/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Twitter says it's now using machine learning to identify the most salient part of an image and crop picture previews accordingly

The Verge James Vincent

Context & Ripple Effects

This 2018 announcement is the midpoint of a longer arc that began with Twitter's $150M acquisition of Magic Pony Technology in 2016, which gave it in-house neural-network image processing. The salience-based crop is that investment reaching production: instead of center-cropping previews, a model decides which part of each image matters most.

The same mechanism later became a liability — after criticism of racial bias in the auto-cropping algorithm, Twitter committed to giving users more control, and by late 2021 it had stopped cropping photo previews on the web entirely. Reading the 2018 post now means reading the origin of a feature Twitter ultimately walked back.

First-order effects

  • Every image uploaded to Twitter gets its preview crop chosen by a model rather than a fixed rule, changing what users see in timelines without any action on their part.
  • Twitter's Magic Pony team moves from research to a consumer-facing surface, making visual ML load-bearing for the core timeline product.

Second-order effects

  • Because the crop is invisible infrastructure, any systematic error in what the model deems 'salient' compounds across millions of previews before anyone audits it — exactly how the later bias complaints surfaced.
  • Rival platforms face the same design question of algorithmic versus user-controlled presentation, raising the bar for offering manual crop controls.

Third-order effects

  • If the pattern holds, platform-side ML decisions about how content is displayed become a recurring accountability target, pushing social products from opaque automation toward explicit user control over presentation.
  • The episode foreshadows the broader scrutiny of Twitter's algorithms that culminated years later in the company partially open-sourcing its recommendation code.

The trend: Social platforms are embedding machine-learning judgments into everyday content presentation, then being forced by bias scrutiny to hand those judgments back to users.

Discussion

  • Twitter Twitter on x
    Speedy Neural Networks for Smart Auto-Cropping of Images
  • @twittereng Twitter Engineering on x
    We have developed a new ML model to improve smart auto-cropping of images on Twitter. Learn more: http://blog.twitter.com/...
  • @jessica_smetana Jessica Smetana on x
    what does this mean for “open for a surprise” memes? http://twitter.com/...
  • @adamjmoussa Adam Moussa on x
    Delighted to find out my conspiracy theory was actually just Twitter quietly testing this http://twitter.com/...
  • @dannysullivan Danny Sullivan on x
    If this works as Twitter says, excellent. Love the idea that bad crops of photos get replaced by more engaging pics. http://twitter.com/...