Skip to content
Oct – Dec 2025 · Internship

Bias first, not last.
Recognized: 31 to 78%.

My impact
Biasly feed card with bias spectrum anchored above the headline
Role
Product design intern
Team
PM, 2 engineers
Tools
Figma, Maze, Notion

69% of participants could not identify the bias of articles they had just spent time reading. Almost all of them said the same thing: they never noticed it.

Overview

They read the headline first, and the bias came too late. The indicator sat at the bottom of every card, arriving after the reader had formed an opinion. I moved it above the headline and made it a spectrum.

Open the prototype
My impact

The same readers who scrolled past the indicator now caught it before the headline sank in.

  1. 01Sessions showed readers forming opinions before the indicator ever appeared.
  2. 02So I killed the overlay in three sessions and moved bias above the headline.
  3. 03Re-tested the same twelve people, same articles, one variable changed.
0%
bias recognition

Up from 31% on an identical task with an identical group. n = 12.

0×
faster recognition

Median time to identify bias fell from 9 seconds to 3.

0%
drop in missed indicators

Observer-coded instances of bias data being overlooked entirely.

Before and after

Same participants, same task, one variable changed.

Same twelve people, same articles, same time on task. The layout was the single change.

bias recognition, before
31%
Readers who could name the bias of an article they had just read.
bias recognition, after
78%
Identical task, identical group. The indicator moved above the headline.
median time to recognise
9s → 3s
Three times faster. The spectrum reads at scroll speed; the text label did not.
missed indicators
−60%
Observer-coded instances of bias data being overlooked entirely.
n = 12, moderated, same cohort across both rounds, on the prototype.
Before · chip under the headline
Before: the bias chip sits under the headlineAfter: the bias chip sits on the image, above the headline
Same card, same data. Toggle to see where the lean moves.
Evidence

An overlay was dismissed as an ad within one second, every time.

Every session dismissed the floating badge inside a second. A decade of the web has trained readers to ignore anything that floats on content. Structural position cannot be trained away.

The most useful finding came from users I was not designing for.

Two politically engaged participants wanted to judge the headline before knowing the leaning. That forced a decision about who the product is for. The header stayed, and the finding rewrote the onboarding copy.

Who it is for

“I usually just read the headline. I didn’t even see the bias tag.”

A politically aware reader · 12 moderated sessions · ages 22 to 45
  • A fast read on the day, not an evaluation exercise
  • Context that arrives before the first impression forms
  • A signal that survives fast scrolling

The anti-persona matters here. A reader who has already decided which sources to trust has no interest in being challenged, and the redesign is deliberately not for them.

Solution 01

Bias arrives before the headline, as a spectrum instead of a word

The pill moved to the card header, first in every scan sequence. The label became a spectrum, so direction and intensity register without reading.

In the header

Every participant began scanning at the top of the card, which makes the header the only position that cannot be skipped.

Before the headline·100%
What the spectrum encodes
Left
blue
Center
purple
Right
red
In the hierarchy

The only way to give bias enough weight to compete with a large image.

Images cut from banners to thumbnails
Full-width bannerConstrained thumbnail
Across surfaces

A code that changes between screens is not a code.

Same encoding in the article view
feed · article · detail
One system to learnNo re-interpretationlearned once, read everywhere
What is not here

Some of this work is not mine to publish

Internship work on a live product. Two things are missing, and the reason is the same for both.

not published
The production feed
I left before the release cycle closed and cannot confirm what shipped. Claiming a launch I did not witness would be the easiest thing here to disprove.
internal
Session recordings and the raw scoring sheet
Participant footage and internal scoring stayed with the company. The figures are from my own notes.
A case study that shows everything either had nothing confidential in it or should not be public.
Beyond the feed

The same rule, a second surface

Following runs on the same card. Once the spectrum earned its place above the headline in the main feed, there was no argument for keeping it below on this one.

The Following page before and after, with the bias chip moved above the headline
Not a new decision. The same one, applied where it had not reached yet.
The encoding

A label you read, a colour you see

Words have to be read. A position on a spectrum is seen.

The bias spectrum colours and the three things that changed between versions
Blue as left and red as right is an American convention. A limit of this design.
Read the full story

Why position, not visibility

The bias indicator was never missing. It sat at the bottom of every card, and 69% of participants could not name the bias of an article they had just spent time reading. Almost all of them said the same thing: they never noticed it. The problem was not that the signal was absent. It was that it arrived after the reader had already formed an opinion.

What the overlay taught me

Before moving the chip, I tested a floating badge that appeared when the reader paused on a card. Every session dismissed it inside a second. A decade of the web has trained readers to ignore anything that floats on content, and that training cannot be undone by a better badge. Structural position cannot be trained away the same way. Three sessions ended that direction.

What this comparison does not prove

Twelve people, moderated, same cohort across both rounds, on the prototype rather than in production. The comparison is worth reading only because nothing else moved: same articles, same time on task, one layout change. It is not evidence that the redesign changes what readers believe, only that they can name the lean afterward. Participant footage and internal scoring stayed with the company.

Reflection
01

The disconfirming finding was the valuable one

Feedback from politically engaged users who found the upfront label distracting was more useful than the data that agreed with me. It clarified who the product is for and changed decisions beyond the feed.

02

Sequence is a design material

The indicator was always visible. It was never early. When something appears is a larger intervention than how it looks.