Twitter says it's now using machine learning to identify the most salient part of an image and crop picture previews accordingly
The machine knows what you want to look at — The lure of machine learning isn't always about big new features; often, what it does best are small tweaks that subtly improve user experience.
Context & Ripple Effects
Twitter's salient-image cropping is the product payoff of a bet it made eighteen months earlier: the acquisition of Magic Pony Technology, a visual-processing machine-learning startup reportedly bought for $150M. The 2018 feature is exactly the kind of deployment that purchase was for — not a headline-grabbing new product, but an invisible tweak where the model decides which part of a photo survives the preview crop.
First-order effects
- Users posting photos lose control over framing: Twitter's model, not the poster, now chooses which region of an image appears in timeline previews.
Second-order effects
- The same automation becomes a liability once it scales — by late 2020, criticism that the cropping showed racial bias forces Twitter to concede the 'salient' framing and promise users more control, before it drops photo-preview cropping on the web entirely in 2021 after mobile did the same that May.
Third-order effects
- The arc from silent ML default to bias complaint to full retreat sketches a template other platforms inherit: automated perception features ship fast, but face a reckoning that ends either in user-facing controls or removal — a cost the original UX gain rarely priced in.
The trend: Platform machine-learning features are migrating from invisible automated decisions toward user control or outright removal as bias scrutiny catches up with them.