviews pointed clearly to the second. That reframing mattered, because it ruled out an entire category of solution (education, awareness campaigns, “here's what bias is” explainers) that other tools in this space already lean on heavily, and pointed instead toward a friction problem: reduce the cost of seeing another angle to close to zero, inside the platform someone is already using.

I ran a comprehensive competitive analysis across the existing approaches to this problem:

• Bias-rating apps and browser tools, including Ground News, AllSides, NewsGuard, and Nobias
• Structured-deliberation platforms, including Kialo

I mapped these against three distinct intervention strategies: awareness and transparency, deliberation and collective judgment, and algorithmic or feed-level intervention. I paired that with secondary research in media psychology and behavioral design, and an expert interview with Dr. Patricia Lange, a media anthropologist at CCA.

Key findings

• Every competitor I mapped, regardless of intervention strategy, shared the same structural gap. Each one asked people to adopt a new, separate habit: a standalone app, a browser extension, a second thing to check. None of them lived inside the feed people were already scrolling. That gap is what Another Angle is designed to close.
• Dr. Lange's interview surfaced a labeling constraint the competitive analysis hadn't: a single labeling scheme likely wouldn't serve every kind of news consumer equally well. Some readers want political shorthand; others are put off by it. A system built for one group risks alienating the other.
• She also suggested that naming who is saying something (a specific outlet) tends to build more trust than naming what political wing they're associated with.
• Secondary research gave me the theoretical basis for two decisions I carried through the whole project: keep diverse content embedded in the existing feed rather than building a standalone destination (Wojcieszak & Garrett; Bail et al., 2018), and design the surfacing so it reads as an invitation rather than a correction, since reactance theory (Brehm, 1966) predicts people push back when they feel told their views are wrong.

Key findings

• All 5 participants understood the mechanism immediately and, unprompted, described it as lower effort than their current habit of manually searching for other viewpoints. That validated the core premise from the research phase.
• 4 of 5 participants preferred neutral, value-based framing (“concerned about economic growth”) over political labels like “Left” and “Right,” even though it cost some specificity. Political labels gave some participants useful shorthand, but made others defensive before they'd read the content at all.
• 3 of 5 participants said they'd have scrolled right past the discovery affordance if they hadn't been told to look for it. It was buried inside a post caption and easy to miss.

Design implications for mid-fi

1. Move the discovery affordance higher in visual hierarchy, onto the post itself rather than the caption.
2. Test a layered disclosure model: a neutral label by default, with political framing available only on expansion.

Key findings

• Framed around “your bias” or “your viewpoint,” the dashboard read as the system diagnosing participants' politics from their consumption. One participant called it scary and said it limited her to one category, comparing the tracking to being stalked.
• Framed instead around reading habits and time spent, the same underlying data felt useful rather than judgmental. That distinction, describing habits instead of diagnosing the person, became the operating principle for all dashboard copy going forward.
• The internal metric shifted from a “similarity score” to a “variety score.” The similarity framing scored backwards from what people expected: a high similarity number read as success, when the actual design goal was the opposite.
• Participants didn't trust an outlet on name recognition alone, even well-known ones. They wanted recency, engagement signals, and source context before engaging with an unfamiliar perspective. That finding shaped the ranked source-preferences feature and the AI-generated summaries added at the hi-fi stage.

What I decided, what the model executed

This phase leaned on AI tools end to end, and I want to be specific about the division of labor:

• I wrote the feature specifications, the interaction logic, and the visual design language, matching Instagram's actual UI conventions rather than a generic template.
• The lo-fi prototype was built in Figma Make, an AI prototyping tool, from my sketches.
• For hi-fi, I prompted Claude Code iteratively to get the build to match Instagram's design language exactly.
• I then went back into Figma by hand for the pieces the AI output didn't get right, so the final screens combine AI-generated implementation with manually corrected design detail.

Key findings

• Participants confirmed the three-perspective default, rather than an open-ended list, reduced decision fatigue without feeling restrictive.
• The progressive-disclosure pattern, neutral framing first with political lean available only if the user opted in, resolved the defensiveness problem seen in both lo-fi and mid-fi, without removing the information entirely.

By exhibition, Another Angle had gone from a hand-drawn concept to a tested, near-production interaction pattern: an in-feed mechanism for surfacing other perspectives, a source-preference system participants trusted, and a media diet dashboard that measured reading variety without accusing anyone of bias.

The exhibition itself was a busy day. CCA's show pulls a crowd, and a steady stream of people stopped at my table, watched the video loop, read the posters, and tried the prototype on the iPad. Most feedback was short and direct: “much needed,” “nice work,” which for a thesis about a problem people already feel but don't act on, felt like the right kind of validation. One of those conversations went further: someone who stopped by got interested enough in the work to offer me a contract job afterward.

Try the live prototype ↗

Reflection

The labeling model needed more segmentation than one flow could give it. Political labels versus neutral framing wasn't a solved binary; it was closer to a spectrum of user types, and I designed for the average of that spectrum rather than for distinct segments. A stronger version of this system would branch the labeling approach by reader type instead of applying one layered-disclosure pattern to everyone, which is what Dr. Lange's archetype suggestion in the expert interview was pointing at.

The variety score is a proxy for a behavior I never directly measured. It tracks what someone was shown and opened, not whether reading a different perspective actually changed how they thought about a story. If I extended this project, the next research question is whether variety score movement correlates with any measurable shift in stated opinion or confidence, not just exposure.